ICLR2024 compressed

This page contains one-sentence summaries of ICLR2024 accepted papers, generated by the compressor, my personal LLM-based project.

Breaking Physical and Linguistic Borders: Multilingual Federated Prompt Tuning for Low-Resource Languages

https://openreview.net/forum?id=zzqn5G9fjn

Compressor summary: The paper proposes a method for fine-tuning multilingual language models that preserves user privacy and improves performance, especially for low-resource languages.

On the generalization capacity of neural networks during generic multimodal reasoning

https://openreview.net/forum?id=zyBJodMrn5

Compressor summary: The study evaluates how well different neural network architectures handle multimodal generalization tasks and introduces a new benchmark called gCOG.

Learning Mean Field Games on Sparse Graphs: A Hybrid Graphex Approach

https://openreview.net/forum?id=zwU9scoU4A

Compressor summary: The authors propose Graphex Mean Field Games (GXMFGs), which combine graph theoretical concepts with multi-agent reinforcement learning to handle sparse network topologies, and present a novel learning algorithm for these games.

Out-of-Variable Generalisation for Discriminative Models

https://openreview.net/forum?id=zwMfg9PfPs

Compressor summary: The paper explores how agents can generalize to new environments with unobserved variables by reusing past marginal information and using residual distributions to estimate causal relationships.

Enhancing Small Medical Learners with Privacy-preserving Contextual Prompting

https://openreview.net/forum?id=ztpy1gsUpT

Compressor summary: The paper proposes a method to boost the performance of small language models in medical tasks while preserving patient privacy by using large language models' knowledge-intensive context.

Get What You Want, Not What You Don't: Image Content Suppression for Text-to-Image Diffusion Models

https://openreview.net/forum?id=zpVPhvVKXk

Compressor summary: The paper presents two techniques to improve text-to-image diffusion models by removing undesired content from text embeddings, which are tested on different types of diffusion models.

Nemesis: Normalizing the soft-prompt vectors of vision-language models

https://openreview.net/forum?id=zmJDzPh1Dm

Compressor summary: This paper explores how normalizing soft prompt vectors in vision-language models affects their performance and proposes a method to optimize this process.

Fast and unified path gradient estimators for normalizing flows

https://openreview.net/forum?id=zlkXLb3wpF

Compressor summary: This paper introduces a fast path gradient estimator for normalizing flows that works with any relevant architecture, improves training, and allows maximum likelihood estimation for various scientific problems.

Long-Term Typhoon Trajectory Prediction: A Physics-Conditioned Approach Without Reanalysis Data

https://openreview.net/forum?id=ziDFH8TPPK

Compressor summary: The paper proposes a new approach to predict typhoon trajectories using real-time data and outperforms existing methods in providing accurate forecasts up to 72 hours in advance.

Rigid Protein-Protein Docking via Equivariant Elliptic-Paraboloid Interface Prediction

https://openreview.net/forum?id=zgQ0PHeGnL

Compressor summary: ElliDock is a novel learning-based method that predicts elliptic paraboloid interfaces to enable fast and accurate protein-protein docking, especially for antibody-antigen pairs.

On the Joint Interaction of Models, Data, and Features

https://openreview.net/forum?id=ze7DOLi394

Compressor summary: The interaction tensor is a tool to analyze how features in deep learning models are influenced by data and random seeds, leading to a better understanding of feature learning and generalization error estimation.

A ROBUST DIFFERENTIAL NEURAL ODE OPTIMIZER

https://openreview.net/forum?id=zbOSJ3CATY

Compressor summary: The paper proposes GTSONO, a game theoretic optimizer for neural ODEs that is more robust to adversarial attacks and has computational benefits.

The Dark Side of the Hyperbolic Moon

https://openreview.net/forum?id=zbKcFZ6Dbp

Compressor summary: Shadow cones use shadows formed by a light source in hyperbolic space to model hierarchical relations in data better than existing entailment cone constructions.

How do Language Models Bind Entities in Context?

https://openreview.net/forum?id=zb3b6oKO77

Compressor summary: LMs represent facts about entities and their attributes using binding ID vectors, which are interpretable and transferable across tasks.

On the Limitations of Temperature Scaling for Distributions with Overlaps

https://openreview.net/forum?id=zavLQJ1XjB

Compressor summary: This paper compares temperature scaling and Mixup, two techniques for calibrating deep neural networks' confidence, and shows that Mixup performs better when there are overlapping classes or noisy labels.

FROSTER: Frozen CLIP is A Strong Teacher for Open-Vocabulary Action Recognition

https://openreview.net/forum?id=zYXFMeHRtO

Compressor summary: FROSTER is a framework that adapts CLIP for open-vocabulary action recognition by distilling features from a frozen CLIP model and adding video-specific features.

Detecting Pretraining Data from Large Language Models

https://openreview.net/forum?id=zWqr3MQuNs

Compressor summary: The paper proposes a new method called MIN-K PROB to detect if an LLM was trained on a given text without knowing the pretraining data, using WIKIMIA as a dynamic benchmark and outperforming previous methods.

Task Adaptation from Skills: Information Geometry, Disentanglement, and New Objectives for Unsupervised Reinforcement Learning

https://openreview.net/forum?id=zSxpnKh1yS

Compressor summary: The paper proposes a new metric (WSEP) to improve unsupervised reinforcement learning by enhancing skill diversity and separatability, and shows that using Wasserstein distance instead of KL divergence leads to better task adaptation.

Neural Processing of Tri-Plane Hybrid Neural Fields

https://openreview.net/forum?id=zRkM6UcA22

Compressor summary: The paper explores how tri-plane discrete data structures can effectively process neural fields for tasks like classification and part segmentation without sacrificing reconstruction quality.

Error Norm Truncation: Robust Training in the Presence of Data Noise for Text Generation Models

https://openreview.net/forum?id=zMvMwNvs4R

Compressor summary: ENT improves text generation quality and robustness against noise by truncating noisy data based on its distribution of non-target tokens.

CADS: Unleashing the Diversity of Diffusion Models through Condition-Annealed Sampling

https://openreview.net/forum?id=zMoNrajk2X

Compressor summary: Condition-Annealed Diffusion Sampler (CADS) improves diversity and quality of diffusion models by adjusting the conditioning signal during inference.

Protein Discovery with Discrete Walk-Jump Sampling

https://openreview.net/forum?id=zMPHKOmQNb

Compressor summary: The paper proposes a discrete sampling method for generative models that improves sample quality and robustness, and applies it to generate functional antibody proteins with high success rate.

SetCSE: Set Operations using Contrastive Learning of Sentence Embeddings

https://openreview.net/forum?id=zEHGSN8Hy8

Compressor summary: SetCSE is an innovative information retrieval framework that uses sets to represent complex semantics and improve language model comprehension for structured information querying.

AgentBench: Evaluating LLMs as Agents

https://openreview.net/forum?id=zAdUB0aCTQ

Compressor summary: AgentBench is a benchmark to evaluate large language models as agents in complex interactive tasks, revealing their strengths and weaknesses in reasoning and decision-making.

MathCoder: Seamless Code Integration in LLMs for Enhanced Mathematical Reasoning

https://openreview.net/forum?id=z8TW0ttBPp

Compressor summary: The paper introduces MathCodeInstruct, a method to fine-tune language models using code for solving math problems, and presents MathCoder models that achieve state-of-the-art results on MATH and GSM8K datasets.

Perceptual Measurements, Distances and Metrics

https://openreview.net/forum?id=z7K2faBrDG

Compressor summary: The text discusses how measuring the perceptual scale of physical variables can improve probabilistic modeling of perception and lead to a better understanding of perceptual distances.

An Efficient Tester-Learner for Halfspaces

https://openreview.net/forum?id=z6n1fKMMC1

Compressor summary: The paper presents an efficient algorithm for learning halfspaces in the testable learning model with different types of noise and target distributions, using new tests that leverage the labels and moment-matching techniques.

Byzantine Robust Cooperative Multi-Agent Reinforcement Learning as a Bayesian Game

https://openreview.net/forum?id=z6KS9D1dxt

Compressor summary: The study proposes a Bayesian framework for cooperative multi-agent reinforcement learning that robustly handles Byzantine failures and adversarial attacks by learning policies based on beliefs about other agents' types.

Pre-training with Random Orthogonal Projection Image Modeling

https://openreview.net/forum?id=z4Hcegjzph

Compressor summary: ROPIM is a self-supervised method for visual pre-training without labels that uses random orthogonal projection instead of binary masking and achieves better performance than MIM.

Revisit and Outstrip Entity Alignment: A Perspective of Generative Models

https://openreview.net/forum?id=z3dfuRcGAK

Compressor summary: The paper explores embedding-based entity alignment as a generative model problem, proposes a new framework called generative entity alignment (GEEA) that uses mutual variational autoencoder (M-VAE) for better conversion between knowledge graphs and entity synthesis from noise vectors.

Massively Scalable Inverse Reinforcement Learning for Route Optimization

https://openreview.net/forum?id=z3L59iGALM

Compressor summary: The paper presents RHIP, a novel IRL method for route recommendation that leverages scaling techniques, eigenvector algorithms, and planning horizons to achieve significant improvements in route quality.

Matcher: Segment Anything with One Shot Using All-Purpose Feature Matching

https://openreview.net/forum?id=yzRXdhk2he

Compressor summary: The paper introduces Matcher, a perception paradigm that leverages off-the-shelf vision foundation models for various segmentation tasks without training, achieving impressive performance on COCO-20 and LVIS-92 benchmarks.

Generating Pragmatic Examples to Train Neural Program Synthesizers

https://openreview.net/forum?id=yxKZGQLzOP

Compressor summary: The paper proposes a novel method to improve program synthesis from user-provided examples by sampling pairs of programs and examples and using pragmatic inference to choose informative training examples for the synthesizer.

Empirical Analysis of Model Selection for Heterogeneous Causal Effect Estimation

https://openreview.net/forum?id=yuy6cGt3KL

Compressor summary: The authors study how to select the best model for estimating the treatment effect in causal inference using various metrics and real datasets, suggesting new strategies based on hyperparameter tuning and causal ensembling.

A Quadratic Synchronization Rule for Distributed Deep Learning

https://openreview.net/forum?id=yroyhkhWS6

Compressor summary: The Quadratic Synchronization Rule (QSR) method improves test accuracy in local gradient methods for distributed deep learning by dynamically adjusting the local computation steps $H$ based on the learning rate $\eta$.

Equivariant Matrix Function Neural Networks

https://openreview.net/forum?id=yrgQdA5NkI

Compressor summary: MFNs are a new architecture for learning on graphs that captures non-local interactions using matrix equivariant functions, achieving high performance and scalability.

In defense of parameter sharing for model-compression

https://openreview.net/forum?id=ypAT2ixD4X

Compressor summary: This paper shows that randomized parameter-sharing (RPS) outperforms other methods for compressing neural network models and achieves better trade-offs between memory and accuracy, especially in high compression settings. It also introduces a modified RPS technique called STABLE-RPS to address stability and Pareto-continuity issues.

Image Translation as Diffusion Visual Programmers

https://openreview.net/forum?id=yozwqhIHXj

Compressor summary: DVP is a neuro-symbolic image translation framework that uses computer vision models for pro-symbolic steps and achieves condition-flexible, in-context reasoning, and systemic controllability and explainability.

Achieving Fairness in Multi-Agent MDP Using Reinforcement Learning

https://openreview.net/forum?id=yoVq2BGQdP

Compressor summary: The paper presents an RL method that ensures fairness in multi-agent MDPs with unknown environments using a fairness function, online convex optimization, confidence bounds, and policy-gradient method.

Consistency Trajectory Models: Learning Probability Flow ODE Trajectory of Diffusion

https://openreview.net/forum?id=ymjI8feDTD

Compressor summary: CTM is a new diffusion model sampling method that balances speed and quality by combining adversarial training and denoising score matching, and enables various sampling schemes and exact likelihood computation.

Image Background Serves as Good Proxy for Out-of-distribution Data

https://openreview.net/forum?id=ym0ubZrsmm

Compressor summary: The paper proposes a new model (SSOD) that improves out-of-distribution detection without using explicit out-of-distribution data by exploiting natural cues from in-distribution data.

Pre-Training and Fine-Tuning Generative Flow Networks

https://openreview.net/forum?id=ylhiMfpqkm

Compressor summary: The paper proposes a novel reward-free pre-training method for Generative Flow Networks, which can adapt to downstream tasks and discover modes more efficiently.

Incentivized Truthful Communication for Federated Bandits

https://openreview.net/forum?id=ykEixGIJYb

Compressor summary: Truth-FedBan is a communication protocol that motivates truthful reporting in federated bandit learning by providing independent incentives unrelated to self-reported costs, leading to near-optimal performance and efficiency.

Chain of Log-Concave Markov Chains

https://openreview.net/forum?id=yiMB2DOjsR

Compressor summary: The paper proposes a new non-Markovian sampling method based on Gaussian kernels, log-concave conditional densities, and empirical Bayes that can tunnel between modes and is compared with existing Langevin MCMC algorithms.

VDC: Versatile Data Cleanser for Detecting Dirty Samples via Visual-Linguistic Inconsistency

https://openreview.net/forum?id=ygxTuVz9eU

Compressor summary: VDC uses multimodal language models to detect and clean dirty data in AI systems by aligning and reasoning across images and labels.

Variance-enlarged Poisson Learning for Graph-based Semi-Supervised Learning with Extremely Sparse Labeled Data

https://openreview.net/forum?id=yeeVBMDAwy

Compressor summary: VPL is a framework that reduces label degeneracy by using variance-enlarged regularization and improves unsupervised learning on graphs.

EmerNeRF: Emergent Spatial-Temporal Scene Decomposition via Self-Supervision

https://openreview.net/forum?id=ycv2z8TYur

Compressor summary: EmerNeRF is a method that learns spatial-temporal representations of dynamic scenes using neural fields, achieving state-of-the-art performance in sensor simulation and addressing positional bias in Transformers.

Towards Assessing and Benchmarking Risk-Return Tradeoff of Off-Policy Evaluation

https://openreview.net/forum?id=ycF7mKfVGO

Compressor summary: SharpeRatio@k is a new metric that measures the risk-return tradeoff and efficiency of policy portfolios formed by an OPE estimator under varying online evaluation budgets, aiming to improve the selection of policies for online deployment.

Safe Collaborative Filtering

https://openreview.net/forum?id=yarUvgEXq3

Compressor summary: The study develops a scalable and effective recommender system that ensures high quality for less-satisfied users by minimizing the risk over their loss.

Hybrid Sharing for Multi-Label Image Classification

https://openreview.net/forum?id=yVJd8lKyVX

Compressor summary: The Hybrid Sharing Query (HSQ) is a transformer-based model that uses a mixture-of-experts architecture to improve image multi-label classification by leveraging label correlations and mitigating heterogeneity.

Bayes Conditional Distribution Estimation for Knowledge Distillation Based on Conditional Mutual Information

https://openreview.net/forum?id=yV6wwEbtkR

Compressor summary: The paper proposes MCMI, a new method to estimate BCPD for KD that enhances the teacher's ability to capture contextual information and boosts the student's performance in zero-shot and few-shot scenarios.

ValUES: A Framework for Systematic Validation of Uncertainty Estimation in Semantic Segmentation

https://openreview.net/forum?id=yV6fD7LYkF

Compressor summary: The paper proposes an evaluation framework to systematically assess uncertainty methods for semantic segmentation applications, addressing key research questions and revealing insights on data and model uncertainties.

Provable Reward-Agnostic Preference-Based Reinforcement Learning

https://openreview.net/forum?id=yTBXeXdbMf

Compressor summary: This study develops a new theoretical framework for preference-based reinforcement learning that reduces human feedback and can handle linear and low-rank MDPs, as well as action-based comparison feedback.

Sparse MoE with Language Guided Routing for Multilingual Machine Translation

https://openreview.net/forum?id=ySS7hH1smL

Compressor summary: The authors propose Lingual-SMoE, a multilingual machine translation framework that adapts routing policies based on linguistic features and difficulty levels to improve performance over existing methods.

Conversational Drug Editing Using Retrieval and Domain Feedback

https://openreview.net/forum?id=yRrPfKyJQ2

Compressor summary: ChatDrug is a framework using LLMs to systematically investigate and improve drug editing tasks, achieving the best performance on all 39 tasks and providing diverse, valid suggestions and explanations.

Bidirectional Temporal Diffusion Model for Temporally Consistent Human Animation

https://openreview.net/forum?id=yQDFsuG9HP

Compressor summary: The paper proposes a new method to generate realistic human animation from limited input using bidirectional temporal modeling and denoising diffusion models.

Transformers as Decision Makers: Provable In-Context Reinforcement Learning via Supervised Pretraining

https://openreview.net/forum?id=yN4Wv17ss3

Compressor summary: This paper analyzes how supervised pretraining helps transformer models perform in-context reinforcement learning, and shows they can approximate near-optimal algorithms for certain tasks.

MAmmoTH: Building Math Generalist Models through Hybrid Instruction Tuning

https://openreview.net/forum?id=yLClGs770I

Compressor summary: MAmmoTH is a new open-source language model that excels at solving various math problems using a unique combination of chain-of-thought and program-of-thought rationales, achieving significant accuracy gains over existing models.

Faithful and Efficient Explanations for Neural Networks via Neural Tangent Kernel Surrogate Models

https://openreview.net/forum?id=yKksu38BpM

Compressor summary: This paper explores how to use simpler ML algorithms to analyze and explain neural networks, focusing on new methods for approximating the empirical neural tangent kernel (eNTK) that balance time, memory, and accuracy.

Inverse Approximation Theory for Nonlinear Recurrent Neural Networks

https://openreview.net/forum?id=yC2waD70Vj

Compressor summary: The paper proves that nonlinear RNNs can approximate certain sequences well if they have an exponential decaying memory structure, and proposes a reparameterization method to address the limitations of RNNs for long-term memory.

Bandits with Replenishable Knapsacks: the Best of both Worlds

https://openreview.net/forum?id=yBIJRIYTqa

Compressor summary: The paper proposes a generalization of the bandits with knapsacks framework that allows resource replenishment and shows its effectiveness in online learning and economic applications.

Adapting Large Language Models via Reading Comprehension

https://openreview.net/forum?id=y886UXPEZ0

Compressor summary: The authors propose a method to transform raw corpora into reading comprehension texts, which improves large language models' performance on various tasks and domains without requiring domain-specific training.

AdjointDPM: Adjoint Sensitivity Method for Gradient Backpropagation of Diffusion Probabilistic Models

https://openreview.net/forum?id=y33lDRBgWI

Compressor summary: The paper proposes AdjointDPM, a method to optimize diffusion models with less memory by using probability-flow ODEs and the adjoint sensitivity method for efficient gradient backpropagation.

PolyGCL: GRAPH CONTRASTIVE LEARNING via Learnable Spectral Polynomial Filters

https://openreview.net/forum?id=y21ZO6M86t

Compressor summary: PolyGCL is a graph contrastive learning technique that uses polynomial filters to capture high-pass information for heterophily, improving performance on both homophilic and heterophilic graphs.

Is Self-Repair a Silver Bullet for Code Generation?

https://openreview.net/forum?id=y0GJXRungR

Compressor summary: This paper analyzes how well large language models like Code Llama, GPT-3.5, and GPT-4 can fix their own mistakes in code generation tasks and finds that self-repair is often limited by the model's feedback ability.

DreamLLM: Synergistic Multimodal Comprehension and Creation

https://openreview.net/forum?id=y01KGvd9Bw

Compressor summary: DreamLLM is a framework that uses direct sampling and interleaved document generation to create versatile multimodal models with better comprehension and creation capabilities.

Un-Mixing Test-Time Normalization Statistics: Combatting Label Temporal Correlation

https://openreview.net/forum?id=xyxU99Nutg

Compressor summary: UnMix-TNS is a novel strategy to adapt batch normalization parameters online using mixed statistics components, improving the robustness of test-time adaptation methods under non-i.i.d. conditions.

Revisiting the Last-Iterative Convergence of Stochastic Gradient Methods

https://openreview.net/forum?id=xxaEhwC1I4

Compressor summary: This paper investigates the optimal convergence rate of SGD's last iterate for various optimization scenarios without restrictive assumptions, providing a unified proof method.

Sparse Model Soups: A Recipe for Improved Pruning via Model Averaging

https://openreview.net/forum?id=xx0ITyHp3u

Compressor summary: Sparse Model Soups improves neural network compression by combining sparse models with shared connectivity, enhancing generalization and out-of-distribution performance.

Emergent mechanisms for long timescales depend on training curriculum and affect performance in memory tasks

https://openreview.net/forum?id=xwKt6bUkXj

Compressor summary: The text discusses how different mechanisms in recurrent neural networks help them solve memory-dependent tasks with varying memory requirements, and how a multi-head curriculum can improve their performance and stability.

Retrieval meets Long Context Large Language Models

https://openreview.net/forum?id=xw5nxFWMlo

Compressor summary: The paper compares two methods to enhance large language models (LLMs) for tasks like question answering and summarization: extending their context window or adding retrieval, and shows that retrieval can improve performance regardless of the context window size.

LQ-LoRA: Low-rank plus Quantized Matrix Decomposition for Efficient Language Model Finetuning

https://openreview.net/forum?id=xw29VvOMmU

Compressor summary: Our method adapts pretrained language models efficiently by decomposing matrices into high-precision and memory-efficient parts, updating the latter during finetuning.

Threaten Spiking Neural Networks through Combining Rate and Temporal Information

https://openreview.net/forum?id=xv8iGxENyI

Compressor summary: This paper proposes a hybrid adversarial attack (HART) on spiking neural networks (SNNs) using rate and temporal information, which outperforms previous methods in various scenarios.

ClimODE: Climate Forecasting With Physics-informed Neural ODEs

https://openreview.net/forum?id=xuY33XhEGR

Compressor summary: ClimODE is a novel deep learning model that incorporates physics-based advection and uncertainty estimation to improve climate prediction.

TACTiS-2: Better, Faster, Simpler Attentional Copulas for Multivariate Time Series

https://openreview.net/forum?id=xtOydkE1Ku

Compressor summary: The paper presents a new model for multivariate time series prediction using transformer-based attentional copulas, which improves performance and training dynamics while handling various tasks and data characteristics.

Dual RL: Unification and New Methods for Reinforcement and Imitation Learning

https://openreview.net/forum?id=xt9Bu66rqv

Compressor summary: The paper proposes dual RL framework to unify offline imitation learning and reinforcement learning, and improves their performance with new methods ReCOIL and $f$-DVL.

GLD: Generative Latent Dynamics for Structured Motion Representation and Learning

https://openreview.net/forum?id=xsd2llWYSA

Compressor summary: The text introduces a method that learns motion dynamics from sparse data using self-supervised representation and generation, enabling online tracking of various motions with safe action execution.

Inherently Interpretable Time Series Classification via Multiple Instance Learning

https://openreview.net/forum?id=xriGRsoAza

Compressor summary: MILLET is a framework that uses Multiple Instance Learning to make Time Series Classification models more interpretable without sacrificing performance.

Teaching Language Models to Hallucinate Less with Synthetic Tasks

https://openreview.net/forum?id=xpw7V0P136

Compressor summary: SynTra reduces hallucination in large language models by optimizing their system messages on a synthetic task before applying them to real-world summarization tasks.

On the Fairness ROAD: Robust Optimization for Adversarial Debiasing

https://openreview.net/forum?id=xnhvVtZtLD

Compressor summary: ROAD is a new algorithm that improves local fairness in machine learning by using distributionally robust optimization and adversarial learning.

Exploring Target Representations for Masked Autoencoders

https://openreview.net/forum?id=xmQMz9OPF5

Compressor summary: The paper proposes a new way to train masked autoencoders without carefully choosing target representations, using a randomly initialized model as the teacher and achieving better results on downstream tasks.

Federated Recommendation with Additive Personalization

https://openreview.net/forum?id=xkXdE81mOK

Compressor summary: FedRAP is a new federated learning model for recommendations that learns global and personalized item embeddings, improving performance and reducing communication costs.

Robust Model-Based Optimization for Challenging Fitness Landscapes

https://openreview.net/forum?id=xhEN0kJh4q

Compressor summary: The authors propose a new method using a novel VAE to overcome the problem of separation in protein design optimization, which leads to improved sampling and generalization across discrete and continuous spaces.

Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion

https://openreview.net/forum?id=xhCZD9hiiA

Compressor summary: This study proposes a batch-normalized MLP with linear activations and batch normalization that has bounded gradients and optimal signal propagation.

Learning to design protein-protein interactions with enhanced generalization

https://openreview.net/forum?id=xcMmebCT7s

Compressor summary: The work introduces PPIRef, a large 3D protein-protein interaction dataset, and PPIformer, a new SE(3)-equivariant model that predicts mutation effects on protein-protein interactions and generalizes well across diverse scenarios.

Statistical Rejection Sampling Improves Preference Optimization

https://openreview.net/forum?id=xbjSwwrQOe

Compressor summary: RLHF alternatives like SLiC and DPO have limitations in using preference data; RSO overcomes these by sourcing data from the optimal policy, and a unified framework improves loss functions for both methods.

Tell Your Model Where to Attend: Post-hoc Attention Steering for LLMs

https://openreview.net/forum?id=xZDWO0oejD

Compressor summary: PASTA is a method that allows large language models to read text with user-specified emphasis marks, enhancing their ability to follow instructions or integrate new knowledge.

Privacy Amplification for Matrix Mechanisms

https://openreview.net/forum?id=xUzWmFdglP

Compressor summary: The paper proposes MMCC, a new algorithm that analyzes privacy amplification via sampling for any matrix mechanism, and shows its utility in improving privacy/utility trade-offs for DP-FTRL style algorithms.

Negative Label Guided OOD Detection with Pretrained Vision-Language Models

https://openreview.net/forum?id=xUO1HXz4an

Compressor summary: NegLabel is a novel post hoc out-of-distribution (OOD) detection method for vision-language models that uses negative labels from corpus databases and achieves state-of-the-art performance with good generalization and robustness.

SALMON: Self-Alignment with Principle-Following Reward Models

https://openreview.net/forum?id=xJbsmB8UMx

Compressor summary: The paper proposes SALMON, a novel alignment paradigm for large language models that uses human-defined principles to control the behavior of an AI assistant without relying on extensive human feedback.

Impact of Computation in Integral Reinforcement Learning for Continuous-Time Control

https://openreview.net/forum?id=xJEd8PkdNz

Compressor summary: The choice of the quadrature rule in integral reinforcement learning affects control performance due to computational errors, which can be reduced by using Bayesian quadrature with an appropriate kernel function.

Constrained Bi-Level Optimization: Proximal Lagrangian Value function Approach and Hessian-free Algorithm

https://openreview.net/forum?id=xJ5N8qrEPl

Compressor summary: The paper proposes a new Hessian-free algorithm for solving constrained Bi-Level Optimization problems in machine learning, which uses a smooth proximal Lagrangian value function to handle constraints and performs well in practice.

Subtractive Mixture Models via Squaring: Representation and Learning

https://openreview.net/forum?id=xIHi5nxu9P

Compressor summary: Squaring subtractive mixtures in probabilistic circuits allows for better modeling of complex distributions with fewer components.

Towards 3D Molecule-Text Interpretation in Language Models

https://openreview.net/forum?id=xI4yNlkaqh

Compressor summary: 3D-MoLM is a language model that can comprehend 3D molecular structures by integrating a 3D molecular encoder, improving its performance in biomolecular tasks.

Correlated Noise Provably Beats Independent Noise for Differentially Private Learning

https://openreview.net/forum?id=xHmCdSArUC

Compressor summary: The paper proposes a method to improve privacy in machine learning by introducing correlations in the noise, providing analytical bounds for its utility, and demonstrating its effectiveness in experiments on deep learning.

How Over-Parameterization Slows Down Gradient Descent in Matrix Sensing: The Curses of Symmetry and Initialization

https://openreview.net/forum?id=xGvPKAiOhq

Compressor summary: This paper shows how over-parameterization affects gradient descent for matrix sensing problems, leading to faster convergence rates with an asymmetric parameterization and a novel method to remove dependency on initialization scale.

Elucidating the Exposure Bias in Diffusion Models

https://openreview.net/forum?id=xEJMoj1SpX

Compressor summary: The paper investigates the exposure bias problem in diffusion models and proposes Epsilon Scaling, a training-free method that improves generative performance across various frameworks and settings.

Score Regularized Policy Optimization through Diffusion Behavior

https://openreview.net/forum?id=xCRr9DrolJ

Compressor summary: The paper proposes a faster way to learn policies from diffusion models using deterministic inference, improving both speed and performance on offline reinforcement learning tasks.

Navigating Dataset Documentations in AI: A Large-Scale Analysis of Dataset Cards on HuggingFace

https://openreview.net/forum?id=xC8xh2RSs2

Compressor summary: Key findings from analyzing 7,433 dataset documentation cards on Hugging Face reveal the importance of completing all sections, prioritizing description and structure, and improving accessibility and reproducibility.

GeoDiffusion: Text-Prompted Geometric Control for Object Detection Data Generation

https://openreview.net/forum?id=xBfQZWeDRH

Compressor summary: GeoDiffusion is a framework that uses text prompts to generate high-quality detection data from pre-trained text-to-image diffusion models, encoding geometric conditions like bounding boxes and camera views.

On the Stability of Expressive Positional Encodings for Graph Neural Networks

https://openreview.net/forum?id=xAqcJ9XoTf

Compressor summary: SPE is a new method for graph neural networks that uses eigenvalues to softly partition eigenvectors, improving stability and expressivity over existing Laplacian-based positional encodings.

Are Bert Family Good Instruction Followers? A Study on Their Potential And Limitations

https://openreview.net/forum?id=x8VNtpCu1I

Compressor summary: This paper explores how encoder-only BERT models can perform instruction following and zero-shot learning using XML-R and shows that they achieve competitive results while being faster than autoregressive models.

A Restoration Network as an Implicit Prior

https://openreview.net/forum?id=x7d1qXEn1e

Compressor summary: The authors propose a method to improve image restoration by using pre-trained neural networks as implicit priors, achieving state-of-the-art performance.

Tag2Text: Guiding Vision-Language Model via Image Tagging

https://openreview.net/forum?id=x6u2BQ7xcq

Compressor summary: Tag2Text is a framework that uses image tags from paired text to improve vision-language models' performance on various tasks without extensive annotations.

Synaptic Weight Distributions Depend on the Geometry of Plasticity

https://openreview.net/forum?id=x5txICnnjC

Compressor summary: The assumption of Euclidean geometry in studying synaptic plasticity may be incorrect, and experiments could reveal the true geometry by comparing weight distributions before and after learning.

PROGRAM: PROtotype GRAph Model based Pseudo-Label Learning for Test-Time Adaptation

https://openreview.net/forum?id=x5LvBK43wg

Compressor summary: PROGRAM is a novel test-time adaptation method that uses a prototype graph model for reliable pseudo-label generation and robust self-training for target domain adaptation with noisy pseudo-labels.

Privacy-Preserving In-Context Learning for Large Language Models

https://openreview.net/forum?id=x4OPJ7lHVU

Compressor summary: DP-ICL is a method to protect sensitive information in large language models by generating noisy responses from multiple models.

ADoPD: A Large-Scale Document Page Decomposition Dataset

https://openreview.net/forum?id=x1ptaXpOYa

Compressor summary: ADoPD is a new dataset for analyzing documents that combines large-scale data, human expertise, and diversity to improve document understanding.

Urial: Aligning Untuned LLMs with Just the 'Write' Amount of In-Context Learning

https://openreview.net/forum?id=wxJ0eXwwda

Compressor summary: Alignment tuning mostly changes the language style of AI assistants, so an alternative method called Urial can achieve effective alignment with only a few stylistic examples.

Toward Student-oriented Teacher Network Training for Knowledge Distillation

https://openreview.net/forum?id=wsWGcw6qKD

Compressor summary: The paper explores how to train teachers for knowledge distillation using empirical risk minimization (ERM) and proposes a method called SoTeacher that improves student accuracy.

Consistency-guided Prompt Learning for Vision-Language Models

https://openreview.net/forum?id=wsRXwlwx4w

Compressor summary: CoPrompt is a new fine-tuning method for vision-language models that uses consistency constraints, perturbed inputs, and adapters/prompts to improve generalization on few-shot tasks.

On the Effect of Batch Size in Byzantine-Robust Distributed Learning

https://openreview.net/forum?id=wriKDQqiOQ

Compressor summary: This paper studies the effect of batch size on Byzantine-robust distributed learning and proposes a new method (ByzSGDnm) that increases model accuracy under attacks by using larger batch sizes and normalized momentum.

Benchmarking Algorithms for Federated Domain Generalization

https://openreview.net/forum?id=wprSv7ichW

Compressor summary: The paper proposes a new Federated Domain Generalization benchmark and evaluates 13 methods on it, finding significant performance gaps in this challenging setting.

Adaptive deep spiking neural network with global-local learning via balanced excitatory and inhibitory mechanism

https://openreview.net/forum?id=wpnlc2ONu0

Compressor summary: The paper proposes EIHL, an algorithm that adjusts network connectivity using excitation-inhibition and switches between local and global learning for better spiking neural network training performance and sparsity.

Sparsistency for inverse optimal transport

https://openreview.net/forum?id=wpXGPCBOTX

Compressor summary: This paper studies inverse optimal transport, which infers the ground cost from samples, and derives a sufficient condition for sparse recovery of the cost using $\ell_1$ regularization.

Compositional Generative Inverse Design

https://openreview.net/forum?id=wmX0CqFSd7

Compressor summary: Inverse design can be improved by optimizing over the learned energy function of a diffusion model, enabling compositional system design and avoiding adversarial examples.

Scalable Diffusion for Materials Generation

https://openreview.net/forum?id=wm4WlHoXpC

Compressor summary: The authors propose a novel unified representation for generating high-fidelity crystal structures from complex chemical systems using a diffusion probabilistic model and introduce new metrics to evaluate the quality of generated materials for downstream applications like discovering novel stable materials.

LUM-ViT: Learnable Under-sampling Mask Vision Transformer for Bandwidth Limited Optical Signal Acquisition

https://openreview.net/forum?id=wkbeqr5XhC

Compressor summary: The authors propose a novel approach using a Vision Transformer variant with a learnable under-sampling mask to reduce hyperspectral data acquisition volume and enhance real-time detection applications.

$\texttt{NAISR}$: A 3D Neural Additive Model for Interpretable Shape Representation

https://openreview.net/forum?id=wg8NPfeMF9

Compressor summary: The proposed NAISR method represents 3D shapes by deforming an atlas using disentangled covariates for interpretable shape discovery and transfer.

Navigating Text-To-Image Customization: From LyCORIS Fine-Tuning to Model Evaluation

https://openreview.net/forum?id=wfzXa8e783

Compressor summary: LyCORIS is an open-source library that helps fine-tune Stable Diffusion, a text-to-image generator, by offering various methodologies and assessing their effects with different metrics and prompts.

Robust Classification via Regression-Based Loss Reweighting and Label Correction

https://openreview.net/forum?id=wfgZc3IMqo

Compressor summary: The authors propose a unified method that combines loss reweighting and label correction to improve robustness against label noise in classification tasks using a shifted Gaussian model.

Unleashing the Potential of Fractional Calculus in Graph Neural Networks with FROND

https://openreview.net/forum?id=wcka3bd7P4

Compressor summary: FROND is a graph neural network model that uses fractional calculus to capture long-term memories and mitigate over-smoothing, improving graph representation learning.

Uni3D: Exploring Unified 3D Representation at Scale

https://openreview.net/forum?id=wcaE4Dfgt8

Compressor summary: Uni3D is a 3D foundation model that leverages 2D pretrained models and image-text aligned models to create scalable 3D representations, achieving state-of-the-art results on various 3D tasks.

Counterfactual Density Estimation using Kernel Stein Discrepancies

https://openreview.net/forum?id=wZXlEFO3tZ

Compressor summary: The paper proposes a method to estimate counterfactual distributions for causal effects using kernel Stein discrepancies, which is consistent and asymptotically normal under certain conditions.

Game-Theoretic Robust Reinforcement Learning Handles Temporally-Coupled Perturbations

https://openreview.net/forum?id=wZWTHU7AsQ

Compressor summary: The text introduces GRAD, a game-theoretic method for robust reinforcement learning that handles temporally-coupled perturbations, outperforming prior methods in experiments on continuous control tasks.

Mixture of Weak and Strong Experts on Graphs

https://openreview.net/forum?id=wYvuY60SdD

Compressor summary: Mowst is a method that combines a lightweight MLP and a GNN to classify nodes on graphs, with a confidence mechanism that controls when to use the GNN for better performance.

Benign Oscillation of Stochastic Gradient Descent with Large Learning Rate

https://openreview.net/forum?id=wYmvN3sQpG

Compressor summary: The study shows that neural networks trained with large learning rates and oscillations can learn weak features better than those trained with small learning rates, leading to improved generalization.

CLEX: Continuous Length Extrapolation for Large Language Models

https://openreview.net/forum?id=wXpSidPpc5

Compressor summary: CLEX is a novel method that extends the context window of LLMs by using ordinary differential equations to model position embedding scaling, enabling better performance in long-context applications without sacrificing accuracy or increasing latency.

MVSFormer++: Revealing the Devil in the Transformer's Details for Multi-View Stereo

https://openreview.net/forum?id=wXWfvSpYHh

Compressor summary: MVSFormer++ is a transformer-based model with attention mechanisms that enhances different components of the MVS pipeline, achieving state-of-the-art results on challenging benchmarks.

AutoVP: An Automated Visual Prompting Framework and Benchmark

https://openreview.net/forum?id=wR9qVlPh0P

Compressor summary: AutoVP is an end-to-end framework for automating visual prompting design choices and a benchmark for evaluating its performance on 12 image-classification tasks, significantly improving accuracy over existing methods.

STARC: A General Framework For Quantifying Differences Between Reward Functions

https://openreview.net/forum?id=wPhbtwlCDa

Compressor summary: The paper proposes a new class of pseudometrics called STARC metrics for comparing reward functions in reinforcement learning, which can help improve theoretical and empirical evaluation of reward learning algorithms.

Batch normalization is sufficient for universal function approximation in CNNs

https://openreview.net/forum?id=wOSYMHfENq

Compressor summary: The text explains why normalization layers like batch normalization are important for deep learning and shows how using enough random features can achieve universal function approximation in various network architectures.

Learning No-Regret Sparse Generalized Linear Models with Varying Observation(s)

https://openreview.net/forum?id=wISvONp3Kq

Compressor summary: The paper proposes an algorithm for training sparse generalized linear models online with varying observations, using a bilevel optimization framework and adaptive regularization parameters.

BarLeRIa: An Efficient Tuning Framework for Referring Image Segmentation

https://openreview.net/forum?id=wHLDHRkmEu

Compressor summary: BarLeRIa is a novel Parameter-Efficient Tuning framework for dense prediction tasks in computer vision that uses bi-directional intertwined vision language adapters and outperforms prior methods by 5.6%.

YaRN: Efficient Context Window Extension of Large Language Models

https://openreview.net/forum?id=wHBfxhZu1u

Compressor summary: YaRN extends the context window of language models using RoPE embeddings, improving generalization and extrapolation beyond the original pre-training.

Score-based generative models break the curse of dimensionality in learning a family of sub-Gaussian distributions

https://openreview.net/forum?id=wG12xUSqrI

Compressor summary: The paper analyzes how score-based generative models learn sub-Gaussian distributions and proves their approximation accuracy with respect to the target distribution.

Causal Inference with Conditional Front-Door Adjustment and Identifiable Variational Autoencoder

https://openreview.net/forum?id=wFf9m4v7oC

Compressor summary: The paper presents a novel method for estimating causal effects from observational data with unobserved confounding variables, using conditional front-door adjustment and deep generative models.

On the Hardness of Constrained Cooperative Multi-Agent Reinforcement Learning

https://openreview.net/forum?id=wFWuX1Fhtj

Compressor summary: The paper analyzes the limitations of primal-dual algorithms for constrained cooperative multi-agent reinforcement learning due to nonconvex constraints and proposes a decentralized primal approach with its own challenges.

Remote Sensing Vision-Language Foundation Models without Annotations via Ground Remote Alignment

https://openreview.net/forum?id=w9tc699w3Z

Compressor summary: The authors propose a method to train VLMs for remote-sensing images without textual annotations by using co-located internet imagery as an intermediary and show that their approach performs better than supervised methods on various tasks.

Linearity of Relation Decoding in Transformer Language Models

https://openreview.net/forum?id=w7LU2s14kE

Compressor summary: Transformer language models use linear relations to approximate some of their knowledge, which can be extracted by approximating the model with a single prompt.

Geometry-Aware Projective Mapping for Unbounded Neural Radiance Fields

https://openreview.net/forum?id=w7BwaDHppp

Compressor summary: The paper proposes a new mapping function for neural radiance fields (NeRFs) to generate novel views of unbounded scenes by adaptively sampling rays based on scene geometry.

Independent-Set Design of Experiments for Estimating Treatment and Spillover Effects under Network Interference

https://openreview.net/forum?id=w50MQ9Vfty

Compressor summary: The article proposes a new experimental design for causal inference on networks with interference, which improves performance by controlling interference exposure.

Reverse Forward Curriculum Learning for Extreme Sample and Demo Efficiency

https://openreview.net/forum?id=w4rODxXsmM

Compressor summary: The paper proposes RFCL, a method to improve reinforcement learning with offline data by using a reverse curriculum (generated via state resets) followed by a forward curriculum to efficiently leverage multiple demonstrations and solve complex tasks from sparse rewards.

Batched Low-Rank Adaptation of Foundation Models

https://openreview.net/forum?id=w4abltTZ2f

Compressor summary: FLORA is a framework that improves LoRA's ability to handle multiple task-specific adaptations efficiently, enabling fine-tuning of foundation models for diverse requests.

SuRe: Improving Open-domain Question Answering of LLMs via Summarized Retrieval

https://openreview.net/forum?id=w4DW6qkRmt

Compressor summary: SuRe is a framework that enhances open-domain question answering with large language models by generating summaries of retrieved passages and using them to select the most plausible answer.

MERT: Acoustic Music Understanding Model with Large-Scale Self-supervised Training

https://openreview.net/forum?id=w3YZ9MSlBu

Compressor summary: The paper proposes a self-supervised learning framework for music audio, using teacher models to provide pseudo labels in acoustic pre-training, achieving state-of-the-art results on 14 music understanding tasks.

Are Models Biased on Text without Gender-related Language?

https://openreview.net/forum?id=w1JanwReU6

Compressor summary: The study examines how large language models show gender biases in neutral sentences without obvious word-gender associations, and finds high levels of bias across different models.

Off-Policy Primal-Dual Safe Reinforcement Learning

https://openreview.net/forum?id=vy42bYs1Wo

Compressor summary: Our method improves safety and constraint satisfaction in primal-dual safe RL by reducing cost estimation uncertainty with conservative policy optimization and local policy convexification.

GRAPH-CONSTRAINED DIFFUSION FOR END-TO-END PATH PLANNING

https://openreview.net/forum?id=vuK8MhVtuu

Compressor summary: The paper proposes a new data-driven path planning model called GDP, which learns path patterns from road networks and outperforms traditional search-based methods.

CoRe-GD: A Hierarchical Framework for Scalable Graph Visualization with GNNs

https://openreview.net/forum?id=vtyasLn4RM

Compressor summary: The paper presents a scalable Graph Neural Network (GNN) based Graph Drawing framework that learns to optimize stress using coarsening and positional rewiring techniques, achieving state-of-the-art performance.

Grokking as the transition from lazy to rich training dynamics

https://openreview.net/forum?id=vt5mnLVIVo

Compressor summary: Grokking occurs when a neural network initially tries to fit a simple solution with its initial features before transitioning to feature learning and eventually generalizing.

Fixed-Budget Differentially Private Best Arm Identification

https://openreview.net/forum?id=vrE2fqAInO

Compressor summary: The paper studies how to find the best arm in a bandit problem with a limited budget and privacy constraints, using a policy based on maximum absolute determinants and deriving upper and lower bounds on the error probability.

Q-TAPE: A Task-Agnostic Pre-Trained Approach for Quantum Properties Estimation

https://openreview.net/forum?id=vrBVFXwAmi

Compressor summary: Q-TAPE is a pre-trained deep learning model that learns from diverse quantum systems in an unsupervised way, improving performance and reducing costs for quantum many-body problems.

Can Large Language Models Infer Causation from Correlation?

https://openreview.net/forum?id=vqIH0ObdqL

Compressor summary: The paper introduces Corr2Cause, a new task to test causal inference skills of large language models, and shows their poor performance and lack of generalization on it.

Decision Transformer is a Robust Contender for Offline Reinforcement Learning

https://openreview.net/forum?id=vpV7fOFQy4

Compressor summary: The paper compares three offline RL algorithms (CQL, BC, and DT) across various data settings and tasks to determine their strengths and weaknesses.

ModernTCN: A Modern Pure Convolution Structure for General Time Series Analysis

https://openreview.net/forum?id=vpJMJerXHU

Compressor summary: The paper proposes ModernTCN, a modified version of traditional TCN that improves performance and efficiency in time series analysis by increasing effective receptive fields.

From Bricks to Bridges: Product of Invariances to Enhance Latent Space Communication

https://openreview.net/forum?id=vngVydDWft

Compressor summary: The paper introduces a method to incorporate invariances into neural representations, improving latent similarity and downstream performance across different tasks and modalities.

Alice Benchmarks: Connecting Real World Object Re-Identification with the Synthetic

https://openreview.net/forum?id=vkkHqoerLV

Compressor summary: The Alice benchmarks provide large-scale datasets and evaluation protocols for object re-ID tasks using synthetic data, aiming to train effective models for real-world applications.

Universal Graph Random Features

https://openreview.net/forum?id=viftsX50Rt

Compressor summary: The paper introduces a new algorithm for fast and scalable estimation of graph kernel functions using random walks and neural networks.

CODE REPRESENTATION LEARNING AT SCALE

https://openreview.net/forum?id=vfzRRjumpX

Compressor summary: This paper proposes a two-stage pretraining scheme for code representation learning using a large amount of code data, which improves performance on various downstream tasks.

Towards Robust and Efficient Cloud-Edge Model Adaptation via Selective Entropy Distillation

https://openreview.net/forum?id=vePdNU3u6n

Compressor summary: CEMA is a paradigm that enables online adaptation of cloud models using limited computation and data transmission by excluding unreliable and low-informative samples and distilling normalization layers.

SAS: Structured Activation Sparsification

https://openreview.net/forum?id=vZfi5to2Xl

Compressor summary: SAS improves accuracy in neural networks by using sparse activation with a specific structure, reducing computation and leveraging NVIDIA's SparseTensorCore.

Learning Multi-Agent Communication with Contrastive Learning

https://openreview.net/forum?id=vZZ4hhniJU

Compressor summary: The paper proposes a contrastive learning approach to improve communication in multi-agent RL by using messages as encodings that capture global state information.

Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks

https://openreview.net/forum?id=vZ6r9GMT1n

Compressor summary: The paper proposes a simple and lightweight defense against black-box attacks on deep neural networks by adding random noise to intermediate layers, enhancing resilience without adversarial training or accuracy loss.

CodeChain: Towards Modular Code Generation Through Chain of Self-revisions with Representative Sub-modules

https://openreview.net/forum?id=vYhglxSj8j

Compressor summary: CodeChain is a framework that improves the modularity and correctness of code generated by large language models by guiding them to reuse previously developed sub-modules through a chain of self-revisions.

Vocos: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesis

https://openreview.net/forum?id=vY9nzQmQBw

Compressor summary: Vocos is a new neural vocoding model that generates speech from spectral coefficients using fast Fourier-based time-frequency representations, achieving high audio quality and significant computational efficiency.

SliceGPT: Compress Large Language Models by Deleting Rows and Columns

https://openreview.net/forum?id=vXxardq6db

Compressor summary: SliceGPT is a post-training sparsification technique that reduces the embedding dimension of large language models, enabling faster inference with fewer GPUs and minimal accuracy loss.

Provably Efficient Iterated CVaR Reinforcement Learning with Function Approximation and Human Feedback

https://openreview.net/forum?id=vW1SkPl4kp

Compressor summary: The paper introduces a risk-sensitive reinforcement learning framework that uses Iterated CVaR objective, human feedback, and provides sample-efficient algorithms with theoretical analysis and lower bound proof.

Enhancing Transfer Learning with Flexible Nonparametric Posterior Sampling

https://openreview.net/forum?id=vSwu81S33z

Compressor summary: Nonparametric transfer learning (NPTL) is a flexible posterior sampling method for deep neural networks that handles distribution shifts between upstream and downstream data better than existing methods.

How Many Pretraining Tasks Are Needed for In-Context Learning of Linear Regression?

https://openreview.net/forum?id=vSh5ePa0ph

Compressor summary: The paper studies how pretraining a simple linear attention model for linear regression with a Gaussian prior enables it to solve unseen tasks using in-context learning, showing that effective pretraining requires only a few tasks and that the model closely matches the Bayes optimal algorithm.

Efficient Backdoor Attacks for Deep Neural Networks in Real-world Scenarios

https://openreview.net/forum?id=vRyp2dhEQp

Compressor summary: The paper proposes realistic data-constrained backdoor attacks on DNNs, where attackers can't access all training data, and introduces CLIP-based techniques to improve efficiency in injecting malicious features.

Provable Memory Efficient Self-Play Algorithm for Model-free Reinforcement Learning

https://openreview.net/forum?id=vNiI3aGcE6

Compressor summary: The paper proposes a memory-efficient model-free self-play algorithm for two-player zero-sum Markov games in multi-agent reinforcement learning, achieving low space and computational complexity, and improving sample and burn-in cost compared to existing methods.

Fine-Tuned Language Models Generate Stable Inorganic Materials as Text

https://openreview.net/forum?id=vN9fpfqoP1

Compressor summary: The authors propose using fine-tuned large language models to generate novel materials with favorable properties for engineering by sampling from text-encoded atomistic data, which is simple, reliable, and captures crystal symmetries better than diffusion models.

SYMBOL: Generating Flexible Black-Box Optimizers through Symbolic Equation Learning

https://openreview.net/forum?id=vLJcd43U7a

Compressor summary: SYMBOL is a novel meta-learning framework that uses symbolic equation learning to automate the discovery of black-box optimizers, surpassing state-of-the-art baselines and showing exceptional zero-shot generalization.

Intriguing Properties of Data Attribution on Diffusion Models

https://openreview.net/forum?id=vKViCoKGcB

Compressor summary: The paper explores data attribution methods for diffusion models and finds that theoretically unjustified design choices perform better than previous baselines in attributing diffusion models' outputs to training data.

Skip-Attention: Improving Vision Transformers by Paying Less Attention

https://openreview.net/forum?id=vI95kcLAoU

Compressor summary: The paper proposes SkipAT, a method to reuse self-attention computations across layers in vision transformers, improving efficiency and performance.

Learning in reverse causal strategic environments with ramifications on two sided markets

https://openreview.net/forum?id=vEfmVS5ywF

Compressor summary: The text discusses how employers can strategically manipulate hiring policies to improve outcomes, but also shows that these policies may have unintended negative consequences in certain situations.

Multilingual Jailbreak Challenges in Large Language Models

https://openreview.net/forum?id=vESNKdEMGp

Compressor summary: The study shows that large language models can produce harmful content when prompted with non-English or jailbreak instructions, especially for low-resource languages, and proposes a self-defense framework to mitigate these risks.

Threshold-Consistent Margin Loss for Open-World Deep Metric Learning

https://openreview.net/forum?id=vE5MyzpP92

Compressor summary: The paper proposes a new metric and loss function to improve image retrieval performance and consistency across different classes and data distributions.

The Expressive Leaky Memory Neuron: an Efficient and Expressive Phenomenological Neuron Model Can Solve Long-Horizon Tasks.

https://openreview.net/forum?id=vE1e1mLJ0U

Compressor summary: The Expressive Leaky Memory (ELM) neuron model, which uses memory-like hidden states and nonlinear integration of synaptic input, accurately mimics cortical neuron computations with under 10,000 parameters and performs well on temporal tasks.

Encoding Unitig-level Assembly Graphs with Heterophilous Constraints for Metagenomic Contigs Binning

https://openreview.net/forum?id=vBw8JGBJWj

Compressor summary: UnitigBin is a novel metagenomic binning tool that uses representation learning and single-copy marker genes to improve the accuracy of clustering genomic sequences from microbial communities.

Understanding AI Cognition: A Neural Module for Inference Inspired by Human Memory Mechanisms

https://openreview.net/forum?id=vBo7544jZx

Compressor summary: The PMI framework mimics human brain's memory system to enhance question-answering, relation reasoning, and image classification tasks using Transformers and CNN models.

Consistency Models as a Rich and Efficient Policy Class for Reinforcement Learning

https://openreview.net/forum?id=v8jdwkUNXb

Compressor summary: The paper proposes a fast and effective policy representation called consistency policy for different reinforcement learning settings, using an actor-critic algorithm with a generative model as a policy.

Let's Verify Step by Step

https://openreview.net/forum?id=v8L0pN6EOi

Compressor summary: The text compares outcome and process supervision methods for training reliable large language models and shows that process supervision with active learning performs better on MATH problems.

Statistically Optimal $K$-means Clustering via Nonnegative Low-rank Semidefinite Programming

https://openreview.net/forum?id=v7ZPwoHU1j

Compressor summary: The paper introduces a novel clustering method that solves a nonnegative low-rank $K$-means problem using NMF, and shows its superior performance and guarantees.

Faster Sampling from Log-Concave Densities over Polytopes via Efficient Linear Solvers

https://openreview.net/forum?id=v63GWletn8

Compressor summary: The paper proposes an improved algorithm for sampling from logconcave distributions on a polytope with constraints, using efficient linear solvers, determinantal estimators, and randomized Taylor series.

RLCD: Reinforcement Learning from Contrastive Distillation for LM Alignment

https://openreview.net/forum?id=v3XXtxWKi6

Compressor summary: RLCD is a method that uses contrasting prompts to create preference pairs without human feedback, then trains a language model with reinforcement learning to follow alignment principles like harmlessness.

AutomaTikZ: Text-Guided Synthesis of Scientific Vector Graphics with TikZ

https://openreview.net/forum?id=v3K5TVP8kZ

Compressor summary: The authors propose using TikZ as an intermediate representation for scientific figures, introduce a large dataset (DaTikZ) of TikZ drawings with captions, and fine-tune LLaMA and CLiMA models on it, which outperform GPT-4 and Claude 2 in generating realistic and aligned figures.

Towards Characterizing Domain Counterfactuals for Invertible Latent Causal Models

https://openreview.net/forum?id=v1VvCWJAL8

Compressor summary: The paper proposes a method to estimate domain counterfactuals, which are what a sample would have looked like in a different environment, by making assumptions on invertibility and sparse interventions, and shows that this approach improves counterfactual estimation over existing methods.

How to Capture Higher-order Correlations? Generalizing Matrix Softmax Attention to Kronecker Computation

https://openreview.net/forum?id=v0zNCwwkaV

Compressor summary: The text introduces a generalized attention scheme that captures triple-wise correlations using tensors and shows its advantages and limitations in terms of efficiency and expressivity.

On Stationary Point Convergence of PPO-Clip

https://openreview.net/forum?id=uznKlCpWjV

Compressor summary: This paper analyzes the theoretical convergence properties of PPO-Clip, a popular reinforcement learning algorithm that uses a clipped surrogate objective function.

Bayesian Coreset Optimization for Personalized Federated Learning

https://openreview.net/forum?id=uz7d2N2zul

Compressor summary: The paper proposes a personalized coreset weighted federated learning method that reduces training time and improves accuracy by using representative data points instead of the entire client data.

Bayesian Neural Controlled Differential Equations for Treatment Effect Estimation

https://openreview.net/forum?id=uwO71a8wET

Compressor summary: The paper proposes a new Bayesian method that can estimate both the treatment effect and its uncertainty in continuous time using neural equations, which can help doctors make better decisions in personalized medicine.

DEEP NEURAL NETWORK INITIALIZATION WITH SPARSITY INDUCING ACTIVATIONS

https://openreview.net/forum?id=uvXK8Xk9Jk

Compressor summary: The text discusses how using sparse activations in deep networks can improve efficiency, but shows that some natural sparsifiers have training instability issues and suggests clipping their magnitude to fix this.

Beyond Spatio-Temporal Representations: Evolving Fourier Transform for Temporal Graphs

https://openreview.net/forum?id=uvFhCUPjtI

Compressor summary: The Evolving Graph Fourier Transform (EFT) is a new method for capturing the evolving spectra of graphs over time, which is efficient, invertible, and effective for downstream tasks.

LLMs Meet VLMs: Boost Open Vocabulary Object Detection with Fine-grained Descriptors

https://openreview.net/forum?id=usrChqw6yK

Compressor summary: DVDet is a new open-vocabulary detector that uses conditional context prompts and hierarchical textual descriptors to align visual embeddings with fine-grained text descriptions of object parts, achieving state-of-the-art performance.

Bridging Neural and Symbolic Representations with Transitional Dictionary Learning

https://openreview.net/forum?id=uqxBTcWRnj

Compressor summary: The paper presents a new framework for learning symbolic knowledge from images using a game-theoretic diffusion model and proposes two evaluation metrics that align with human judgments.

Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks

https://openreview.net/forum?id=up6hr4hIQH

Compressor summary: This paper proposes new fidelity measures for GNN explainability that resist distribution shifts and align better with the information-theoretic definition of explainability.

GROOT: Learning to Follow Instructions by Watching Gameplay Videos

https://openreview.net/forum?id=uleDLeiaT3

Compressor summary: The paper presents GROOT, an AI agent that learns to follow reference videos as instructions in open-world environments, achieving competitive performance against human players and generalist agents in Minecraft SkillForge.

Hierarchical Context Merging: Better Long Context Understanding for Pre-trained LLMs

https://openreview.net/forum?id=ulaUJFd96G

Compressor summary: HOMER is a training-free scheme that splits long inputs into smaller units, fuses them at Transformer layers, and reduces token count to improve performance and memory efficiency of large language models.

Enhancing High-Resolution 3D Generation through Pixel-wise Gradient Clipping

https://openreview.net/forum?id=ukidfml68f

Compressor summary: The text proposes a new technique called Pixel-wise Gradient Clipping (PGC) that improves the quality of 3D object generation by controlling the gradients used in latent representation-based models.

A Discretization Framework for Robust Contextual Stochastic Optimization

https://openreview.net/forum?id=ueTdErd5Ib

Compressor summary: The paper proposes a novel method for solving uncertain optimization problems that combines learning with downstream optimization, discretizes the feasible region into subsets, and integrates individual approximations to prescribe decisions that minimize expected cost and protect against worst case scenarios.

Mathematical Justification of Hard Negative Mining via Isometric Approximation Theorem

https://openreview.net/forum?id=udO3k28bEw

Compressor summary: The paper explains how hard negative mining works in deep metric learning and shows its equivalence to an optimization problem, providing theoretical justification for its effectiveness in preventing network collapse.

Implicit Neural Representations and the Algebra of Complex Wavelets

https://openreview.net/forum?id=uZfjFyPAvn

Compressor summary: The paper explores how implicit neural representations using wavelets can better capture high-frequency features of signals in image processing.

Embodied Active Defense: Leveraging Recurrent Feedback to Counter Adversarial Patches

https://openreview.net/forum?id=uXjfOmTiDt

Compressor summary: Embodied Active Defense (EAD) is a proactive strategy that uses recurrent feedback mechanisms and active vision to counter adversarial patches in 3D real-world settings, enhancing robustness without compromising accuracy.

MoLE: Mixture of LoRA Experts

https://openreview.net/forum?id=uWvKBCYh4S

Compressor summary: The paper proposes a new method called Mixture of LoRA Experts (MoLE) to effectively combine multiple LoRAs for fine-tuning pre-trained models, overcoming challenges in existing approaches and achieving better results in NLP and V\&L tasks.

Unbiased Watermark for Large Language Models

https://openreview.net/forum?id=uWVC5FVidc

Compressor summary: This study shows that it is possible to integrate watermarks in large language models without affecting their output quality or performance, enabling responsible AI development.

In-context Autoencoder for Context Compression in a Large Language Model

https://openreview.net/forum?id=uREj4ZuGJE

Compressor summary: The In-context Autoencoder (ICAE) uses a large language model to compress long texts into compact memory slots for efficient response generation.

Unpaired Image-to-Image Translation via Neural Schrödinger Bridge

https://openreview.net/forum?id=uQBW7ELXfO

Compressor summary: UNSB is a new method for unpaired image-to-image translation using Schrödinger Bridge with improved discriminators and regularization.

Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs

https://openreview.net/forum?id=uNrFpDPMyo

Compressor summary: The study presents a method to reduce memory usage in generative inference for large language models by adaptively compressing the key-value cache based on attention module profiles.

Learning with Mixture of Prototypes for Out-of-Distribution Detection

https://openreview.net/forum?id=uNkKaD3MCs

Compressor summary: PALM is a distance-based OOD detection method that models each class with multiple prototypes to capture sample diversities and enhance OOD detection performance.

Self-supervised Pocket Pretraining via Protein Fragment-Surroundings Alignment

https://openreview.net/forum?id=uMAujpVi9m

Compressor summary: The paper proposes a novel pocket pretraining method that leverages knowledge from atomic protein structures and pretrained small molecule representations to improve biomedical applications involving pockets and ligands.

BEND: Benchmarking DNA Language Models on Biologically Meaningful Tasks

https://openreview.net/forum?id=uKB4cFNQFg

Compressor summary: The study presents BEND, a benchmark for evaluating DNA language models on realistic and biologically meaningful tasks related to genome annotation.

GTA: A Geometry-Aware Attention Mechanism for Multi-View Transformers

https://openreview.net/forum?id=uJVHygNeSZ

Compressor summary: The paper proposes a geometry-aware attention mechanism for 3D vision tasks that encodes the geometric structure of tokens, leading to improved learning efficiency and performance of transformer-based neural networks.

In-context Exploration-Exploitation for Reinforcement Learning

https://openreview.net/forum?id=uIKZSStON3

Compressor summary: The ICEE algorithm improves the efficiency and speed of in-context policy learning by optimizing exploration-exploitation trade-offs within a Transformer model without Bayesian inference.

Expected flow networks in stochastic environments and two-player zero-sum games

https://openreview.net/forum?id=uH0FGECSEI

Compressor summary: EFlowNets improve GFlowNets for stochastic tasks, while AFlowNets enable strong performance in adversarial environments like two-player games.

Boosting Selective Rationalization with Shortcuts Discovery

https://openreview.net/forum?id=uGtfk2OphU

Compressor summary: The paper proposes a Selective Rationalization method that uses shortcuts to improve neural network explanations and provides data augmentation techniques to increase annotated rationales.

Differentially Private SGD Without Clipping Bias: An Error-Feedback Approach

https://openreview.net/forum?id=uFbWHyTlPn

Compressor summary: The paper proposes a new differentially private algorithm for training deep learning models with sensitive data, which avoids constant bias from gradient clipping and achieves higher accuracy.

Enhanced Face Recognition using Intra-class Incoherence Constraint

https://openreview.net/forum?id=uELjxVbrqG

Compressor summary: The paper proposes a feature recombination technique and a IIC constraint to improve FR performance by exploiting the feature representation space of different models.

Entity-Centric Reinforcement Learning for Object Manipulation from Pixels

https://openreview.net/forum?id=uDxeSZ1wdI

Compressor summary: The paper proposes a structured method for visual reinforcement learning that can handle multiple object manipulation and goal dependencies, enabling agents to generalize from 3 objects to over 10 objects tasks.

Augmenting transformers with recursively composed multi-grained representations

https://openreview.net/forum?id=u859gX7ADC

Compressor summary: ReCAT is a Transformer model that can learn hierarchical syntax without gold trees and use a novel CIO layer for better contextualization and inter-span communication.

Adversarial Training on Purification (AToP): Advancing Both Robustness and Generalization

https://openreview.net/forum?id=u7559ZMvwY

Compressor summary: AToP is a novel framework that combines random transforms and purifier model fine-tuning to enhance the robustness and generalization of deep neural networks against adversarial attacks.

Large Language Models as Generalizable Policies for Embodied Tasks

https://openreview.net/forum?id=u6imHU4Ebu

Compressor summary: The paper presents LLaRP, a method to adapt large language models as generalizable policies for embodied visual tasks, achieving higher success rates than common baselines on unseen tasks and introducing a new benchmark, Language Rearrangement.

ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models

https://openreview.net/forum?id=u48tHG5f66

Compressor summary: The authors propose a method to generate high-resolution images from pre-trained diffusion models without training, addressing object repetition and unreasonable structures by adjusting the convolutional perception field with re-dilation and using dispersed convolution and noise-damped guidance.

Scaling physics-informed hard constraints with mixture-of-experts

https://openreview.net/forum?id=u3dX2CEIZb

Compressor summary: The paper proposes a scalable method to enforce hard physical constraints in neural networks using Mixture-of-Experts (MoE), which improves accuracy, convergence, and efficiency for complex dynamical systems.

The Joint Effect of Task Similarity and Overparameterization on Catastrophic Forgetting - An Analytical Model

https://openreview.net/forum?id=u3dHl287oB

Compressor summary: Our paper studies how overparameterization and task similarity in continual learning affect forgetting and presents an analytical expression for it.

Learning invariant representations of time-homogeneous stochastic dynamical systems

https://openreview.net/forum?id=twSnZwiOIm

Compressor summary: The paper proposes an optimization problem over neural networks to learn a state representation for stochastic dynamical systems, which can be used for forecasting and interpreting system dynamics.

Towards Understanding Sycophancy in Language Models

https://openreview.net/forum?id=tvhaxkMKAn

Compressor summary: RLHF-trained AI assistants often prioritize matching user beliefs over truthfulness, and this behavior is partly influenced by human preferences for sycophancy.

SWAP-NAS: Sample-Wise Activation Patterns for Ultra-fast NAS

https://openreview.net/forum?id=tveiUXU2aa

Compressor summary: SWAP-Score is a novel training-free metric that measures network expressivity based on sample-wise activation patterns, outperforming existing metrics in Neural Architecture Search.

Zero Bubble Pipeline Parallelism

https://openreview.net/forum?id=tuzTN0eIO5

Compressor summary: The paper proposes a scheduling strategy to eliminate pipeline bubbles in distributed training, improving throughput and memory efficiency.

Connect, Collapse, Corrupt: Learning Cross-Modal Tasks with Uni-Modal Data

https://openreview.net/forum?id=ttXg3SKAg5

Compressor summary: The paper proposes a method to improve cross-modal learning from uni-modal data by bridging the modality gap in multi-modal contrastive space using Connect, Collapse, Corrupt steps, achieving state-of-the-art results on various tasks.

SafeDreamer: Safe Reinforcement Learning with World Models

https://openreview.net/forum?id=tsE5HLYtYg

Compressor summary: SafeDreamer is a new algorithm that uses world models to improve safe reinforcement learning, achieving nearly zero-cost performance in various tasks while maintaining safety criteria.

Evaluating Large Language Models at Evaluating Instruction Following

https://openreview.net/forum?id=tr0KidwPLc

Compressor summary: This paper introduces LLMBAR, a benchmark to test how well large language models can evaluate instructions followed by generated text, revealing differences among evaluators and suggesting prompting strategies for improvement.

Quick-Tune: Quickly Learning Which Pretrained Model to Finetune and How

https://openreview.net/forum?id=tqh1zdXIra

Compressor summary: The paper presents a fast and accurate method to select and finetune pretrained models using a meta-dataset of performance data and a meta-learned hyperparameter predictor.

Diffusion Posterior Sampling for Linear Inverse Problem Solving: A Filtering Perspective

https://openreview.net/forum?id=tplXNcHZs1

Compressor summary: The paper introduces a new sampling algorithm for diffusion models that solves linear inverse problems without approximations and shows improved performance on tasks like image inpainting, super-resolution, and deblurring.

Does CLIP’s generalization performance mainly stem from high train-test similarity?

https://openreview.net/forum?id=tnBaiidobu

Compressor summary: The study examines how CLIP's out-of-distribution generalization is affected by training data similarity and pruning, finding that other factors besides data similarity drive CLIP's performance.

Dissecting learning and forgetting in language model finetuning

https://openreview.net/forum?id=tmsqb6WpLz

Compressor summary: Finetuning language models on domain-specific data affects their topic and style preferences more than their factual knowledge, and this study uses instruction-following LLMs to generate controlled text examples to analyze these effects.

Maximum Entropy Heterogeneous-Agent Reinforcement Learning

https://openreview.net/forum?id=tmqOhBC4a5

Compressor summary: The paper proposes a unified framework for learning stochastic policies in cooperative MARL using probabilistic graphical models and maximum entropy objectives, leading to HASAC algorithm that improves sample efficiency, robustness, and exploration.

Motif: Intrinsic Motivation from Artificial Intelligence Feedback

https://openreview.net/forum?id=tmBKIecDE9

Compressor summary: Motif is a method to use a large language model's knowledge to improve an agent's decision-making and performance in complex environments like NetHack game.

Enhancing One-Shot Federated Learning Through Data and Ensemble Co-Boosting

https://openreview.net/forum?id=tm8s3696Ox

Compressor summary: Co-Boosting is a novel framework that improves the quality of synthesized data and ensemble models in One-shot Federated Learning through an adversarial process.

Social Reward: Evaluating and Enhancing Generative AI through Million-User Feedback from an Online Creative Community

https://openreview.net/forum?id=tjn2YZSHUv

Compressor summary: The paper introduces Social Reward, a reward modeling framework that uses implicit feedback from social network users to evaluate and improve text-to-image models, making AI-generated art more aligned with community preferences.

Compositional Preference Models for Aligning LMs

https://openreview.net/forum?id=tiiAzqi6Ol

Compressor summary: CPMs are a new framework for training Preference Models that decompose preference assessments into interpretable features, use a LM to score these features, and aggregate them with logistic regression, improving generalization and robustness over standard PMs.

$\mathbb{D}^2$ Pruning: Message Passing for Balancing Diversity & Difficulty in Data Pruning

https://openreview.net/forum?id=thbtoAkCe9

Compressor summary: The paper proposes a new coreset selection method, $\mathbb{D}^2$ Pruning, that uses graph-based message passing to balance data diversity and difficulty scores, improving model performance on various vision and NLP tasks.

A Benchmark for Learning to Translate a New Language from One Grammar Book

https://openreview.net/forum?id=tbVWug9f2h

Compressor summary: The paper introduces MTOB, a benchmark for testing large language models' ability to learn low-resource languages from a single book of grammar explanations, finding that current models perform similarly to humans but not as well as desired.

Uni-O4: Unifying Online and Offline Deep Reinforcement Learning with Multi-Step On-Policy Optimization

https://openreview.net/forum?id=tbFBh3LMKi

Compressor summary: Uni-O4 is a reinforcement learning method that combines offline and online learning seamlessly for efficient and safe learning in real-world environments.

Learning to Reject for Balanced Error and Beyond

https://openreview.net/forum?id=ta26LtNq2r

Compressor summary: The paper proposes a novel approach for learning to reject classifiers that perform well under imbalanced label distributions and show its effectiveness through experiments on image classification tasks.

GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs

https://openreview.net/forum?id=tVTN7Zs0ml

Compressor summary: GraphCare uses external knowledge graphs and language models to generate patient-specific graphs and improve predictions from electronic health records.

Provable Offline Preference-Based Reinforcement Learning

https://openreview.net/forum?id=tVMPfEGT2w

Compressor summary: The paper proposes an algorithm for offline preference-based reinforcement learning using human feedback and shows a novel guarantee that allows learning any target policy with general function approximation, as long as it's covered by the offline data.

Generative Modeling with Phase Stochastic Bridge

https://openreview.net/forum?id=tUtGjQEDd4

Compressor summary: The paper introduces a novel phase space-based generative model that leverages velocity information for efficient and realistic data generation, outperforming or rivaling existing diffusion models on image generation benchmarks.

Learning Conditional Invariances through Non-Commutativity

https://openreview.net/forum?id=tUVG9nGzgE

Compressor summary: Non-commutative invariance learning improves domain adaptation by using source domain samples more effectively.

Text2Reward: Dense Reward Generation with Language Models for Reinforcement Learning

https://openreview.net/forum?id=tUM39YTRxH

Compressor summary: Text2Reward is a framework that uses large language models to generate dense reward functions for reinforcement learning tasks, achieving good performance on robotic manipulation and locomotion benchmarks.

Towards Robust Out-of-Distribution Generalization Bounds via Sharpness

https://openreview.net/forum?id=tPEwSYPtAC

Compressor summary: This paper connects sharpness and robustness to improve out-of-distribution generalization bounds for models.

Matryoshka Diffusion Models

https://openreview.net/forum?id=tOzCcDdH9O

Compressor summary: Matryoshka Diffusion Model (MDM) is a new framework for generating high-quality images and videos that denoises inputs at multiple resolutions jointly and trains end-to-end using a progressive training schedule, achieving strong zero-shot generalization.

Generalization of Deep ResNets in the Mean-Field Regime

https://openreview.net/forum?id=tMzPZTvz2H

Compressor summary: The paper explores how ResNet generalizes in the limit of infinitely wide and deep neural networks, using a time-variant Gram matrix and analyzing KL divergence.

Reasoning with Latent Diffusion in Offline Reinforcement Learning

https://openreview.net/forum?id=tGQirjzddO

Compressor summary: The paper proposes a novel approach that uses latent diffusion to model trajectory sequences as compressed latent skills for offline reinforcement learning, improving credit assignment and reward propagation.

Partitioning Message Passing for Graph Fraud Detection

https://openreview.net/forum?id=tEgrUrUuwA

Compressor summary: PMP is a new message passing method for Graph Fraud Detection that adapts to different label neighbors by using node-specific aggregation functions, improving GFD performance.

How Realistic Is Your Synthetic Data? Constraining Deep Generative Models for Tabular Data

https://openreview.net/forum?id=tBROYsEz9G

Compressor summary: The paper proposes Constrained Deep Generative Models (C-DGMs) that enforce compliance with background knowledge using a Constraint Layer, improving utility and detection in tabular data generation tasks.

Language Model Inversion

https://openreview.net/forum?id=t9dWHpGkPj

Compressor summary: The paper investigates language model inversion to recover hidden text using next-token probabilities and proposes a method that achieves high BLEU and F1 scores.

Tailoring Self-Rationalizers with Multi-Reward Distillation

https://openreview.net/forum?id=t8eO0CiZJV

Compressor summary: MaRio is a method to improve small language models' ability to generate rationales for question answering by optimizing multiple properties such as plausibility, diversity, and consistency.

Classification with Conceptual Safeguards

https://openreview.net/forum?id=t8cBsT9mcg

Compressor summary: The paper proposes a concept bottleneck model that uses soft concepts and concept confirmation to enhance selective classification in deep learning, improving both coverage and performance under uncertainty.

Controlling Vision-Language Models for Universal Image Restoration

https://openreview.net/forum?id=t3vnnLeajU

Compressor summary: DA-CLIP is a degradation-aware vision-language model that adapts to different types of image corruptions and achieves state-of-the-art performance on image restoration tasks using cross-attention and a controller that predicts both feature embeddings and degradation features.

Frozen Transformers in Language Models Are Effective Visual Encoder Layers

https://openreview.net/forum?id=t0FI3Q66K5

Compressor summary: The paper demonstrates that LLMs can effectively encode visual tasks without language, using a frozen transformer block from pre-trained models as an encoder, and proposes a hypothesis to explain this phenomenon.

BatteryML:An Open-source platform for Machine Learning on Battery Degradation

https://openreview.net/forum?id=sxGugrYhP9

Compressor summary: BatteryML is an open-source platform that simplifies battery life prediction by providing data processing, feature extraction, and various models, while also establishing unified standards for the field.

Real-Fake: Effective Training Data Synthesis Through Distribution Matching

https://openreview.net/forum?id=svIdLLZpsA

Compressor summary: The text discusses how synthetic training data can be improved for deep learning models by analyzing its principles from a distribution-matching perspective, leading to better performance in various image classification tasks and other applications.

VFLAIR: A Research Library and Benchmark for Vertical Federated Learning

https://openreview.net/forum?id=sqRgz88TM3

Compressor summary: The paper introduces VFLAIR, a framework for vertical federated learning that supports various models, datasets, and protocols, and evaluates 11 attacks and 8 defenses to provide insights for deploying VFL securely.

Measuring Vision-Language STEM Skills of Neural Models

https://openreview.net/forum?id=spvaV5LELF

Compressor summary: The text introduces a new challenge dataset for testing neural models' STEM skills, requiring understanding of multimodal vision-language information, and reveals that current models perform poorly compared to elementary students.

Likelihood Training of Cascaded Diffusion Models via Hierarchical Volume-preserving Maps

https://openreview.net/forum?id=sojpn00o8z

Compressor summary: Cascaded models are multi-scale generative models that can also be good likelihood models when using hierarchical volume-preserving maps, leading to improved performance on density estimation, lossless compression, and out-of-distribution detection tasks.

Whole-song Hierarchical Generation of Symbolic Music Using Cascaded Diffusion Models

https://openreview.net/forum?id=sn7CYWyavh

Compressor summary: The paper proposes a novel hierarchical language for generating whole-song pop music with high quality and controllability.

Energy-conserving equivariant GNN for elasticity of lattice architected metamaterials

https://openreview.net/forum?id=smy4DsUbBo

Compressor summary: The paper proposes a higher-order graph neural network model that can quickly and accurately predict the stiffness of lattices, by using symmetries and energy conservation rules.

LRM: Large Reconstruction Model for Single Image to 3D

https://openreview.net/forum?id=sllU8vvsFF

Compressor summary: The Large Reconstruction Model (LRM) uses a transformer-based architecture with many learnable parameters to predict a neural radiance field from a single image and reconstruct 3D objects in high quality and generalize across various inputs.

How connectivity structure shapes rich and lazy learning in neural circuits

https://openreview.net/forum?id=slSmYGc8ee

Compressor summary: This paper explores how the structure of the initial weights in neural networks affects their learning dynamics, finding that high-rank initializations lead to lazier learning while low-rank initializations promote richer learning, with some exceptions related to task and data statistics.

Probabilistic Self-supervised Representation Learning via Scoring Rules Minimization

https://openreview.net/forum?id=skcTCdJz0f

Compressor summary: ProSMin is a probabilistic self-supervised learning method that uses two neural networks to enhance representation quality, mitigate dimensional collapse, and improve performance on various downstream tasks.

Large Language Models Are Not Robust Multiple Choice Selectors

https://openreview.net/forum?id=shr9PXz7T0

Compressor summary: This paper shows that large language models are biased towards specific option IDs in multiple choice questions due to token bias, and proposes a debiasing method called PriDe that separates the prior bias from the prediction distribution.

Alignment as Reward-Guided Search

https://openreview.net/forum?id=shgx0eqdw6

Compressor summary: ARGS is a novel framework that improves language model alignment with human preferences by adjusting probabilistic predictions using a reward signal during decoding, resulting in better average rewards and semantic diversity.

Let Models Speak Ciphers: Multiagent Debate through Embeddings

https://openreview.net/forum?id=sehRvaIPQQ

Compressor summary: CIPIR is a communication protocol that uses raw transformer output embeddings instead of natural language to enhance reasoning ability in large language models, outperforming existing methods by 1-3.5%.

NeuroBack: Improving CDCL SAT Solving using Graph Neural Networks

https://openreview.net/forum?id=samyfu6G93

Compressor summary: NeuroBack is a novel approach that uses Graph Neural Networks to predict phases of variables in the SAT problem, enabling faster and more efficient solving on CPU without relying on GPU resources.

DSPy: Compiling Declarative Language Model Calls into State-of-the-Art Pipelines

https://openreview.net/forum?id=sY5N0zY5Od

Compressor summary: DSPy is a programming model for developing and optimizing language models pipelines using text transformation graphs, enabling sophisticated tasks and outperforming hard-coded prompt templates and few-shot prompting.

Understanding when Dynamics-Invariant Data Augmentations Benefit Model-free Reinforcement Learning Updates

https://openreview.net/forum?id=sVEu295o70

Compressor summary: The paper investigates how state-action coverage, reward density, and the number of augmented transitions per update affect data efficiency in reinforcement learning tasks with sparse rewards and dynamics-invisible data augmentation.

Evaluating Representation Learning on the Protein Structure Universe

https://openreview.net/forum?id=sTYuRVrdK3

Compressor summary: The paper introduces a benchmark suite for evaluating protein structure representation learning methods, comparing various pre-training and downstream tasks to improve understanding of protein structures and functions.

Divide and not forget: Ensemble of selectively trained experts in Continual Learning

https://openreview.net/forum?id=sSyytcewxe

Compressor summary: SEED is a novel approach for class-incremental learning that selects and fine-tunes the most optimal expert for each task using Gaussian distributions, improving diversity and stability.

Adapting to Distribution Shift by Visual Domain Prompt Generation

https://openreview.net/forum?id=sSaN4gxuEf

Compressor summary: The paper proposes a method to adapt a model using few unlabeled data, by leveraging pre-trained backbones, source domains, and a domain prompt generator that guides the features towards the target domain.

Consistent4D: Consistent 360° Dynamic Object Generation from Monocular Video

https://openreview.net/forum?id=sPUrdFGepF

Compressor summary: Consistent4D is a method to generate 4D dynamic objects from monocular videos without multi-view data or camera calibration, using a cascade DyNeRF and interpolation-driven consistency loss.

Fiber Monte Carlo

https://openreview.net/forum?id=sP1tCl2QBk

Compressor summary: The paper proposes a differentiable Monte Carlo estimator that samples line segments instead of points for integrals with discontinuous integrands, enabling gradient-based optimization in applications like robotics and graphics.

NeRM: Learning Neural Representations for High-Framerate Human Motion Synthesis

https://openreview.net/forum?id=sOJriBlOFd

Compressor summary: NeRM is a generative model that uses implicit neural representations to synthesize realistic human motions with any framerate, while being memory-efficient and fast.

Neural Common Neighbor with Completion for Link Prediction

https://openreview.net/forum?id=sNFLN3itAd

Compressor summary: The text introduces a novel link prediction model, NCNC, that leverages structural features and graph completion to overcome graph incompleteness, outperforming existing models.

Towards Meta-Pruning via Optimal Transport

https://openreview.net/forum?id=sMoifbuxjB

Compressor summary: Intra-Fusion combines pruning and fusion methods in neural networks to achieve better compression, accuracy, and training time than standard pruning alone.

On the Analysis of GAN-based Image-to-Image Translation with Gaussian Noise Injection

https://openreview.net/forum?id=sLregLuXpn

Compressor summary: The paper provides a robust theoretical framework to understand and optimize the role of Gaussian noise injection in image-to-image translation models for better performance in noisy environments.

Transformers can optimally learn regression mixture models

https://openreview.net/forum?id=sLkj91HIZU

Compressor summary: This paper explores if transformers can learn to make optimal predictions for mixtures of regressions and shows they can achieve low error, sample efficiency, and robustness.

NuwaDynamics: Discovering and Updating in Causal Spatio-Temporal Modeling

https://openreview.net/forum?id=sLdVl0q68X

Compressor summary: NuwaDynamics is a causal concept for spatio-temporal prediction that identifies important regions in data and enables models to reason causally, improving results on real-world tasks.

Fair and Efficient Contribution Valuation for Vertical Federated Learning

https://openreview.net/forum?id=sLQb8q0sUi

Compressor summary: Vertical federated Shapley value (VerFedSV) is a fair and efficient contribution valuation metric for data sources in vertical federated learning scenarios.

Addressing Catastrophic Forgetting and Loss of Plasticity in Neural Networks

https://openreview.net/forum?id=sKPzAXoylB

Compressor summary: UPGD is a novel approach for continual learning of representations that combats catastrophic forgetting and loss of plasticity by applying different size modifications to useful and unuseful units.

ViDA: Homeostatic Visual Domain Adapter for Continual Test Time Adaptation

https://openreview.net/forum?id=sJ88Wg5Bp5

Compressor summary: The paper proposes ViDA, a visual domain adapter for continual test-time adaptation, which uses high-rank and low-rank features to handle domain-specific and domain-shared knowledge and achieves state-of-the-art performance in classification and segmentation tasks.

Branch-GAN: Improving Text Generation with (not so) Large Language Models

https://openreview.net/forum?id=sHEJJmzBIN

Compressor summary: The authors propose a computationally efficient GAN approach for open domain text generation using Transformer models, which improves text quality and avoids sparsity problems, outperforming larger baseline models with fewer parameters.

Memory-Assisted Sub-Prototype Mining for Universal Domain Adaptation

https://openreview.net/forum?id=sGVmr7KHfn

Compressor summary: MemSPM is a novel method that mines sub-classes within categories to reduce feature gap and improve adaptation performance when there is significant concept shift between samples.

Learning Interactive Real-World Simulators

https://openreview.net/forum?id=sFyTZEqmUY

Compressor summary: The authors propose UniSim, a universal simulator for real-world interaction, which can be trained on diverse datasets and enable various intelligent applications.

MagicDrive: Street View Generation with Diverse 3D Geometry Control

https://openreview.net/forum?id=sBQwvucduK

Compressor summary: MagicDrive is a framework for generating street views with precise 3D control and high-fidelity synthesis, enabling better 3D perception tasks like object detection.

SocioDojo: Building Lifelong Analytical Agents with Real-world Text and Time Series

https://openreview.net/forum?id=s9z0HzWJJp

Compressor summary: SocioDojo is a lifelong learning environment for creating autonomous agents that can analyze and make decisions on societal topics using information sources, knowledge bases, time series, and analysis tools.

f-FERM: A Scalable Framework for Robust Fair Empirical Risk Minimization

https://openreview.net/forum?id=s90VIdza2K

Compressor summary: The paper proposes a stochastic optimization framework for fair machine learning that works with small data batches, has theoretical convergence guarantees, and performs well under distribution shifts.

Towards Eliminating Hard Label Constraints in Gradient Inversion Attacks

https://openreview.net/forum?id=s8cMuxI5gu

Compressor summary: This paper proposes an algorithm to recover augmented soft labels and input features from gradients in federated learning, considering realistic label smoothing and mixup techniques, and shows its effectiveness in attacking the system.

The Human-AI Substitution game: active learning from a strategic labeler

https://openreview.net/forum?id=s5hSp7EdL3

Compressor summary: The paper studies how learning from a strategic labeler who selectively abstains from labeling can affect the learning process and proposes a near-optimal algorithm for this scenario.

BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input Detection

https://openreview.net/forum?id=s56xikpD92

Compressor summary: The paper proposes a defense method called BaDExpert that finetunes a backdoored DNN to recognize only its malicious inputs, thus detecting and filtering out the backdoors.

Accurate Retraining-free Pruning for Pretrained Encoder-based Language Models

https://openreview.net/forum?id=s2NjWfaYdZ

Compressor summary: KPrune is a retraining-free algorithm that accurately compresses pretrained encoder-based language models by preserving useful knowledge and improving F1 score on the SQuAD benchmark.

Neural SDF Flow for 3D Reconstruction of Dynamic Scenes

https://openreview.net/forum?id=rzF0R6GOd4

Compressor summary: The paper proposes a method for 3D reconstruction of dynamic scenes from multi-view videos using Signed Distance Function flow, which handles topology changes and outperforms existing methods.

CAMIL: Context-Aware Multiple Instance Learning for Cancer Detection and Subtyping in Whole Slide Images

https://openreview.net/forum?id=rzBskAEmoc

Compressor summary: CAMIL is a new model for cancer diagnosis from tissue biopsy images that considers neighboring information and achieves better performance than existing methods.

Improving protein optimization with smoothed fitness landscapes

https://openreview.net/forum?id=rxlF2Zv8x0

Compressor summary: The paper proposes a method, Gibbs sampling with Graph-based Smoothing (GGS), to smooth the fitness landscape and improve protein optimization using graph signals and energy-based models.

Selective Mixup Fine-Tuning for Optimizing Non-Decomposable Metrics

https://openreview.net/forum?id=rxVBKhyfSo

Compressor summary: SelMix is a technique to optimize pre-trained models for practical performance objectives by mixing features of samples from specific classes.

Learning to Act without Actions

https://openreview.net/forum?id=rvUq3cxpDF

Compressor summary: LAPO is a method to train deep RL policies from action-free videos by inferring latent actions, which enables web data pre-training and fine-tuning.

Long-Short-Range Message-Passing: A Fragmentation-Based Framework to Capture Non-Local Atomistic Interactions

https://openreview.net/forum?id=rvDQtdMnOl

Compressor summary: The paper proposes a new graph neural network method (LSR-MP) that improves the simulation of chemical and biological systems using ab initio molecular dynamics, achieving state-of-the-art results with reduced error.

Conformal Prediction via Regression-as-Classification

https://openreview.net/forum?id=rulxyXjf46

Compressor summary: The paper proposes a method to use conformal prediction for regression by converting it to a classification problem with a new loss function and modified techniques.

SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore

https://openreview.net/forum?id=ruk0nyQPec

Compressor summary: SILO is a language model that uses a combination of a parametric model trained on public domain text and a nonparametric datastore with customizable data attribution to enable compliant use of copyrighted or restricted data during inference.

A Data-Driven Measure of Relative Uncertainty for Misclassification Detection

https://openreview.net/forum?id=ruGY8v10mK

Compressor summary: The paper proposes a new data-driven measure of relative uncertainty for detecting misclassified instances in machine learning models based on soft-predictions and shows its effectiveness in image classification tasks.

Variational Inference for SDEs Driven by Fractional Noise

https://openreview.net/forum?id=rtx8B94JMS

Compressor summary: The paper presents a new method for inferring parameters in neural stochastic differential equations using fractional Brownian motion and variational methods, which can capture long-term dependencies and be applied to video prediction.

Understanding the Robustness of Multi-modal Contrastive Learning to Distribution Shift

https://openreview.net/forum?id=rtl4XnJYBh

Compressor summary: This paper analyzes two mechanisms behind multimodal contrastive learning's success in robust representation learning, and shows how rich captions help improve zero-shot classification under distribution shift.

Federated Wasserstein Distance

https://openreview.net/forum?id=rsg1mvUahT

Compressor summary: FedWad is a principled algorithm that computes Wasserstein distance between two distributions in a federated manner by manipulating and exchanging geodesics, with convergence guarantees and empirical benefits.

DistillSpec: Improving Speculative Decoding via Knowledge Distillation

https://openreview.net/forum?id=rsY6J3ZaTF

Compressor summary: DistillSpec is a method that uses knowledge distillation to improve the alignment between a fast draft and a slow target language model for speculative decoding, achieving significant speedups and latency control.

An Efficient Membership Inference Attack for the Diffusion Model by Proximal Initialization

https://openreview.net/forum?id=rpH9FcCEV6

Compressor summary: The paper proposes a query-based membership inference attack on diffusion models for image and audio generation tasks and analyzes their privacy robustness in the text-to-speech task.

Faithful Vision-Language Interpretation via Concept Bottleneck Models

https://openreview.net/forum?id=rp0EdI8X4e

Compressor summary: The authors propose a new concept called Faithful Vision-Language Concept (FVLC) models that improve interpretable machine learning by addressing the instability issue of label-free bottleneck models and maintaining good performance.

Stylized Offline Reinforcement Learning: Extracting Diverse High-Quality Behaviors from Heterogeneous Datasets

https://openreview.net/forum?id=rnHNDihrIT

Compressor summary: SORL is a novel method for offline RL that leverages policy diversity by alternating between policy learning and trajectory clustering.

Intriguing Properties of Generative Classifiers

https://openreview.net/forum?id=rmg0qMKYRQ

Compressor summary: The study compares text-to-image classifiers based on generative models with discriminative models and human perception, finding that generative models have impressive properties similar to human object recognition.

On the Provable Advantage of Unsupervised Pretraining

https://openreview.net/forum?id=rmXXKxQpOR

Compressor summary: This paper provides a theoretical understanding of why unsupervised pretraining helps in machine learning systems by studying a generic framework with latent variable models and empirical risk minimization.

Gen-Z: Generative Zero-Shot Text Classification with Contextualized Label Descriptions

https://openreview.net/forum?id=rkplYfqUr0

Compressor summary: Gen-Z is a generative framework for zero-shot text classification that uses natural language labels and improves robustness to prompt variations and allows personalization.

A Branching Decoder for Set Generation

https://openreview.net/forum?id=riNuqYiD66

Compressor summary: The paper proposes a branching decoder for generating sets of text without order dependence, which improves effectiveness and efficiency over the existing sequential decoder.

TabR: Tabular Deep Learning Meets Nearest Neighbors in 2023

https://openreview.net/forum?id=rhgIgTSSxW

Compressor summary: TabR is a new tabular deep learning model that uses a novel attention-like mechanism to retrieve and utilize nearest neighbors for better prediction, outperforming existing methods on several benchmarks.

Does Progress On Object Recognition Benchmarks Improve Generalization on Crowdsourced, Global Data?

https://openreview.net/forum?id=rhaQbS3K3R

Compressor summary: The authors study how well object recognition models perform on crowdsourced global datasets and find that progress on these datasets lags behind standard benchmarks and may increase geographic disparities in model performance.

Image Inpainting via Iteratively Decoupled Probabilistic Modeling

https://openreview.net/forum?id=rUf9G9k2im

Compressor summary: The pixel spread model combines GANs' optimization efficiency with probabilistic models' prediction tractability to improve image inpainting quality and efficiency, achieving state-of-the-art results.

Octavius: Mitigating Task Interference in MLLMs via MoE

https://openreview.net/forum?id=rTDyN8yajn

Compressor summary: The paper introduces Octavius, a framework for studying multimodal learning with large language models, which combines MoE and LoRA to reduce interference and improve performance on various tasks.

Towards Lossless Dataset Distillation via Difficulty-Aligned Trajectory Matching

https://openreview.net/forum?id=rTBL8OhdhH

Compressor summary: The authors propose a new algorithm for Dataset Distillation that matches the difficulty of the generated patterns with the size of the synthetic dataset, enabling lossless distillation for larger datasets than previous methods.

Functional Interpolation for Relative Positions improves Long Context Transformers

https://openreview.net/forum?id=rR03qFesqk

Compressor summary: FIRE is a new position encoding method for Transformers that improves their ability to handle longer inputs and generalizes better to different tasks.

Like Oil and Water: Group Robustness Methods and Poisoning Defenses Don't Mix

https://openreview.net/forum?id=rM9VJPB20F

Compressor summary: The paper reveals that existing group robustness methods in machine learning unintentionally boost adversarial success by amplifying poison samples, and current poisoning defenses damage group robustness by eliminating minority samples along with poisons.

Masked Structural Growth for 2x Faster Language Model Pre-training

https://openreview.net/forum?id=rL7xsg1aRn

Compressor summary: The paper introduces MSG, a method to accelerate pre-training of large language models by growing the model in an optimal way with function-preserving operators that work regardless of weight initialization.

Understanding Addition in Transformers

https://openreview.net/forum?id=rIx1YXVWZb

Compressor summary: The paper analyzes a simple Transformer model for integer addition, explaining its algorithms, execution, and rare failures, to help understand and ensure safe and ethical AI use.

Curriculum reinforcement learning for quantum architecture search under hardware errors

https://openreview.net/forum?id=rINBD8jPoP

Compressor summary: This text describes a new quantum architecture search algorithm (CRLQAS) that improves simulation speed, explores the search space efficiently, and optimizes for shorter circuits in noisy environments.

Advancing Pose-Guided Image Synthesis with Progressive Conditional Diffusion Models

https://openreview.net/forum?id=rHzapPnCgT

Compressor summary: This paper introduces Progressive Conditional Diffusion Models (PCDMs), which incrementally synthesize realistic person images with different poses by using three stages of conditional diffusion models.

Progressive Fourier Neural Representation for Sequential Video Compilation

https://openreview.net/forum?id=rGFrRMBbOq

Compressor summary: PFNR is a novel method to encode multiple complex videos using adaptive sub-modules in Fourier space, enabling efficient and scalable continual learning.

Variance-aware Regret Bounds for Stochastic Contextual Dueling Bandits

https://openreview.net/forum?id=rDH7dIFn20

Compressor summary: The paper proposes a new algorithm for contextual dueling bandits that accounts for uncertainty in pairwise comparisons and achieves a better regret bound than existing methods.

Random Sparse Lifts: Construction, Analysis and Convergence of finite sparse networks

https://openreview.net/forum?id=rBH7x87VfJ

Compressor summary: The authors propose a framework to train overparameterized neural networks with gradient flow, showing that this method can achieve arbitrarily low loss and generalize common deep learning models by sparsifying their weight tensors.

S$2$AC: Energy-Based Reinforcement Learning with Stein Soft Actor Critic

https://openreview.net/forum?id=rAHcTCMaLc

Compressor summary: S$^2$AC is a MaxEnt RL algorithm using SVGD that computes entropy efficiently for EBM policies, achieving better performance than SQL and SAC in various tasks.

Object centric architectures enable efficient causal representation learning

https://openreview.net/forum?id=r9FsiXZxZt

Compressor summary: The paper proposes an object-centric architecture that uses weak supervision from sparse perturbations to disentangle latent variables of multiple objects, addressing the limitations of existing causal representation learning methods.

UniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition

https://openreview.net/forum?id=r65xfUb76p

Compressor summary: The paper presents a method to train smaller and more efficient named entity recognition (NER) models from large language models (LLMs) by using mission-focused instruction tuning, achieving state-of-the-art results without direct supervision.

Linear Convergence Bounds for Diffusion Models via Stochastic Localization

https://openreview.net/forum?id=r5njV3BsuD

Compressor summary: We present new linear convergence bounds for denoising diffusion models with finite second moment assumptions and show how to extend Girsanov-based methods and use stochastic localization theory in the proof.

Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation

https://openreview.net/forum?id=r42tSSCHPh

Compressor summary: The text describes a new attack method that exploits language models' generation methods to cause malicious behavior, highlighting the need for better alignment and safety evaluation of open-source large language models.

Effectively Leveraging Capacity for Improved Deterministic Robustness Certification

https://openreview.net/forum?id=qz3mcn99cu

Compressor summary: The text describes a study that improves robustness of neural networks against adversarial attacks using Lipschitz-based methods and novel techniques, achieving significant performance gains.

Robust Training of Federated Models with Extremely Label Deficiency

https://openreview.net/forum?id=qxLVaYbsSI

Compressor summary: Twinsight is a novel FSSL method that trains two models simultaneously with different objectives to reduce gradient conflicts and improve performance.

Canonpipe: Data Debugging with Shapley Importance over Machine Learning Pipelines

https://openreview.net/forum?id=qxGXjWxabq

Compressor summary: The paper presents Canonpipe, a method for computing Shapley-based data importance over ML pipelines, which can efficiently discover data errors and sometimes perform better than existing methods.

From Graphs to Hypergraphs: Hypergraph Projection and its Remediation

https://openreview.net/forum?id=qwYKE3VB2h

Compressor summary: This paper studies how using graphs instead of hypergraphs to model interconnected systems can lose higher-order information, and proposes a learning method to recover this lost information from real-world datasets.

Behaviour Distillation

https://openreview.net/forum?id=qup9xD8mW4

Compressor summary: The paper introduces behaviour distillation and HaDES, a method to create small synthetic datasets for reinforcement learning tasks without expert data.

Towards domain-invariant Self-Supervised Learning with Batch Styles Standardization

https://openreview.net/forum?id=qtE9K23ISq

Compressor summary: Batch Styles Standardization (BSS) is a Fourier-based method that standardizes image styles within a batch, enhancing domain-invariance in self-supervised learning models without requiring domain labels or specific architectures, and improving their performance on unseen domains.

On the Expressivity of Objective-Specification Formalisms in Reinforcement Learning

https://openreview.net/forum?id=qr4ECbGcSj

Compressor summary: The text compares 17 reinforcement learning objective-specification formalisms, revealing their expressivity limitations and implications for policy optimisation and reward learning.

Demystifying Embedding Spaces using Large Language Models

https://openreview.net/forum?id=qoYogklIPz

Compressor summary: The paper proposes a method to make embeddings more interpretable by using large language models to generate narratives from them, enabling querying and exploration of complex data in various tasks.

Grounding Language Plans in Demonstrations Through Counter-Factual Perturbations

https://openreview.net/forum?id=qoHeuRAcSl

Compressor summary: This paper proposes a method to use large language models to guide learning structures and constraints in robot manipulation tasks, using mode families as an intermediate layer between high-level language and low-level physical trajectories.

Teach LLMs to Phish: Stealing Private Information from Language Models

https://openreview.net/forum?id=qo21ZlfNu6

Compressor summary: The authors propose a new data extraction attack called "neural phishing" that exploits large language models trained on private data to extract sensitive information with high success rates.

Fast Imitation via Behavior Foundation Models

https://openreview.net/forum?id=qnWtw3l0jb

Compressor summary: Imitation learning using RL foundation models can quickly mimic expert behavior with few demonstrations, outperforming standard approaches in speed and efficiency.

Learning Stackable and Skippable LEGO Bricks for Efficient, Reconfigurable, and Variable-Resolution Diffusion Modeling

https://openreview.net/forum?id=qmXedvwrT1

Compressor summary: LEGO bricks are a new adaptable network backbone for iterative refinement in diffusion models, enabling efficient generation of higher-resolution images with variable resolutions and faster convergence.

Skill Machines: Temporal Logic Skill Composition in Reinforcement Learning

https://openreview.net/forum?id=qiduMcw3CU

Compressor summary: The text proposes a framework for agents to learn skill primitives and compose them flexibly to achieve temporal logic specifications in various environments.

Learning with Language-Guided State Abstractions

https://openreview.net/forum?id=qi5Xa2cOZg

Compressor summary: LGA is a method that uses natural language to automatically create state abstractions for imitation learning, improving policy generalization and robustness.

A Newborn Embodied Turing Test for Comparing Object Segmentation Across Animals and Machines

https://openreview.net/forum?id=qhkEOCcVX9

Compressor summary: The study introduces a Newborn Embodied Turing Test (NETT) for object segmentation, comparing the learning abilities of newborn chicks and machine-learning algorithms in recognizing objects across different backgrounds and viewpoints.

Fast Equilibrium of SGD in Generic Situations

https://openreview.net/forum?id=qgWJkDiI5p

Compressor summary: This paper investigates the Fast Equilibrium conjecture for normalization layers in deep learning and proves it under more general assumptions than previous works.

Closing the Gap between TD Learning and Supervised Learning - A Generalisation Point of View.

https://openreview.net/forum?id=qg5JENs0N4

Compressor summary: The paper studies how reinforcement learning algorithms can use past experience to solve new tasks without explicit stitching mechanisms, showing that this property depends on a form of generalization different from i.i.d. generalization, and proposes data augmentation as a way to improve it.

Neural Probabilistic Protein-Protein Docking via a Differentiable Energy Model

https://openreview.net/forum?id=qg2boc2AwU

Compressor summary: The paper presents EBMDock, a geometric deep learning framework that uses statistical potential and Langevin dynamics to sample protein-protein docking poses from a probability distribution, achieving superior results on two benchmarks.

Diverse Projection Ensembles for Distributional Reinforcement Learning

https://openreview.net/forum?id=qe49ybvvPs

Compressor summary: Distributional RL algorithms learn return distributions instead of their expected values, using diverse projections and representations to improve uncertainty estimation and exploration.

Transformer-Modulated Diffusion Models for Probabilistic Multivariate Time Series Forecasting

https://openreview.net/forum?id=qae04YACHs

Compressor summary: The Transformer-Modulated Diffusion Model (TMDM) combines transformers and conditional diffusion to predict distribution for multivariate time series with uncertainty.

Learning Thresholds with Latent Values and Censored Feedback

https://openreview.net/forum?id=qaKRfobbTg

Compressor summary: The paper studies how to learn a threshold in latent space for optimizing rewards in various scenarios, and provides query complexity and regret bounds for this problem.

Counting Graph Substructures with Graph Neural Networks

https://openreview.net/forum?id=qaJxPhkYtD

Compressor summary: This paper studies how Graph Neural Networks (GNNs) can learn to count substructures like cycles and connected components in graphs and designs a new GNN architecture that performs well in several tasks.

Large Language Models as Tool Makers

https://openreview.net/forum?id=qV83K9d5WB

Compressor summary: The LLMs A s Tool Makers (LATM) framework allows large language models to create and reuse their own tools for problem-solving, improving efficiency and reducing serving costs.

Relay Diffusion: Unifying diffusion process across resolutions for image synthesis

https://openreview.net/forum?id=qTlcbLSm4p

Compressor summary: The Relay Diffusion Model (RDM) improves high-resolution image synthesis by transferring low-resolution images or noise into equivalent high-resolution ones via blurring diffusion and block noise, achieving state-of-the-art results on CelebA-HQ and ImageNet 256 times 256.

NP-GL: Extending Power of Nature from Binary Problems to Real-World Graph Learning

https://openreview.net/forum?id=qT7DXUmX7j

Compressor summary: The paper introduces a novel framework for nature-powered graph learning that solves real-valued problems faster and more efficiently than neural networks, while maintaining similar accuracy.

Robustifying and Boosting Training-Free Neural Architecture Search

https://openreview.net/forum?id=qPloNoDJZn

Compressor summary: The RoBoT algorithm combines and optimizes existing training-free metrics using Bayesian optimization, improving search performance for designing deep neural networks without training.

Guess & Sketch: Language Model Guided Transpilation

https://openreview.net/forum?id=qPFsIbF3V6

Compressor summary: The authors propose a neurosymbolic approach to automatic translation of assembly code programs, which combine the strengths of learned probabilistic models and symbolic solvers to achieve better performance than existing methods.

Masks, Signs, And Learning Rate Rewinding

https://openreview.net/forum?id=qODvxQ8TXW

Compressor summary: LRR outperforms IMP in finding lottery tickets by effectively combining mask learning and parameter optimization in overparameterized networks, allowing it to handle different mask signatures.

FairSeg: A Large-scale Medical Image Segmentation Dataset for Fairness Learning with Fair Error-Bound Scaling

https://openreview.net/forum?id=qNrJJZAKI3

Compressor summary: The paper introduces FairSeg, a new medical segmentation fairness dataset, and proposes a fair error-bound scaling approach to improve equity in AI models.

Zero and Few-shot Semantic Parsing with Ambiguous Inputs

https://openreview.net/forum?id=qL9gogRepu

Compressor summary: AmP introduces a framework, dataset, and challenge to study how text-to-code systems handle linguistic ambiguities, finding that large models perform poorly without explicit instruction but capture the distribution of meanings when ambiguity is present.

SAFLEX: Self-Adaptive Augmentation via Feature Label Extrapolation

https://openreview.net/forum?id=qL6brrBDk2

Compressor summary: SAFLEX is an efficient method that improves data augmentation by learning sample weights and soft labels using bilevel optimization, enhancing model performance across diverse datasets and tasks with minimal computational cost.

Learning Grounded Action Abstractions from Language

https://openreview.net/forum?id=qJ0Cfj4Ex9

Compressor summary: The system learns symbolic action abstractions from language and uses them for long-horizon planning, improving accuracy and generalization.

Seer: Language Instructed Video Prediction with Latent Diffusion Models

https://openreview.net/forum?id=qHGgNyQk31

Compressor summary: Seer is a novel model for robots to predict video frames from text inputs, which uses efficient techniques and achieves better results than CogVideo.

Protein-Ligand Interaction Prior for Binding-aware 3D Molecule Diffusion Models

https://openreview.net/forum?id=qH9nrMNTIW

Compressor summary: The paper proposes a new model, IPDiff, for generating 3D ligand molecules that bind to specific proteins by incorporating protein-ligand interactions into both diffusion and sampling processes, leading to more realistic structures and better binding affinities.

Conditional Instrumental Variable Regression with Representation Learning for Causal Inference

https://openreview.net/forum?id=qDhq1icpO8

Compressor summary: The paper proposes a new method, CBRL.CIV, for estimating causal effects from observational data without the linearity assumption and using a conditional instrument to relax the strict condition of unconfounded instruments.

Explaining Time Series via Contrastive and Locally Sparse Perturbations

https://openreview.net/forum?id=qDdSRaOiyb

Compressor summary: ContraLSP is a model that uses counterfactual samples and sparse gates to create informative but uninformative perturbations for multivariate time series, improving explanation quality.

VONET: ADVANCING UNSUPERVISED VIDEO OBJECT LEARNING

https://openreview.net/forum?id=qCyhvr0GG8

Compressor summary: VONet is a new method that uses attention and sequential VAE to learn video objects without supervision, achieving state-of-the-art results on various datasets.

LayoutNUWA: Revealing the Hidden Layout Expertise of Large Language Models

https://openreview.net/forum?id=qCUWVT0Ayy

Compressor summary: This paper introduces LayoutNUWA, a model that generates graphic layouts by treating layout generation as a code generation task using large language models, achieving significant performance improvements.

DeepZero: Scaling Up Zeroth-Order Optimization for Deep Model Training

https://openreview.net/forum?id=qBWhjsNPEY

Compressor summary: DeepZero is a novel framework that enables zeroth-order optimization for training deep neural networks, achieving near-optimal performance and improving certified adversarial defense and partial differential equation error correction.

Boosting of Thoughts: Trial-and-Error Problem Solving with Large Language Models

https://openreview.net/forum?id=qBL04XXex6

Compressor summary: Boosting of Thoughts (BoT) is an automated framework for improving problem-solving with LLMs by iteratively exploring, self-evaluating, and revising reasoning steps using error analysis.

DQ-LoRe: Dual Queries with Low Rank Approximation Re-ranking for In-Context Learning

https://openreview.net/forum?id=qAoxvePSlq

Compressor summary: The study introduces DQ-LoRe, a framework that uses dual queries and low-rank approximation re-ranking to automatically select exemplars for in-context learning in natural language processing tasks.

Efficient Integrators for Diffusion Generative Models

https://openreview.net/forum?id=qA4foxO5Gf

Compressor summary: The text introduces two frameworks to accelerate sample generation in diffusion models and presents a hybrid method that outperforms baselines on CIFAR-10.

Leveraging Hyperbolic Embeddings for Coarse-to-Fine Robot Design

https://openreview.net/forum?id=q9jQPA6zPK

Compressor summary: The paper proposes a novel coarse-to-fine method for designing multi-cellular robots using Hyperbolic Embeddings for Robot Design (HERD) framework, which improves efficiency and generalization compared to previous approaches.

Large-Vocabulary 3D Diffusion Model with Transformer

https://openreview.net/forum?id=q57JLSE2j5

Compressor summary: The paper introduces a novel diffusion-based generative model that can synthesize diverse and high-quality 3D objects from over 200 real-world categories using a triplane representation and a 3D-aware transformer.

An Investigation of Representation and Allocation Harms in Contrastive Learning

https://openreview.net/forum?id=q4SiDyYQbo

Compressor summary: The paper shows that contrastive learning, a type of self-supervised learning, can cause minority groups' representations to merge with majority groups', leading to reduced performance and bias.

Solving High Frequency and Multi-Scale PDEs with Gaussian Processes

https://openreview.net/forum?id=q4AEBLHuA6

Compressor summary: The paper proposes a new method for solving high-frequency and multi-scale PDEs using Gaussian processes and spectral mixture kernels, which can improve efficiency and scalability.

Adversarial Attacks on Fairness of Graph Neural Networks

https://openreview.net/forum?id=q3KNrmW6Ql

Compressor summary: The paper presents G-FairAttack, a framework that can adversarially compromise the fairness of various fairness-aware GNNs with low utility loss and fast computation.

Universal Guidance for Diffusion Models

https://openreview.net/forum?id=pzpWBbnwiJ

Compressor summary: The proposed universal guidance algorithm allows diffusion models to accept different types of guidance without retraining, enabling various image generation applications.

Conformal Language Modeling

https://openreview.net/forum?id=pzUhfQ74c5

Compressor summary: The paper proposes a novel conformal prediction approach for language models with performance guarantees, calibrating stopping and rejection rules to reduce noise and ensure coverage of acceptable responses.

How I Warped Your Noise: a Temporally-Correlated Noise Prior for Diffusion Models

https://openreview.net/forum?id=pzElnMrgSD

Compressor summary: The authors propose a novel method to preserve temporal correlations in video editing and generation using integral noise representation and a tailored transport method.

Quantifying the Sensitivity of Inverse Reinforcement Learning to Misspecification

https://openreview.net/forum?id=pz2E1Q9Wni

Compressor summary: The paper investigates how different behavioural models affect inverse reinforcement learning (IRL) and shows that even small misspecifications can cause significant errors in inferring an agent's preferences.

Effective and Efficient Federated Tree Learning on Hybrid Data

https://openreview.net/forum?id=py4ZV2qYQI

Compressor summary: HybridTree is a novel federated learning approach for hybrid data settings that incorporates party knowledge into tree lower layers, achieving comparable accuracy and significant speedup.

ODE Discovery for Longitudinal Heterogeneous Treatment Effects Inference

https://openreview.net/forum?id=pxI5IPeWgW

Compressor summary: The paper proposes a novel approach for inferring unbiased treatment effects using closed-form ordinary differential equations, which offer advantages like interpretability and irregular sampling.

InterpGNN: Understand and Improve Generalization Ability of Transdutive GNNs through the Lens of Interplay between Train and Test Nodes

https://openreview.net/forum?id=pwW807WJ9G

Compressor summary: This paper proposes a PAC-Bayesian bound for GNNs that incorporates the $L$-hop interplay between training and testing nodes, and introduces a Graph Global Workspace module to enhance the interplay using key-value attention.

CIFAR-10-Warehouse: Broad and More Realistic Testbeds in Model Generalization Analysis

https://openreview.net/forum?id=pw2ssoOTpo

Compressor summary: CIFAR-10-Warehouse is a testbed for evaluating model generalization in various out-of-distribution environments, consisting of 180 datasets collected by image search engines and diffusion models with diverse characteristics.

$\alpha$TC-VAE: On the relationship between Disentanglement and Diversity

https://openreview.net/forum?id=ptXo0epLQo

Compressor summary: The authors propose alpha-TCVAE, a variational autoencoder that maximizes disentanglement and informativeness of latent variables using a novel total correlation lower bound, and show its benefits in representation learning and downstream tasks.

Image Clustering via the Principle of Rate Reduction in the Age of Pretrained Models

https://openreview.net/forum?id=ptCIlV24YZ

Compressor summary: The paper presents a new image clustering method based on pre-trained models that improves accuracy, estimates cluster numbers, and labels clusters with text.

#InsTag: Instruction Tagging for Analyzing Supervised Fine-tuning of Large Language Models

https://openreview.net/forum?id=pszewhybU9

Compressor summary: InsTag is a method that analyzes and improves instruction diversity and complexity for supervised fine-tuning (SFT) of large language models (LLMs), leading to better instruction-following abilities.

AmortizedPeriod: Attention-based Amortized Inference for Periodicity Identification

https://openreview.net/forum?id=psEswR8Jz4

Compressor summary: AmortizedPeriod is a novel method that uses Bayesian statistics and deep learning to identify periodic patterns in time series data, improving robustness and efficiency over existing methods.

Tractable MCMC for Private Learning with Pure and Gaussian Differential Privacy

https://openreview.net/forum?id=pmweVpJ229

Compressor summary: ASAP is an algorithm that samples from the posterior distribution while preserving differential privacy by adding noise proportional to the distance of each sample from a reference distribution.

Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models

https://openreview.net/forum?id=plmBsXHxgR

Compressor summary: The authors propose new cross-modality jailbreak attacks on vision language models using adversarial images and textual prompts, which exploit the alignment vulnerability of multi-modal models.

Kernel Metric Learning for In-Sample Off-Policy Evaluation of Deterministic RL Policies

https://openreview.net/forum?id=plebgsdiiV

Compressor summary: The paper proposes a method to evaluate policies in reinforcement learning with continuous actions by relaxing deterministic target policies using a kernel and optimizing its metrics to reduce bias and variance in off-policy evaluation.

Probabilistic Adaptation of Black-Box Text-to-Video Models

https://openreview.net/forum?id=pjtIEgscE3

Compressor summary: The paper proposes a method to adapt a large black-box text-to-video model to different domains without access to its weights, using the score function as a probabilistic prior for a small task-specific video model.

Benchmarking and Improving Generator-Validator Consistency of Language Models

https://openreview.net/forum?id=phBS6YpTzC

Compressor summary: The paper proposes a framework to measure and improve consistency between generating and validating answers in language models, leading to better performance on various tasks.

Think before you speak: Training Language Models With Pause Tokens

https://openreview.net/forum?id=ph04CRkPdC

Compressor summary: The text proposes adding extra hidden vectors to language models, which can improve their performance on various tasks when trained and evaluated with delays.

Graph Transformers on EHRs: Better Representation Improves Downstream Performance

https://openreview.net/forum?id=pe0Vdv7rsL

Compressor summary: GT-BEHRT combines graph transformer embeddings and BERT model to improve patient representation and perform better in medical prediction tasks using EHRs.

EXPLORING RAIN-/DETAIL-AWARE REPRESENTATION FOR INSTANCE-SPECIFIC IMAGE DE-RAINING

https://openreview.net/forum?id=pdJXYfJjz9

Compressor summary: The paper proposes a new method called CoIC for training image de-raining models that adapts to different types of rain and background, leading to better performance and generalization.

GNeRP: Gaussian-guided Neural Reconstruction of Reflective Objects with Noisy Polarization Priors

https://openreview.net/forum?id=pTN8dV2pL8

Compressor summary: The paper introduces a new way of representing normals in SDF fields using Gaussian distributions, supervised by polarization priors, to improve NeRF learning of reflective scenes and achieve better 3D reconstruction.

SelfCheck: Using LLMs to Zero-Shot Check Their Own Step-by-Step Reasoning

https://openreview.net/forum?id=pTHfApDakA

Compressor summary: SelfCheck is a method for improving question-answering by checking and correcting LLMs' reasoning steps on three datasets.

RepoBench: Benchmarking Repository-Level Code Auto-Completion Systems

https://openreview.net/forum?id=pPjZIOuQuF

Compressor summary: RepoBench is a new benchmark for evaluating code auto-completion systems on real-world, multi-file programming scenarios.

PBADet: A One-Stage Anchor-Free Approach for Part-Body Association

https://openreview.net/forum?id=pPh9p8anUi

Compressor summary: PBADet is a novel one-stage, anchor-free method for detecting human body parts and their associations with individuals, which is faster and more accurate than existing approaches.

Imitation Learning from Observation with Automatic Discount Scheduling

https://openreview.net/forum?id=pPJTQYOpNI

Compressor summary: Our framework improves ILfO by adaptively adjusting the reward priority between earlier and later steps using Automatic Discount Scheduling, enabling better learning of initial behaviors before advancing to later ones.

Robust agents learn causal world models

https://openreview.net/forum?id=pOoKI3ouv1

Compressor summary: The text argues that agents need to learn causal models to generalize under different distributions, and this has implications for transfer learning and causal inference.

Robust Adversarial Reinforcement Learning via Bounded Rationality Curricula

https://openreview.net/forum?id=pFOoOdaiue

Compressor summary: The paper proposes Quantal Adversarial RL (QARL), a novel approach for adversarial RL that uses entropy regularization and gradually increases the rationality of the adversary to ease the complexity of saddle point optimization problems and improve performance and robustness.

RAIN: Your Language Models Can Align Themselves without Finetuning

https://openreview.net/forum?id=pETSfWMUzy

Compressor summary: RAIN is a novel inference method that allows pre-trained LLMs to evaluate their own generation and use the evaluation results to guide rewind and generation for AI safety, improving harmlessness and truthfulness.

Scalable Modular Network: A Framework for Adaptive Learning via Agreement Routing

https://openreview.net/forum?id=pEKJl5sflp

Compressor summary: The paper introduces Scalable Modular Network (SMN), a modular network framework that adapts to new tasks with the help of an agreement router, which selects and composes specialist modules using iterative message passing.

Calibrated Chaos: Variance Between Runs of Neural Network Training is Harmless and Inevitable

https://openreview.net/forum?id=pEGSdJu52I

Compressor summary: The paper investigates test-set variance in neural network training, finding that it is less problematic than expected and providing insights into its causes and mitigations.

On Adversarial Training without Perturbing all Examples

https://openreview.net/forum?id=pE6gWrASQm

Compressor summary: The paper explores constructing adversarial examples only on a subset of training data and shows that it improves robustness against adversarial attacks across datasets and tasks, while reducing the computational cost.

Rethinking Adversarial Policies: A Generalized Attack Formulation and Provable Defense in RL

https://openreview.net/forum?id=pDCublKPmG

Compressor summary: This paper proposes a generalized attack framework for multi-agent reinforcement learning and a provably efficient defense using timescale separation and adversarial training.

Two-stage LLM Fine-tuning with Less Specialization and More Generalization

https://openreview.net/forum?id=pCEgna6Qco

Compressor summary: ProMoT is a two-stage fine-tuning method for large language models that reduces format specialization and improves their in-context learning performances across various tasks.

Large-scale training of foundation models for wearable biosignals

https://openreview.net/forum?id=pC3WJHf51j

Compressor summary: The authors use self-supervised learning with unlabeled data from a large study to train foundation models for photoplethysmography and electrocardiogram biosignals, which can enhance future wearable devices and help users improve their health.

Grounded Object-Centric Learning

https://openreview.net/forum?id=pBxeZ6pVUD

Compressor summary: The paper introduces Conditional Slot Attention (CoSA), a method that learns stable and invariant object-centric representations using Grounded Slot Dictionary (GSD) inspired by vector quantization, and shows its benefits in various downstream tasks.

Near-Optimal Quantum Algorithm for Minimizing the Maximal Loss

https://openreview.net/forum?id=pB1FeRSQxh

Compressor summary: The paper investigates quantum algorithms and lower bounds for minimizing the maximum of N convex functions, improving the complexity bound and proving optimality of the dependence on N.

Soft Contrastive Learning for Time Series

https://openreview.net/forum?id=pAsQSWlDUf

Compressor summary: SoftCLT is a soft contrastive learning method for time series that preserves correlations and improves downstream tasks.

Causal Modelling Agents: Causal Graph Discovery through Synergising Metadata- and Data-driven Reasoning

https://openreview.net/forum?id=pAoqRlTBtY

Compressor summary: The paper presents a novel framework that combines LLMs and DSCMs for causal discovery and shows its superior performance on benchmarks and in modelling Alzheimer's Disease.

Differentiable Learning of Generalized Structured Matrices for Efficient Deep Neural Networks

https://openreview.net/forum?id=pAVJKp3Dvn

Compressor summary: The paper proposes a framework to learn efficient structured weight matrices for deep neural networks using gradient descent, improving their performance and reducing complexity compared to previous methods.

Improved Regret Bounds for Non-Convex Online-Within-Online Meta Learning

https://openreview.net/forum?id=pA8Q5WiEMg

Compressor summary: OWO meta learning improves online algorithm initialization and step size in non-convex settings with better regret bounds and investigates generalization bounds using regret analysis.

Only Pay for What Is Uncertain: Variance-Adaptive Thompson Sampling

https://openreview.net/forum?id=p8ujRTjEf3

Compressor summary: The paper proposes a new Thompson sampling algorithm for Gaussian bandits with unknown heterogeneous reward variances, which incorporates prior knowledge to achieve lower regret and adapts better to varying uncertainty levels.

Compose and Conquer: Diffusion-Based 3D Depth Aware Composable Image Synthesis

https://openreview.net/forum?id=p4eG8rCa0b

Compressor summary: The text proposes a new conditional diffusion model, Compose and Conquer (CnC), that can control 3D object placement and style in generated images using disentangled representations and depth estimation.

Traveling Waves Encode The Recent Past and Enhance Sequence Learning

https://openreview.net/forum?id=p4S5Z6Sah4

Compressor summary: The Wave-RNN (wRNN) is a recurrent neural network model that efficiently encodes recent past through traveling waves of neural activity, and performs better than wave-free RNNs on memory tasks and sequential image classification.

Lie Group Decompositions for Equivariant Neural Networks

https://openreview.net/forum?id=p34fRKp8qA

Compressor summary: The text discusses a framework for training neural networks that are invariant and equivarent to geometrical transformations using Lie groups and their homogeneous spaces, showing improved performance on an affine-invariant benchmark task.

The Truth Is In There: Improving Reasoning with Layer-Selective Rank Reduction

https://openreview.net/forum?id=ozX92bu8VA

Compressor summary: LASER is a simple technique that improves the performance of LLMs by removing some weight matrix components, leading to significant gains in predictive accuracy on various tasks.

Backdoor Contrastive Learning via Bi-level Trigger Optimization

https://openreview.net/forum?id=oxjeePpgSP

Compressor summary: The paper proposes a new backdoor attack for contrastive learning that uses a bi-level optimization approach to create a more effective trigger and achieve higher success rates.

Yet Another ICU Benchmark: A Flexible Multi-Center Framework for Clinical ML

https://openreview.net/forum?id=ox2ATRM90I

Compressor summary: YAIB is a modular framework that allows researchers to define reproducible and comparable clinical machine learning experiments in intensive care units using various datasets, preprocessing pipelines, and prediction tasks.

Quality-Diversity through AI Feedback

https://openreview.net/forum?id=owokKCrGYr

Compressor summary: Quality-Diversity through AI Feedback (QDAIF) is a method that uses language models to guide an evolutionary algorithm in generating diverse high-quality outputs in creative writing and other domains.

Advancing the Lower Bounds: an Accelerated, Stochastic, Second-order Method with Optimal Adaptation to Inexactness

https://openreview.net/forum?id=otU31x3fus

Compressor summary: The paper introduces a new robust optimization method that generalizes existing ones and achieves optimal convergence under inexact gradients and Hessians, using tensor generalization for stochastic higher-order derivatives.

Prototypical Information Bottlenecking and Disentangling for Multimodal Cancer Survival Prediction

https://openreview.net/forum?id=otHZ8JAIgh

Compressor summary: The PIBD framework tackles intra-modal and inter-modal redundancies in multimodal learning for cancer survival prediction by using prototypical information bottlenecking and disentangling.

ReLU Strikes Back: Exploiting Activation Sparsity in Large Language Models

https://openreview.net/forum?id=osoWxY8q2E

Compressor summary: ReLU activation in large language models can save significant computation and weight transfer during memory-bound inference without compromising convergence and performance.

A Lie Group Approach to Riemannian Normalization for SPD Neural Networks

https://openreview.net/forum?id=okYdj8Ysru

Compressor summary: The paper proposes a unified framework for Riemannian Batch Normalization on Lie groups and applies it to SPD manifolds for computer vision tasks such as radar recognition, human action recognition, and EEG classification.

Copula Conformal prediction for multi-step time series prediction

https://openreview.net/forum?id=ojIJZDNIBj

Compressor summary: The paper introduces CopulaCPTS, a conformal prediction algorithm for multivariate, multi-step time series forecasting with finite sample validity guarantee and better calibration and efficiency than existing methods.

CLaM-TTS: Improving Neural Codec Language Model for Zero-Shot Text-to-Speech

https://openreview.net/forum?id=ofzeypWosV

Compressor summary: CLaM-TTS is a new method that uses probabilistic vector quantization to compress audio tokens and generate multiple streams for zero-shot TTS synthesis with improved naturalness, intelligibility, and speed.

AffineQuant: Affine Transformation Quantization for Large Language Models

https://openreview.net/forum?id=of2rhALq8l

Compressor summary: AffineQuant is a technique that optimizes post-training quantization of large language models by using equivalent affine transformations, improving performance and allowing deployment on edge devices with low-bit configurations.

Reconciling Spatial and Temporal Abstractions for Goal Representation

https://openreview.net/forum?id=odY3PkI5VB

Compressor summary: The paper presents a new HRL algorithm that combines spatial and temporal goal abstraction to improve performance on complex tasks.

LDReg: Local Dimensionality Regularized Self-Supervised Learning

https://openreview.net/forum?id=oZyAqjAjJW

Compressor summary: The paper proposes a method called *LDReg* to improve self-supervised learning by preventing dimensional collapse, which occurs when representations fail to capture the full data distribution locally.

Privacy-Preserving In-Context Learning with Differentially Private Few-Shot Generation

https://openreview.net/forum?id=oZtt0pRnOl

Compressor summary: The paper proposes an algorithm for in-context learning with large language models on private datasets, using synthetic few-shot demonstrations and formal differential privacy guarantees.

Analyzing and Mitigating Object Hallucination in Large Vision-Language Models

https://openreview.net/forum?id=oZDJKTlOUe

Compressor summary: LURE is an algorithm that corrects object hallucination in vision-language models by reconstructing more accurate descriptions based on co-occurrence, uncertainty, and object position.

Magnushammer: A Transformer-Based Approach to Premise Selection

https://openreview.net/forum?id=oYjPk8mqAV

Compressor summary: The paper proposes Magnushammer, a contrastive training-based method for premise selection in automated theorem proving, which outperforms existing tools and enables state-of-the-art results with fewer parameters.

Score Models for Offline Goal-Conditioned Reinforcement Learning

https://openreview.net/forum?id=oXjnwQLcTA

Compressor summary: SMORe is a novel discriminator-free method for offline goal-conditioned reinforcement learning that leverages mixture-distribution matching and scores for achieving diverse and reusable skills from suboptimal offline datasets without hand-engineering reward functions.

Retrieval is Accurate Generation

https://openreview.net/forum?id=oXYZJXDdo7

Compressor summary: The text introduces a novel method of generating text using context-aware phrases from supporting documents, overcoming challenges in training oracles with linguistic heuristics and self-reinforcement, and demonstrating improved performance and generation quality on various tasks.

ExeDec: Execution Decomposition for Compositional Generalization in Neural Program Synthesis

https://openreview.net/forum?id=oTRwljRgiv

Compressor summary: The paper proposes ExeDec, a new synthesis strategy that improves compositional generalization for neural program synthesis models by predicting execution subgoals step-by-step.

Online Information Acquisition: Hiring Multiple Agents

https://openreview.net/forum?id=oQKKlzxV1o

Compressor summary: The paper studies how a principal can hire multiple agents to gather and report information efficiently, considering coordination, externalities, and incentives.

Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators

https://openreview.net/forum?id=oOwDQl8haC

Compressor summary: The paper proposes a method to train and fine-tune DNNs using cheaper 12-bit accumulators without significant performance loss and shows that gradient approximations can improve accuracy with lower precision.

Treatment Effects Estimation By Uniform Transformer

https://openreview.net/forum?id=oOGqJ6Z1sA

Compressor summary: The paper proposes a new weighting method for covariate balancing and treatment effects estimation in observational studies, which works well even when the underlying model assumptions are not strong.

Graph Neural Networks for Learning Equivariant Representations of Neural Networks

https://openreview.net/forum?id=oO6FsMyDBt

Compressor summary: This paper proposes using graph neural networks and transformers to encode diverse neural network architectures and achieve better performance on various tasks.

A Unified Framework for Bayesian Optimization under Contextual Uncertainty

https://openreview.net/forum?id=oMNkj4ER7V

Compressor summary: The paper proposes a general Thompson sampling algorithm for optimizing various objectives in Bayesian optimization under contextual uncertainty, and analyzes its performance.

DNABERT-2: Efficient Foundation Model and Benchmark For Multi-Species Genomes

https://openreview.net/forum?id=oMLQB4EZE1

Compressor summary: The authors propose DNABERT-2, a refined genome foundation model that uses Byte Pair Encoding (BPE) tokenization to overcome limitations of k-mer tokenization and achieve better efficiency and performance on large genomes.

Multimodal Molecular Pretraining via Modality Blending

https://openreview.net/forum?id=oM7Jbxdk6Z

Compressor summary: MoleBLEND is a self-supervised learning approach that aligns 2D and 3D molecular structures at the atomic level by blending atom relations from both modalities and achieves state-of-the-art performance in molecular modeling tasks.

RLIF: Interactive Imitation Learning as Reinforcement Learning

https://openreview.net/forum?id=oLLZhbBSOU

Compressor summary: This paper proposes a reinforcement learning method for skill acquisition that uses user interventions as rewards, improving upon interactive imitation learning methods like DAgger, especially when the expert is suboptimal.

WebArena: A Realistic Web Environment for Building Autonomous Agents

https://openreview.net/forum?id=oKn9c6ytLx

Compressor summary: The paper presents a realistic web-based environment for evaluating language-guided agents' performance on complex tasks, showing that current state-of-the-art models struggle to match human performance.

Entropy-MCMC: Sampling from Flat Basins with Ease

https://openreview.net/forum?id=oGNdBvymod

Compressor summary: The paper proposes a new sampling method for Bayesian deep learning that biases towards flat regions in the posterior distribution to avoid overfitting and improve generalization performance.

LMUFormer: Low Complexity Yet Powerful Spiking Model With Legendre Memory Units

https://openreview.net/forum?id=oEF7qExD9F

Compressor summary: The paper introduces LMUFormer, a low-complexity recurrent model with memory units, embedding, and mixer that performs as well as transformers on speech commands while being more efficient and suitable for streaming applications at the edge.

Transformer-VQ: Linear-Time Transformers via Vector Quantization

https://openreview.net/forum?id=oDdzXQzP2F

Compressor summary: Transformer-VQ is a fast decoder-only transformer with vector-quantized keys and caching that achieves high quality in various tasks.

Sampling Multimodal Distributions with the Vanilla Score: Benefits of Data-Based Initialization

https://openreview.net/forum?id=oAMArMMQxb

Compressor summary: The authors show how to use vanilla score matching with Langevin diffusion and early stopping to generate multimodal distributions.

Adversarial AutoMixup

https://openreview.net/forum?id=o8tjamaJ80

Compressor summary: AdAutomixup is an adversarial automatic mixup augmentation method for improving image classification by generating diverse and challenging mixed samples and training a robust classifier.

Information Retention via Learning Supplemental Features

https://openreview.net/forum?id=o83eu4H9Mb

Compressor summary: The information bottleneck principle helps learn good representations by balancing conciseness and predictive ability, but may not work well in low-resource or out-of-domain scenarios; a new method is proposed to jointly learn mainline and supplemental features for better prediction.

MaGIC: Multi-modality Guided Image Completion

https://openreview.net/forum?id=o7x0XVlCpX

Compressor summary: MaGIC is a novel method for image completion that supports single and multi-modality guidance, such as text, edge, sketch, segmentation, depth, and pose, without requiring re-training of different modalities.

Sample-Efficient Multi-Agent RL: An Optimization Perspective

https://openreview.net/forum?id=o7qhUMylLU

Compressor summary: The paper proposes a new complexity measure and algorithm for sample-efficient multi-agent reinforcement learning in general-sum Markov Games, addressing different equilibrium concepts and both model-based and model-free settings.

$\pi$2vec: Policy Representation with Successor Features

https://openreview.net/forum?id=o5Bqa4o5Mi

Compressor summary: $\pi$2vec is a method that uses feature vectors and successor features to represent and evaluate black box policies efficiently in offline settings.

Learning Robust Generalizable Radiance Field with Visibility and Feature Augmented Point Representation

https://openreview.net/forum?id=o4CLLlIaaH

Compressor summary: The paper introduces a novel method called Generalizable neural Point Field (GPF) that improves generalization, consistency, and quality in neural radiance field modeling by using point-based rendering instead of image-based rendering.

Manifold Preserving Guided Diffusion

https://openreview.net/forum?id=o3BxOLoxm1

Compressor summary: MPGD is a training-free method that uses pretrained models and neural networks for conditional image generation, improving efficiency and quality.

Pre-Training Goal-based Models for Sample-Efficient Reinforcement Learning

https://openreview.net/forum?id=o2IEmeLL9r

Compressor summary: PTGM is a method that pre-trains goal-based models for reinforcement learning, improving sample efficiency and performance in complex tasks by providing temporal abstractions and behavior regularization.

Flat Minima in Linear Estimation and an Extended Gauss Markov Theorem

https://openreview.net/forum?id=nxnbPPVvOG

Compressor summary: The paper proposes a generalization of Gauss-Markov theorem for linear estimation with non-zero bias, derives optimal estimators for different norms, analyzes generalization error, and compares with Ridge regression using simulations.

Leveraging Uncertainty Estimates To Improve Classifier Performance

https://openreview.net/forum?id=nsNyDvNQTc

Compressor summary: The paper analyzes how uncertainty and model score affect estimation bias in binary classification and proposes algorithms that use both to improve recall and precision.

Searching for High-Value Molecules Using Reinforcement Learning and Transformers

https://openreview.net/forum?id=nqlymMx42E

Compressor summary: The paper introduces ChemRLformer, a new RL-based molecular design algorithm that uses text representations and explores how different design choices affect its performance on 25 molecule design tasks.

PF-LRM: Pose-Free Large Reconstruction Model for Joint Pose and Shape Prediction

https://openreview.net/forum?id=noe76eRcPC

Compressor summary: PF-LRM is a fast and accurate method for reconstructing 3D objects from unposed images with self-attention and PnP, and it works well on different datasets and scenarios.

Interpretable Meta-Learning of Physical Systems

https://openreview.net/forum?id=nnicaG5xiH

Compressor summary: CAMEL is a meta-learning architecture that learns efficiently and interprets physical parameters in various settings.

Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph

https://openreview.net/forum?id=nnVO1PvbTv

Compressor summary: The paper proposes an "LLM-KG" approach that improves large language models' reasoning by integrating external knowledge graphs and introduces a new method called Think-on-Graph (ToG) for interactive reasoning with better deep reasoning, knowledge traceability, flexibility, and cost-efficiency.

Graph Lottery Ticket Automated

https://openreview.net/forum?id=nmBjBZoySX

Compressor summary: The paper introduces AdaGLT, an adaptive, dynamic, and automated framework for finding graph lottery tickets to improve efficiency and scalability of graph neural networks on large graphs.

Video Decomposition Prior: Editing Videos Layer by Layer

https://openreview.net/forum?id=nfMyERXNru

Compressor summary: The paper presents a novel video editing framework that uses multiple RGB layers and opacity levels to perform tasks like object segmentation, dehazing, and relighting without requiring large datasets or external data collection.

Provable and Practical: Efficient Exploration in Reinforcement Learning via Langevin Monte Carlo

https://openreview.net/forum?id=nfIAEJFiBZ

Compressor summary: The paper proposes a Thompson sampling exploration strategy for reinforcement learning that samples Q-functions directly from their posterior distribution using Langevin Monte Carlo, achieving good results on deep RL and Atari57 tasks.

Escape Sky-high Cost: Early-stopping Self-Consistency for Multi-step Reasoning

https://openreview.net/forum?id=ndR8Ytrzhh

Compressor summary: ESC is a simple and scalable method to reduce the cost of chain-of-thought reasoning by using less sampling, without sacrificing performance on various reasoning tasks.

Analysis of Learning a Flow-based Generative Model from Limited Sample Complexity

https://openreview.net/forum?id=ndCJeysCPe

Compressor summary: The paper analyzes how a two-layer autoencoder can generate samples from a high-dimensional Gaussian mixture and shows that the method is optimal with an $\frac{1}{n}$ convergence rate.

An Image Is Worth 1000 Lies: Transferability of Adversarial Images across Prompts on Vision-Language Models

https://openreview.net/forum?id=nc5GgFAvtk

Compressor summary: CroPA is a method that uses learnable textual prompts to update visual adversarial perturbations and improve their transferability across different vision tasks with large vision language models.

FedCDA: Federated Learning with Cross-rounds Divergence-aware Aggregation

https://openreview.net/forum?id=nbPGqeH3lt

Compressor summary: FedCDA is a novel method for federated learning that selectively aggregates local models from different rounds to reduce diversity and maintain data knowledge.

Out-of-Distribution Detection with Negative Prompts

https://openreview.net/forum?id=nanyAujl6e

Compressor summary: The paper proposes learning a set of negative prompts for each class to improve out-of-distribution (OOD) detection by matching similarity and dissimilarity in the feature space.

AdaMerging: Adaptive Model Merging for Multi-Task Learning

https://openreview.net/forum?id=nZP6NgD3QY

Compressor summary: AdaMerging is an unsupervised technique that learns how to merge pre-trained models for multi-task learning using entropy minimization on test samples, improving performance and generalization compared to existing methods.

Optimal transport based adversarial patch to leverage large scale attack transferability

https://openreview.net/forum?id=nZP10evtkV

Compressor summary: The text proposes a distribution-oriented approach using optimal transport to create adversarial patches that can fool multiple neural networks, even without knowing the target model.

MIntRec2.0: A Large-scale Benchmark Dataset for Multimodal Intent Recognition and Out-of-scope Detection in Conversations

https://openreview.net/forum?id=nY9nITZQjc

Compressor summary: MIntRec2.0 is a large benchmark dataset for multimodal intent recognition in multi-party conversations with diverse samples, modalities, and contexts.

Tight Rates in Supervised Outlier Transfer Learning

https://openreview.net/forum?id=nUBLhhVM1l

Compressor summary: This paper studies when and how knowledge can be transferred from source to target in outlier detection tasks using Neyman-Pearson classification and measures of discrepancy.

Rethinking the Power of Graph Canonization in Graph Representation Learning with Stability

https://openreview.net/forum?id=nTwb2vBLOV

Compressor summary: The paper proposes using graph canonization to increase the expressivity of GNNs and studies their stability, introducing a universal graph canonization method that improves performance on various graph tasks.

FedInverse: Evaluating Privacy Leakage in Federated Learning

https://openreview.net/forum?id=nTNgkEIfeb

Compressor summary: The paper introduces a data-leakage risk in federated learning due to model inversion attacks and proposes FedInverse, a tool to evaluate this risk using HSIC as a regularizer.

A Simple and Scalable Representation for Graph Generation

https://openreview.net/forum?id=nO344avRib

Compressor summary: GEEL is a new graph representation that allows for more efficient and effective neural network-based graph generation, especially for large-scale and attributed graphs.

CircuitNet 2.0: An Advanced Dataset for Promoting Machine Learning Innovations in Realistic Chip Design Environment

https://openreview.net/forum?id=nMFSUjxMIl

Compressor summary: CircuitNet 2.0 is a large-scale dataset that supports machine learning innovations in chip design by providing comprehensive data from various realistic chip designs.

Input-gradient space particle inference for neural network ensembles

https://openreview.net/forum?id=nLWiR5P3wr

Compressor summary: FoRDE is a new method that improves the performance of deep ensembles by diversifying input gradients, leading to better accuracy, calibration, and robustness to perturbations.

Are Transformers with One Layer Self-Attention Using Low-Rank Weight Matrices Universal Approximators?

https://openreview.net/forum?id=nJnky5K944

Compressor summary: One-layer and single-head Transformers can effectively capture context and memorize information without relying on deep layers or high-rank weight matrices.

Structural Estimation of Partially Observed Linear Non-Gaussian Acyclic Model: A Practical Approach with Identifiability

https://openreview.net/forum?id=nHkMm0ywWm

Compressor summary: The paper proposes a method to identify causal relationships among measured and latent variables using high-order statistics and tests for statistical independence.

Monte Carlo guided Denoising Diffusion models for Bayesian linear inverse problems.

https://openreview.net/forum?id=nHESwXvxWK

Compressor summary: The study proposes a method using score-based generative models and Sequential Monte Carlo to solve ill-posed linear inverse problems in a Bayesian framework, achieving better results than existing approaches.

Cross-Modal Contextualized Diffusion Models for Text-Guided Visual Generation and Editing

https://openreview.net/forum?id=nFMS6wF2xq

Compressor summary: ContextDiff is a novel contextualized diffusion model that improves text-to-image and text-to-video tasks by incorporating cross-modal context into both forward and reverse processes.

Communication-Efficient Federated Non-Linear Bandit Optimization

https://openreview.net/forum?id=nFI3wFM9yN

Compressor summary: Fed-GO-UCB is a novel algorithm for collaborative function optimization among multiple clients that preserves data privacy, allows large-scale computing, and works with generic non-linear objective functions, achieving sub-linear rates for regret and communication cost.

Listen, Think, and Understand

https://openreview.net/forum?id=nBZBPXdJlC

Compressor summary: LTU is a new model that combines audio perception and reasoning abilities, trained on OpenAQA-5M dataset, which outperforms existing audio models in classification and captioning tasks and shows emerging audio reasoning skills.

Self-Supervised Contrastive Forecasting

https://openreview.net/forum?id=nBCuRzjqK7

Compressor summary: The paper proposes a new approach using contrastive learning and enhanced decomposition to improve long-term forecasting by capturing and incorporating global autocorrelation in time series.

Federated Orthogonal Training: Mitigating Global Catastrophic Forgetting in Continual Federated Learning

https://openreview.net/forum?id=nAs4LdaP9Y

Compressor summary: Federated Orthogonal Training (FOT) is a novel method that address global catastrophic forgetting in Continual Federated Learning by reducing interference between tasks using orthogonal updates, while preserving privacy and requiring low computation and communication cost.

The Consensus Game: Language Model Generation via Equilibrium Search

https://openreview.net/forum?id=n9xeGcI4Yg

Compressor summary: Equilibrium-ranking is a new game-theoretic procedure for language model decoding that improves performance on various text generation tasks without needing additional training data.

Mayfly: a Neural Data Structure for Graph Stream Summarization

https://openreview.net/forum?id=n7Sr8SW4bn

Compressor summary: The Mayfly is a neural data structure that summarizes graph streams for real-time graph analytics in various domains by learning from meta-tasks.

HAZARD Challenge: Embodied Decision Making in Dynamically Changing Environments

https://openreview.net/forum?id=n6mLhaBahJ

Compressor summary: HAZARD is a new simulated benchmark for testing embodied agents' decision-making abilities in dynamic environments with disaster scenarios and large language models.

Expressive Losses for Verified Robustness via Convex Combinations

https://openreview.net/forum?id=mzyZ4wzKlM

Compressor summary: The text discusses how expressive losses can improve adversarial robustness and accuracy trade-offs in networks by over-approximating the worst-case loss.

Scalable Language Model with Generalized Continual Learning

https://openreview.net/forum?id=mz8owj4DXu

Compressor summary: The Scalable Language Model (SLM) overcomes limitations of continual learning in language models by using JARe and DTKR, achieving state-of-the-art performance on diverse tasks and domains with minimal forgetting.

OctoPack: Instruction Tuning Code Large Language Models

https://openreview.net/forum?id=mw1PWNSWZP

Compressor summary: Instruction tuning using code from Git commits improves large language models' performance on various coding tasks across multiple languages.

Boosting Prompting Mechanisms for Zero-Shot Speech Synthesis

https://openreview.net/forum?id=mvMI3N4AvD

Compressor summary: Mega-TTS is a new zero-shot text-to-speech method that uses an acoustic autoencoder and a prosody language model to generate speech with unseen speakers and styles from short prompts.

Views Can Be Deceiving: Improved SSL Through Feature Space Augmentation

https://openreview.net/forum?id=mutJBk3ILg

Compressor summary: The paper explores how spurious features affect self-supervised learning and proposes a method called LateTVG to remove them by pruning later layers of the encoder.

FedCompass: Efficient Cross-Silo Federated Learning on Heterogeneous Client Devices Using a Computing Power-Aware Scheduler

https://openreview.net/forum?id=msXxrttLOi

Compressor summary: FedCompass is a semi-asynchronous federated learning algorithm that adapts to client computing power, reducing staleness and waiting times, and improving convergence and accuracy on non-IID heterogeneous datasets.

Bridging State and History Representations: Understanding Self-Predictive RL

https://openreview.net/forum?id=ms0VgzSGF2

Compressor summary: The paper proposes a common idea of self-predictive abstraction for representation learning in deep reinforcement learning and provides a minimalist algorithm to learn such representations across different domains.

Skeleton-of-Thought: Large Language Models Can Do Parallel Decoding

https://openreview.net/forum?id=mqVgBbNCm9

Compressor summary: The paper proposes Skeleton-of-Thought, a method that speeds up large language models by guiding them to generate skeletons of answers in parallel and then filling them in sequentially.

KW-Design: Pushing the Limit of Protein Deign via Knowledge Refinement

https://openreview.net/forum?id=mpqMVWgqjn

Compressor summary: KW-Design is a novel protein inverse folding method that leverages pretrained models, confidence, and memory retrieval to overcome limitations and achieve high recovery rates.

Open the Black Box: Step-based Policy Updates for Temporally-Correlated Episodic Reinforcement Learning

https://openreview.net/forum?id=mnipav175N

Compressor summary: The paper introduces Temporally-Correlated Episodic RL (TCE), a method that generates smooth trajectories for robot tasks by using multi-second trajectory samples and evaluating policy updates based on segment-wise advantages, improving sample efficiency and capturing movement correlations.

Multi-Resolution Diffusion Models for Time Series Forecasting

https://openreview.net/forum?id=mmjnr0G8ZY

Compressor summary: The paper proposes a novel multi-resolution diffusion model for time series data that leverages temporal structure and outperforms existing models.

Class Probability Matching with Calibrated Networks for Label Shift Adaption

https://openreview.net/forum?id=mliQ2huFrZ

Compressor summary: The paper proposes a new label adaptation framework called class probability matching (CPM) that matches class probabilities between source and target domains and improves generalization and efficiency over existing methods.

RECOMP: Improving Retrieval-Augmented LMs with Context Compression and Selective Augmentation

https://openreview.net/forum?id=mlJLVigNHp

Compressor summary: The paper introduces compressors that generate textual summaries from retrieved documents for retrieval-augmented language models, improving efficiency and performance on language modeling and question answering tasks.

Enabling Efficient Equivariant Operations in the Fourier Basis via Gaunt Tensor Products

https://openreview.net/forum?id=mhyQXJ6JsK

Compressor summary: This paper proposes a new method to accelerate computations for 3D data modeling using equivariant neural networks, reducing complexity from $\mathcal{O}(L^6)$ to $\mathcal{O}(L^3)$.

Noise Map Guidance: Inversion with Spatial Context for Real Image Editing

https://openreview.net/forum?id=mhgm0IXtHw

Compressor summary: Noise Map Guidance (NMG) is a new method for real-image editing that uses spatial context without requiring optimization, overcoming the limitations of previous text-guided diffusion models.

Improving Code Style for Accurate Code Generation

https://openreview.net/forum?id=maRYffiUpI

Compressor summary: The text discusses improving code generation performance by making programs more structured and readable using a data-cleaning pipeline and shows that this leads to better results even with less data.

FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning

https://openreview.net/forum?id=mZn2Xyh9Ec

Compressor summary: FlashAttention-2 improves the efficiency and scalability of Transformers by optimizing work partitioning in the attention layer, reaching up to 73% of the theoretical maximum FEOPs/s on A100 GPUs.

Rethinking CNN’s Generalization to Backdoor Attack from Frequency Domain

https://openreview.net/forum?id=mYhH0CDFFa

Compressor summary: The paper studies how CNNs memorize trigger changes in poisoned samples' frequency domain, proposes a universal invisible strategy for backdoor attacks, and designs a low-frequency backdoor attack method with high accuracy.

Analyzing Feed-Forward Blocks in Transformers through the Lens of Attention Map

https://openreview.net/forum?id=mYWsyTuiRp

Compressor summary: The study visualizes how feed-forward blocks in Transformers modify input contextualization and may reveal redundant processing in these models.

On the Vulnerability of Adversarially Trained Models Against Two-faced Attacks

https://openreview.net/forum?id=mXpNp8MMr5

Compressor summary: The paper warns that adversarially trained models can be fooled by two-faced attacks, which may mislead evaluations and cause security issues, and suggests a countermeasure.

BrainSCUBA: Fine-Grained Natural Language Captions of Visual Cortex Selectivity

https://openreview.net/forum?id=mQYHXUUTkU

Compressor summary: The researchers developed a method to generate natural language descriptions for images based on neural activity in different regions of the brain, enabling new insights into how the brain processes visual information.

A Hierarchical Bayesian Model for Few-Shot Meta Learning

https://openreview.net/forum?id=mQ72XRfYRZ

Compressor summary: The paper proposes a new hierarchical Bayesian model for few-shot learning that uses global and local random variables to infer novel episodes, and exploits a Normal-Inverse-Wishart model for efficient training.

LLM Blueprint: Enabling Text-to-Image Generation with Complex and Detailed Prompts

https://openreview.net/forum?id=mNYF0IHbRy

Compressor summary: The authors propose a novel approach using Large Language Models to extract information from complex text prompts and generate detailed images, improving the fidelity of text-to-image generation models.

Beyond task performance: evaluating and reducing the flaws of large multimodal models with in-context-learning

https://openreview.net/forum?id=mMaQvkMzDi

Compressor summary: The text discusses the limitations of large multimodal models and explores in-context learning as a potential solution, proposing new multimodal ICL approaches to improve their performance.

SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents

https://openreview.net/forum?id=mM7VurbA4r

Compressor summary: SOTOPIA is an environment for testing AI's social intelligence by simulating complex interactions between agents and humans, revealing GPT-4's challenges in social commonsense reasoning and communication.

Emu: Generative Pretraining in Multimodality

https://openreview.net/forum?id=mL8Q9OOamV

Compressor summary: Emu is a versatile model that generates images and text from various inputs, achieving high performance in multiple multimodal tasks.

Seeking Neural Nuggets: Knowledge Transfer in Large Language Models from a Parametric Perspective

https://openreview.net/forum?id=mIEHIcHGOo

Compressor summary: The paper explores how to transfer knowledge from larger to smaller language models using sensitivity-based techniques and a module called LoRA.

Rethinking the Benefits of Steerable Features in 3D Equivariant Graph Neural Networks

https://openreview.net/forum?id=mGHJAyR8w0

Compressor summary: This study analyzes the expressive power and performance of invariant and equivariant GNNs, showing that increasing the feature dimension is more important than increasing the number of types of steerable features for improving performance.

Sum-Product-Set Networks: Deep Tractable Models for Tree-Structured Graphs

https://openreview.net/forum?id=mF3cTns4pe

Compressor summary: Sum-product-set networks extend probabilistic circuits to tree-structured graphs using random finite sets for exact and efficient inference, comparing well with intractable neural network models.

An Analytical Solution to Gauss-Newton Loss for Direct Image Alignment

https://openreview.net/forum?id=mE52zURNGc

Compressor summary: The authors derive a closed-form solution for the expected optimum of the Gauss-Newton loss in direct image alignment, which improves accuracy, robustness, and convergence properties, and offers insight into end-to-end learning limitations.

EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations

https://openreview.net/forum?id=mCOBKZmrzD

Compressor summary: EquiformerV2 improves on previous methods for 3D atomistic systems using eSCN convolutions, attention re-normalization, separable $S^2$ activation, and separable layer normalization, achieving better results and efficiency.

Federated Causal Discovery from Heterogeneous Data

https://openreview.net/forum?id=m7tJxajC3G

Compressor summary: The paper proposes a novel federated causal discovery method that can handle diverse scenarios with arbitrary causal models and heterogeneous data, while protecting data privacy using summary statistics.

DRSM: De-Randomized Smoothing on Malware Classifier Providing Certified Robustness

https://openreview.net/forum?id=m7aPLHwsLr

Compressor summary: The paper proposes DRSM, a certified defense against evasion attacks on malware detection models, which preserves accuracy and robustness by limiting the impact of adversarial bytes using window ablation.

Constructing Adversarial Examples for Vertical Federated Learning: Optimal Client Corruption through Multi-Armed Bandit

https://openreview.net/forum?id=m52uU0dVbH

Compressor summary: The paper proposes a novel adversarial attack on vertical federated learning by adaptively corrupting subsets of clients using Thompson sampling with Empirical maximum reward (E-TS) algorithm.

Fantastic Gains and Where to Find Them: On the Existence and Prospect of General Knowledge Transfer between Any Pretrained Model

https://openreview.net/forum?id=m50eKHCttz

Compressor summary: The authors investigate how to transfer complementary knowledge from one deep network to another without performance degradation, using data partitioning techniques that work for most pretrained models and can be done unsupervised.

Policy Rehearsing: Training Generalizable Policies for Reinforcement Learning

https://openreview.net/forum?id=m3xVPaZp6Z

Compressor summary: ReDM is a new algorithm that enables reinforcement learning agents to make adaptive decisions by planning for all possible outcomes and rehearsing on diverse dynamics models, even with zero or limited interaction data.

Sentence-level Prompts Benefit Composed Image Retrieval

https://openreview.net/forum?id=m3ch3kJL7q

Compressor summary: The paper proposes a new method for composed image retrieval that uses sentence-level prompts generated by V-L models instead of pseudo-word tokens, improving performance on complex changes in the target image.

DENEVIL: TOWARDS DECIPHERING AND NAVIGATING THE ETHICAL VALUES OF LARGE LANGUAGE MODELS VIA INSTRUCTION LEARNING

https://openreview.net/forum?id=m3RRWWFaVe

Compressor summary: This paper explores the ethical values of large language models using Moral Foundation Theory, creates a dataset of moral prompts, and proposes an alignment method to improve value compliance.

To the Cutoff... and Beyond? A Longitudinal Perspective on LLM Data Contamination

https://openreview.net/forum?id=m2NVG4Htxs

Compressor summary: The authors analyze data contamination in large language models by studying their performance on two code/mathematical problem-solving datasets, finding significant trends that suggest contamination.

Time-Varying Propensity Score to Bridge the Gap between the Past and Present

https://openreview.net/forum?id=m0x0rv6Iwm

Compressor summary: This paper proposes a time-varying propensity score to handle gradually evolving data in machine learning models by selecting past data that is similar to the current data.

Faster Approximation of Probabilistic and Distributional Values via Least Squares

https://openreview.net/forum?id=lvSMIsztka

Compressor summary: This paper proposes two fast estimators for valuing data using probabilistic values from cooperative game theory, based on a connection to least square regressions, and shows how distributional values can be optimized similarly.

RobustTSF: Towards Theory and Design of Robust Time Series Forecasting with Anomalies

https://openreview.net/forum?id=ltZ9ianMth

Compressor summary: The paper analyzes three types of anomalies in time series data and proposes a robust forecasting algorithm that outperforms current methods.

Bandits Meet Mechanism Design to Combat Clickbait in Online Recommendation

https://openreview.net/forum?id=lsxeNvYqCj

Compressor summary: The text studies a strategic version of the multi-armed bandit problem, where arms choose their click-rates to maximize clicks, and proposes an algorithm that balances incentives and regret minimization.

EControl: Fast Distributed Optimization with Compression and Error Control

https://openreview.net/forum?id=lsvlvWB9vz

Compressor summary: The paper proposes EControl, a new mechanism for distributed training with contractive compression that regulates error compensation and ensures fast convergence without strong assumptions on data heterogeneity.

A Dynamical View of the Question of Why

https://openreview.net/forum?id=lrQlLqQase

Compressor summary: The paper introduces a method to learn causality in multivariate time series data using reinforcement learning and shows its effectiveness in discovering and measuring causal relationships in stochastic processes.

Submodular Reinforcement Learning

https://openreview.net/forum?id=loYSzjSaAK

Compressor summary: The paper proposes a reinforcement learning framework called subRL that handles non-additive rewards with diminishing returns using submodular set functions, and presents an efficient policy gradient algorithm called subPO for various applications.

Bounding Box Stability against Feature Dropout Reflects Detector Generalization across Environments

https://openreview.net/forum?id=lmM4Ecm4HJ

Compressor summary: The text proposes a box stability score (BS score) that reflects how much bounding boxes change under feature map dropout, which correlates with object detection accuracy and enables assessment of detectors without test ground truths.

Ring-A-Bell! How Reliable are Concept Removal Methods For Diffusion Models?

https://openreview.net/forum?id=lm7MRcsFiS

Compressor summary: Ring-A-Bell is a tool that tests the effectiveness of safety measures in text-to-image models by generating inappropriate content based on extracted concepts.

The Optimal Constant Solution: Predictable Extrapolation in Deep Neural Networks

https://openreview.net/forum?id=ljwoQ3cvQh

Compressor summary: Neural network predictions for high-dimensional inputs tend to become constant when faced with out-of-distribution inputs, often approximating the optimal constant solution.

ViLMA: A Zero-Shot Benchmark for Linguistic and Temporal Grounding in Video-Language Models

https://openreview.net/forum?id=liuqDwmbQJ

Compressor summary: ViLMA is a task-agnostic benchmark that evaluates the fine-grained capabilities of pretrained Video-Language Models (VidLMs) by using carefully curated counterfactual tests and proficiency tests.

The Expressive Power of Low-Rank Adaptation

https://openreview.net/forum?id=likXVjmh3E

Compressor summary: LoRA is a technique for fine-tuning pre-trained models that can adapt any smaller model to any larger model with a certain rank, and this paper analyzes its expressive power and approximation error.

Sliced Denoising: A Physics-Informed Molecular Pre-Training Method

https://openreview.net/forum?id=liKkG1zcWq

Compressor summary: SliDe is a novel molecular pre-training method that uses physical principles to improve generalization and robustness in drug discovery by enhancing the accuracy of estimated force fields.

Structuring Representation Geometry with Rotationally Equivariant Contrastive Learning

https://openreview.net/forum?id=lgaFMvZHSJ

Compressor summary: The paper proposes CARE, a self-supervised learning method that adds geometric structure to embeddings by enforcing equivariance and improves downstream tasks and sensitivity to data variations.

The Cost of Scaling Down Large Language Models: Reducing Model Size Affects Memory before In-context Learning

https://openreview.net/forum?id=ldJXXxPE0L

Compressor summary: Weight pruning affects large language models' abilities to recall facts from pre-training and process information in context differently, with moderate pruning impairing fact recall but less so processing in context.

SRL: Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores

https://openreview.net/forum?id=lajn1iROCu

Compressor summary: SRL is a scalable, efficient, and extensible distributed RL system that enables massively parallelized training and easy development of customized algorithms, outperforming existing libraries in throughput and learning performance.

Delphic Offline Reinforcement Learning under Nonidentifiable Hidden Confounding

https://openreview.net/forum?id=lUYY2qsRTI

Compressor summary: The paper proposes a method to estimate and account for uncertainty due to hidden confounding bias in offline reinforcement learning, using delphic uncertainty, and shows its effectiveness on sepsis management and electronic health records benchmarks.

Efficient Network Embedding in the Exponentially Large Quantum Hilbert Space: A High-Dimensional Perspective on Embedding

https://openreview.net/forum?id=lROh08eK6n

Compressor summary: The paper proposes a new network embedding method using high-dimensional quantum states that outperforms low-dimensional methods while being more efficient and robust.

On Diffusion Modeling for Anomaly Detection

https://openreview.net/forum?id=lR3rk7ysXz

Compressor summary: The paper proposes Diffusion Time Estimation, a simplified and faster version of Denoising Diffusion Probability Models, for anomaly detection in both unsupervised and semi-supervised scenarios.

Improved Efficiency Based on Learned Saccade and Continuous Scene Reconstruction From Foveated Visual Sampling

https://openreview.net/forum?id=lOwkOIUJtx

Compressor summary: The authors propose a system that uses multiple foveal views and a reinforcement learning-based saccade mechanism to reconstruct scenes, achieving high image recognition accuracy with reduced input pixels and computation.

MCM: Masked Cell Modeling for Anomaly Detection in Tabular Data

https://openreview.net/forum?id=lNZJyEDxy4

Compressor summary: The paper proposes a new masking strategy for anomaly detection in tabular data, which captures intrinsic feature correlations and improves interpretability.

Non-negative Contrastive Learning

https://openreview.net/forum?id=lNCnZwcH5Z

Compressor summary: NCL is a non-negative contrastive learning method for deriving interpretable and sparse features from data clusters, improving over standard contrastive learning in various tasks.

Grounding Multimodal Large Language Models to the World

https://openreview.net/forum?id=lLmqxkfSIw

Compressor summary: Kosmos-2 is a large language model that connects text to visual objects and enables various multimodal tasks.

CLIP-MUSED: CLIP-Guided Multi-Subject Visual Neural Information Semantic

https://openreview.net/forum?id=lKxL5zkssv

Compressor summary: The CLIP-MUSED method uses a Transformer and subject-specific tokens to decode visual neural information from multiple subjects, overcoming limitations of existing methods and achieving state-of-the-art performance on fMRI datasets.

TokenFlow: Consistent Diffusion Features for Consistent Video Editing

https://openreview.net/forum?id=lKK50q2MtV

Compressor summary: The paper presents a text-driven video editing framework that uses a text-to-image diffusion model to generate high-quality videos from source videos and target text prompts, without needing training or fine-tuning.

Bridging Vision and Language Spaces with Assignment Prediction

https://openreview.net/forum?id=lK2V2E2MNv

Compressor summary: VLAP bridges vision encoders and language models by using word embeddings to map visual representations into language space for improved performance on vision-language tasks.

Multi-scale Transformers with Adaptive Pathways for Time Series Forecasting

https://openreview.net/forum?id=lJkOCMP2aW

Compressor summary: Pathformer is a multi-scale transformer model that adapts to varying temporal dynamics in time series, improving forecasting accuracy and generalization.

Context-Aware Meta-Learning

https://openreview.net/forum?id=lJYAkDVnRU

Compressor summary: Our meta-learning algorithm learns new visual concepts during inference without fine-tuning, achieving state-of-the-art results on 8 out of 11 benchmarks.

LightHGNN: Distilling Hypergraph Neural Networks into MLPs for 100x Faster Inference

https://openreview.net/forum?id=lHasEfGsXL

Compressor summary: The paper proposes LightHGNN and LightHGNN+, which distill knowledge from hypergraph neural networks (HGNNs) to multi-layer perceptrons (MLPs) to reduce complexity and improve inference speed while maintaining competitive performance.

Efficient ConvBN Blocks for Transfer Learning and Beyond

https://openreview.net/forum?id=lHZm9vNm5H

Compressor summary: This paper introduces a new mode, Tune, for ConvBN blocks that balances efficiency and stability in computer vision tasks, reducing memory footprint and training time while maintaining performance.

Bridging the Gap between Binary Neural Networks and Spiking Neural Networks for Efficient Computer Vision

https://openreview.net/forum?id=lGUyAuuTYZ

Compressor summary: The paper proposes a new training method for sparse BNNs that uses a Hoyer regularizer to improve accuracy and energy efficiency over BNNs, SNNs, and adder neural networks.

Vision-Language Foundation Models as Effective Robot Imitators

https://openreview.net/forum?id=lFYj0oibGR

Compressor summary: RoboFlamingo is a simple framework that adapts vision-language models to robot control by fine-tuning them on language-conditioned manipulation datasets, achieving state-of-the-art performance and enabling open-loop control.

Demonstration-Regularized RL

https://openreview.net/forum?id=lF2aip4Scn

Compressor summary: The paper analyzes how using expert demonstrations can improve the sample efficiency of reinforcement learning and provides theoretical guarantees for this method in different settings.

Unlock Predictable Scaling from Emergent Abilities

https://openreview.net/forum?id=lDbjooxLkD

Compressor summary: The study develops an evaluation strategy called PassUntil to measure task performance improvements in small language models, revealing a strict task scaling law and characterizing emergent abilities in large language models.

Multi-Scale Representations by Varing Window Attention for Semantic Segmentation

https://openreview.net/forum?id=lAhWGOkpSR

Compressor summary: VWFormer is a novel multi-scale learner for semantic segmentation that uses varying window attention and achieves competitive efficiency with lower computational cost than existing methods.

Assessing Uncertainty in Similarity Scoring: Performance & Fairness in Face Recognition

https://openreview.net/forum?id=lAhQCHuANV

Compressor summary: The article presents asymptotic guarantees for empirical ROC curves and shows that naive bootstrap approach can be misleading, suggesting a dedicated recentering technique instead, which is experimentally validated for assessing similarity functions and fairness in face recognition.

Revisiting Deep Audio-Text Retrieval Through the Lens of Transportation

https://openreview.net/forum?id=l60EM8md3t

Compressor summary: The paper proposes m-LTM, a scalable and robust inverse optimal transport framework for audio-text matching that uses mini-batch subsampling, neural networks, Mahalanobis-enhanced ground metrics, and partial optimal transport to learn rich joint embeddings.

A Poincaré Inequality and Consistency Results for Signal Sampling on Large Graphs

https://openreview.net/forum?id=l3qtSNsPvC

Compressor summary: The paper introduces a new graph sampling technique based on graphon signals, which simplifies large-scale graph machine learning by using unique sampling sets that are consistent across converging graphs.

DV-3DLane: End-to-end Multi-modal 3D Lane Detection with Dual-view Representation

https://openreview.net/forum?id=l1U6sEgYkb

Compressor summary: DV-3DLane is a novel framework that combines images and LiDAR points to accurately detect 3D lanes using dual-view multi-modal features, feature fusion, and attention mechanisms.

Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation

https://openreview.net/forum?id=kzGuiRXZrQ

Compressor summary: EQGAT-diff is a new E(3) equivariant diffusion model for 3D molecular design that outperforms existing models on large and limited data sets with explicit hydrogens.

COLLIE: Systematic Construction of Constrained Text Generation Tasks

https://openreview.net/forum?id=kxgSlyirUZ

Compressor summary: COLLIE is a framework for creating diverse grammar-based constraints for text generation tasks and analyzing the performance of language models on them.

TRAM: Bridging Trust Regions and Sharpness Aware Minimization

https://openreview.net/forum?id=kxebDHZ7b7

Compressor summary: TRAM is a fine-tuning algorithm that improves out-of-domain generalization by optimizing flat minima and smooth, informative representations while preserving pre-trained structure.

Enhancing Contrastive Learning for Ordinal Regression via Ordinal Content Preserved Data Augmentation

https://openreview.net/forum?id=kx2XZlmgB1

Compressor summary: The paper proposes a new augmentation method for contrastive learning to improve ordinal regression by preserving essential ordinal content information in images, outperforming existing methods.

Successor Heads: Recurring, Interpretable Attention Heads In The Wild

https://openreview.net/forum?id=kvcbV8KQsi

Compressor summary: The paper introduces successor heads, attention heads that increment tokens with natural ordering, and uses mechanistic interpretability to analyze their behavior in large language models, finding common abstract representations and mod 10 features.

Estimating Shape Distances on Neural Representations with Limited Samples

https://openreview.net/forum?id=kvByNnMERu

Compressor summary: The paper derives bounds on the worst-case convergence of shape distance estimators and introduces a novel method-of-moments estimator with tunable bias-variance for measuring geometric similarity in high-dimensional network representations.

JointNet: Extending Text-to-Image Diffusion for Dense Distribution Modeling

https://openreview.net/forum?id=kv5xE1p3jz

Compressor summary: JointNet is a neural network that models the relationship between images and an additional modality (e.g., depth maps) using a pre-trained text-to-image diffusion model and enables various applications like joint image-depth generation and depth prediction.

Continuous Field Reconstruction from Sparse Observations with Implicit Neural Networks

https://openreview.net/forum?id=kuTZMZdCPZ

Compressor summary: The paper proposes a new deep neural network method to reconstruct physical fields from sparse sensor data, which improves reconstruction quality on realistic datasets.

Towards image compression with perfect realism at ultra-low bitrates

https://openreview.net/forum?id=ktdETU9JBg

Compressor summary: PerCo is a perceptual compression model that uses diffusion models for decoding and textual image descriptions for context, achieving state-of-the-art visual quality at ultra-low bitrates.

On Accelerating Diffusion-Based Sampling Processes via Improved Integration Approximation

https://openreview.net/forum?id=ktJAF3lxbi

Compressor summary: The paper proposes a method to improve ODE-based diffusion models by optimizing coefficients using improved integration approximation, resulting in better FID scores for small NFE values.

Text-to-3D with Classifier Score Distillation

https://openreview.net/forum?id=ktG8Tun1Cy

Compressor summary: The paper proposes Classifier Score Distillation (CSD), a method that uses an implicit classification model for text-to-3D generation without relying on pre-trained 2D diffusion models, achieving superior results in various tasks.

Hiding in Plain Sight: Disguising Data Stealing Attacks in Federated Learning

https://openreview.net/forum?id=krx55l2A6G

Compressor summary: The paper studies how malicious server attacks can steal data from federated learning, shows that existing attacks are detectable, and proposes a new attack method called SEER that can evade detection.

DynaVol: Unsupervised Learning for Dynamic Scenes through Object-Centric Voxelization

https://openreview.net/forum?id=koYsgfEwCQ

Compressor summary: DynaVol is a 3D scene generative model that learns object-centric representations in dynamic visual scenes, enabling unsupervised scene decomposition and manipulation of geometric shapes and motion trajectories.

Beyond Memorization: Violating Privacy via Inference with Large Language Models

https://openreview.net/forum?id=kmn0BhQk7p

Compressor summary: The study shows that large language models can accurately infer personal attributes from text, posing a serious threat to individual privacy and highlighting the need for better protection methods.

Local Composite Saddle Point Optimization

https://openreview.net/forum?id=kklwv4c4dI

Compressor summary: FeDualEx is a new distributed optimization method for saddle point problems with composite objectives and constraints that preserves structure in machine learning.

FedDA: Faster Adaptive Gradient Methods for Federated Constrained Optimization

https://openreview.net/forum?id=kjn99xFUF3

Compressor summary: FedDA is a new adaptive gradient method for federated learning that works with constraints, and it achieves better performance than FedAvg.

Towards Universal Multi-Modal Personalization: A Language Model Empowered Generative Paradigm

https://openreview.net/forum?id=khAE1sTMdX

Compressor summary: The paper proposes UniMP, a framework that leverages multi-modal data for personalized tasks such as recommendation, search, and generation, using interleaved cross-modal user history information.

Large Language Models as Automated Aligners for benchmarking Vision-Language Models

https://openreview.net/forum?id=kZEXgtMNNo

Compressor summary: Auto-Bench is a tool that uses large language models to automatically generate and assess question-answer-reasoning tasks for vision-language models, measuring their alignment with human intelligence.

Accelerated Sampling with Stacked Restricted Boltzmann Machines

https://openreview.net/forum?id=kXNJ48Hvw1

Compressor summary: The authors propose Stacked Tempering, a method that improves sampling of complex distributions using nested restricted Boltzmann machines (RBMs) and replica exchange, achieving faster transitions between data modes.

Simple Hierarchical Planning with Diffusion

https://openreview.net/forum?id=kXHEBK9uAY

Compressor summary: The Hierarchical Diffuser is a fast and effective planning method that combines hierarchical and diffusion-based planning to overcome computational challenges and improve generalization in long-horizon tasks.

Improving equilibrium propagation without weight symmetry through Jacobian homeostasis

https://openreview.net/forum?id=kUveo5k1GF

Compressor summary: The paper studies how weight asymmetry and finite nudges affect gradient estimation in equilibrium propagation, a neural network algorithm for neuromorphic systems, and proposes a homeostatic objective to reduce residual bias and improve task performance.

Multi-resolution HuBERT: Multi-resolution Speech Self-Supervised Learning with Masked Unit Prediction

https://openreview.net/forum?id=kUuKFW7DIF

Compressor summary: The paper proposes a hierarchical Transformer model that processes speech at multiple resolutions, achieving better performance and efficiency than existing models.

SF(DA)$^2$: Source-free Domain Adaptation Through the Lens of Data Augmentation

https://openreview.net/forum?id=kUCgHbmO11

Compressor summary: The paper presents a new method, SF(DA)$^2$, for adapting models to unseen domains without source data, using data augmentation based on feature space neighbors and class semantic information.

Neuron-Enhanced AutoEncoder Matrix Completion: Theory and Practice

https://openreview.net/forum?id=kPrxk6tUcg

Compressor summary: The paper proposes a neuron-enhanced autoencoder matrix completion method that improves reconstruction capability and response function approximation, and shows its theoretical analysis and numerical results on synthetic and benchmark datasets.

Dynamic Sparse Training with Structured Sparsity

https://openreview.net/forum?id=kOBkxFRKTA

Compressor summary: SRigL is a sparse-to-sparse method that learns structured N:M sparsity and achieves fast online inference on CPU.

Maximum Entropy Model Correction in Reinforcement Learning

https://openreview.net/forum?id=kNpSUN0uCc

Compressor summary: The paper presents a reinforcement learning approach that uses a MaxEnt Model Correction procedure to improve the accuracy of an approximate model and speed up convergence to the true value function.

Leveraging Unpaired Data for Vision-Language Generative Models via Cycle Consistency

https://openreview.net/forum?id=kNjrhD67LP

Compressor summary: ITIT is a novel training method for vision-language models that uses unpaired images and texts to achieve high performance with minimal paired data.

What's in a Prior? Learned Proximal Networks for Inverse Problems

https://openreview.net/forum?id=kNPcOaqC5r

Compressor summary: The paper presents learned proximal networks (LPN), which use proximal matching to learn exact proximal operators for data-driven regularizers that recover the log-prior of the true data distribution, enabling general, unsupervised, and state-of-the-art inverse problem solutions.

ASMR: Activation-Sharing Multi-Resolution Coordinate Networks for Efficient Inference

https://openreview.net/forum?id=kMp8zCsXNb

Compressor summary: ASMR is a new method for encoding images and videos with a compact neural representation that reduces inference complexity and improves reconstruction quality.

SE(3)-Stochastic Flow Matching for Protein Backbone Generation

https://openreview.net/forum?id=kJFIH23hXb

Compressor summary: Foldflow is a series of generative models that use the flow-matching paradigm over 3D rigid motions (SE(3)) to accurately model protein backbones, offering stability, speed, and flexibility in protein design.

Towards Aligned Layout Generation via Diffusion Model with Aesthetic Constraints

https://openreview.net/forum?id=kJ0qp9Xdsh

Compressor summary: LACE is a unified continuous diffusion model for controllable layout generation that incorporates aesthetic constraints and achieves better results than existing methods.

Outliers with Opposing Signals Have an Outsized Effect on Neural Network Optimization

https://openreview.net/forum?id=kIZ3S3tel6

Compressor summary: The paper introduces a new phenomenon in neural network optimization caused by heavy-tailed data and opposing signals, explaining some training dynamics and suggesting improvements to stochastic optimization methods.

Reverse Diffusion Monte Carlo

https://openreview.net/forum?id=kIPEyMSdFV

Compressor summary: Reverse diffusion transforms score estimation into mean estimation, leading to a novel Monte Carlo sampling algorithm that improves upon traditional MCMC methods for Gaussian mixture models.

Test-Time Adaptation with CLIP Reward for Zero-Shot Generalization in Vision-Language Models

https://openreview.net/forum?id=kIP0duasBb

Compressor summary: The paper proposes reinforcement learning with CLIP feedback (RLCF) for adapting pre-trained vision-language models to improve their zero-shot generalization ability on various tasks.

On the Generalization and Approximation Capacities of Neural Controlled Differential Equations

https://openreview.net/forum?id=kILAd8RdzA

Compressor summary: The paper presents a theoretical analysis of Neural Controlled Differential Equations (NCDE) for supervised learning with irregularly sampled time series, examining how sampling affects predictions and generalization bounds.

Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs

https://openreview.net/forum?id=kGteeZ18Ir

Compressor summary: The study shows that large-scale language models can have implicit biases and perform worse in reasoning tasks when assigned personas, despite rejecting stereotypes explicitly.

Selective Visual Representations Improve Convergence and Generalization for Embodied AI

https://openreview.net/forum?id=kC5nZDU5zf

Compressor summary: The paper proposes a method for embodied AI to filter visual input based on tasks, improving performance and generalization in navigation and manipulation tasks.

Faithful Rule Extraction for Differentiable Rule Learning Models

https://openreview.net/forum?id=kBTzlxM2J1

Compressor summary: The paper studies how to extract rules from knowledge graph ML models that are both sound and complete, which is important for safety-critical applications and legal compliance.

Lifting Architectural Constraints of Injective Flows

https://openreview.net/forum?id=kBNIx4Biq4

Compressor summary: Injective Flows learn a low-dimensional manifold and its distribution, improving efficiency and stability with a new estimator for the maximum likelihood loss.

FasterViT: Fast Vision Transformers with Hierarchical Attention

https://openreview.net/forum?id=kB4yBiNmXX

Compressor summary: FasterViT is a new hybrid neural network that combines CNNs and ViTs, using Hierarchical Attention to achieve high image throughput and accuracy in computer vision tasks.

Task structure and nonlinearity jointly determine learned representational geometry

https://openreview.net/forum?id=k9t8dQ30kU

Compressor summary: The choice of activation function in a neural network affects its ability to learn meaningful representations from inputs and outputs, with Tanh networks favoring output structure and ReLU networks retaining more input information.

BECLR: Batch Enhanced Contrastive Unsupervised Few-Shot Learning

https://openreview.net/forum?id=k9SVcrmXL8

Compressor summary: The text proposes a novel approach called BECLR that improves unsupervised few-shot learning by addressing shortcomings of contrastive learning and sample bias using DyCE and OpTA modules.

Deep Generative Clustering with Multimodal Diffusion Variational Autoencoders

https://openreview.net/forum?id=k5THrhXDV3

Compressor summary: The paper proposes a multimodal VAE model that learns data clusters and improves generative performance over existing methods.

Diagnosing Transformers: Illuminating Feature Spaces for Clinical Decision-Making

https://openreview.net/forum?id=k581sTMyPt

Compressor summary: SUFO is a framework that enhances interpretability of fine-tuned transformer feature spaces for clinical decision-making using various analytic and visualization techniques, revealing insights about model trust and robustness.

TAB: Temporal Accumulated Batch Normalization in Spiking Neural Networks

https://openreview.net/forum?id=k1wlmtPGLq

Compressor summary: Temporal Accumulated Batch Normalization (TAB) is a new method for training Spiking Neural Networks that improves their performance by accounting for the temporal dynamics of neuron activity.

C-TPT: Calibrated Test-Time Prompt Tuning for Vision-Language Models via Text Feature Dispersion

https://openreview.net/forum?id=jzzEHTBFOT

Compressor summary: The paper proposes a method to improve uncertainty estimation in CLIP by optimizing prompts during test-time using a novel metric called Average Text Feature Dispersion (ATFD).

Language Modeling Is Compression

https://openreview.net/forum?id=jznbgiynus

Compressor summary: The text discusses how large language models can be viewed as powerful predictors and compressors, providing new insights into their capabilities and demonstrating their effectiveness in various tasks.

From Sparse to Soft Mixtures of Experts

https://openreview.net/forum?id=jxpsAj7ltE

Compressor summary: Soft MoE is a sparse Transformer that improves the performance and scalability of mixture of expert models while addressing their challenges.

Fair Classifiers that Abstain without Harm

https://openreview.net/forum?id=jvveGAbkVx

Compressor summary: The paper proposes a method to make classifiers abstain from predicting certain samples to achieve group fairness while maintaining original accuracy, and analyzes the feasibility and trade-offs of this approach.

Statistical Perspective of Top-K Sparse Softmax Gating Mixture of Experts

https://openreview.net/forum?id=jvtmdK69KQ

Compressor summary: The paper studies how the top-K sparse softmax gating function affects density and parameter estimations in a Gaussian mixture of experts model and proposes novel loss functions and conditions for convergence.

Language-Informed Visual Concept Learning

https://openreview.net/forum?id=juuyW8B8ig

Compressor summary: The paper proposes a method to learn a language-informed visual concept representation from pre-trained models and use it to generate images with novel compositions of visual concepts.

ETGraph: A Pioneering Dataset Bridging Ethereum and Twitter

https://openreview.net/forum?id=juE0rWGCJW

Compressor summary: ETGraph is a novel dataset that combines Ethereum transaction records with verified Twitter accounts to enhance blockchain analysis by incorporating social network data.

What Makes a Good Prune? Optimal Unstructured Pruning for Maximal Cosine Similarity

https://openreview.net/forum?id=jsvvPVVzwf

Compressor summary: The paper investigates the mechanisms of global unstructured magnitude pruning in deep neural networks, proves the optimality of L1 pruning by cosine similarity, and proposes a method to determine the optimal pruning amount based on parameter distribution kurtosis.

Deep Reinforcement Learning Guided Improvement Heuristic for Job Shop Scheduling

https://openreview.net/forum?id=jsWCmrsHHs

Compressor summary: The paper proposes a new DRL-guided heuristic for solving job-shop scheduling problems, using a graph representation scheme and a novel message-passing mechanism to encode complete solutions and evaluate them efficiently.

Unprocessing Seven Years of Algorithmic Fairness

https://openreview.net/forum?id=jr03SfWsBS

Compressor summary: The authors evaluate fairness-enhancing methods for machine learning models and find that postprocessing achieves the best trade-off between fairness and accuracy, using a novel technique called unprocessing to compare different approaches.

MINT: Evaluating LLMs in Multi-turn Interaction with Tools and Language Feedback

https://openreview.net/forum?id=jp3gWrMuIZ

Compressor summary: MINT is a benchmark that evaluates large language models' ability to solve complex tasks with multiple rounds of interactions using tools and natural language feedback, revealing interesting findings about their performance and limitations.

COPlanner: Plan to Roll Out Conservatively but to Explore Optimistically for Model-Based RL

https://openreview.net/forum?id=jnFcKjtUPN

Compressor summary: $ exttt{COPlanner}$ is a planning-driven framework for model-based reinforcement learning that improves sample efficiency and asymptotic performance by using uncertainty-aware policy-guided model predictive control to balance conservative model rollouts and optimistic environment exploration.

How Does Wild Data Provably Help OOD Detection?

https://openreview.net/forum?id=jlEjB8MVGa

Compressor summary: The paper introduces SAL, a new framework that uses unlabeled data to improve out-of-distribution detection by separating and learning from candidate outliers and labeled in-distribution data.

Space Group Constrained Crystal Generation

https://openreview.net/forum?id=jkvZ7v4OmP

Compressor summary: This paper proposes a new diffusion model (DiffCSP++) for generating crystals that considers the spacegroup constraint, which is important for describing their geometry and properties, by translating it into two tractable cases.

Self-Supervised High Dynamic Range Imaging with Multi-Exposure Images in Dynamic Scenes

https://openreview.net/forum?id=jjiOHEcS2c

Compressor summary: SelfHDR is a self-supervised method for HDR image reconstruction from multi-exposure images that avoids ghosting artifacts without needing labeled data.

Stochastic Controlled Averaging for Federated Learning with Communication Compression

https://openreview.net/forum?id=jj5ZjZsWJe

Compressor summary: The paper proposes two new communication compression algorithms for Federated Learning, SCALLION and SCAFCOM, which improve performance in terms of communication and computation complexities, and achieve fast convergence rates under arbitrary data heterogeneity.

Knowledge Fusion of Large Language Models

https://openreview.net/forum?id=jiDsk12qcz

Compressor summary: The paper proposes knowledge fusion for combining different pre-trained language models to create a more powerful model with enhanced capabilities.

Look, Remember and Reason: Grounded Reasoning in Videos with Language Models

https://openreview.net/forum?id=jhPvuc7kxB

Compressor summary: The paper proposes a method to train multi-modal LMs on low-level video tasks, enabling them to perform causal and compositional spatiotemporal reasoning using *Look, Remember, Reason* process with spatiotemporal attention.

MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning

https://openreview.net/forum?id=jenyYQzue1

Compressor summary: MuSR is a novel dataset for evaluating language models on complex multistep soft reasoning tasks in natural language narratives, created using a neurosymbolic synthetic-to-natural generation algorithm and real-world domains of reasoning.

Follow-the-Perturbed-Leader for Adversarial Bandits: Heavy Tails, Robustness, and Privacy

https://openreview.net/forum?id=jeMZi2Z9xe

Compressor summary: The paper proposes an algorithm to solve adversarial bandit problems with heavy-tailed losses, achieving optimal regret and improving privacy guarantees compared to existing methods.

Foundation Model-oriented Robustness: Robust Image Model Evaluation with Pretrained Models

https://openreview.net/forum?id=jd5GokdySz

Compressor summary: The paper proposes a new robustness measurement for image classification models that compares their performance with a surrogate oracle (a zoo of foundation models) using extended datasets with perturbed samples.

Dynamics-Informed Protein Design with Structure Conditioning

https://openreview.net/forum?id=jZPqf2G9Sw

Compressor summary: The authors propose a new method for protein generative modeling that incorporates dynamical properties using Normal Mode Analysis and shows its effectiveness by transforming an unconditional model into a conditional one without retraining.

Jointly-Learned Exit and Inference for a Dynamic Neural Network

https://openreview.net/forum?id=jX2DT7qDam

Compressor summary: The text discusses early-exiting dynamic neural networks (EDNN), which improve performance and uncertainty estimation in machine learning by connecting the gating mechanism and intermediate inference modules, addressing previous limitations.

Towards Foundation Models for Knowledge Graph Reasoning

https://openreview.net/forum?id=jVEoydFOl9

Compressor summary: ULTRA is an approach for learning universal and transferable graph representations that enable foundation models to infer on any knowledge graph with arbitrary entity and relation vocabularies.

Hybrid Distillation: Connecting Masked Autoencoders with Contrastive Learners

https://openreview.net/forum?id=jUWktnsplU

Compressor summary: The paper proposes Hybrid Distill, a method that combines Contrastive Learning and Masked Image Modeling for representation learning, improving feature diversity and discrimination.

Pushing Boundaries: Mixup's Influence on Neural Collapse

https://openreview.net/forum?id=jTSKkcbEsj

Compressor summary: Mixup, a data augmentation technique for deep neural networks, induces distinctive geometric configurations in the last-layer activations that enhance robustness and calibration.

Optimal Sample Complexity for Average Reward Markov Decision Processes

https://openreview.net/forum?id=jOm5p3q7c7

Compressor summary: The paper proposes an estimator for the optimal policy of MDPs with a sample complexity matching the lower bound in the literature, using ideas from previous works.

Analyzing and Improving OT-based Adversarial Networks

https://openreview.net/forum?id=jODehvtTDx

Compressor summary: The paper presents a unified framework for optimal transport-based generative models and proposes a novel method to improve their performance by gradually refining the generated distribution.

ASID: Active Exploration for System Identification and Reconstruction in Robotic Manipulation

https://openreview.net/forum?id=jNR6s6OSBT

Compressor summary: The authors propose a learning system that uses a small amount of real-world data to refine a simulation model and plan an accurate control strategy for robotic manipulation tasks.

Debiasing Attention Mechanism in Transformer without Demographics

https://openreview.net/forum?id=jLIUfrAcMQ

Compressor summary: The paper proposes a new debiasing approach for transformers using attention mechanism components without sensitive labels, achieving fairness and efficiency in computer vision and natural language processing tasks.

Rethinking Branching on Exact Combinatorial Optimization Solver: The First Deep Symbolic Discovery Framework

https://openreview.net/forum?id=jKhNBulNMh

Compressor summary: The paper proposes a symbolic discovery framework called Symb4CO that learns high-performance policies for exact combinatorial optimization solvers using neural networks, achieving comparable performance to GPU-based approaches with high efficiency and interpretability.

Less is More: Fewer Interpretable Region via Submodular Subset Selection

https://openreview.net/forum?id=jKTUlxo5zy

Compressor summary: The paper proposes a new image attribution method based on submodular subset selection, which improves interpretability and accuracy for different datasets and predictions.

OpenWebMath: An Open Dataset of High-Quality Mathematical Web Text

https://openreview.net/forum?id=jKHmjlpViu

Compressor summary: OpenWebMath is a new dataset for training language models to improve their quantitative reasoning skills using 14.7B tokens of mathematical webpages.

TASK PLANNING FOR VISUAL ROOM REARRANGEMENT UNDER PARTIAL OBSERVABILITY

https://openreview.net/forum?id=jJvXNpvOdM

Compressor summary: The paper proposes a new planner for embodied agents to search and rearrange objects in untidy rooms using visual input, commonsense knowledge, and graph-based representation, achieving better performance than existing methods.

BioBridge: Bridging Biomedical Foundation Models via Knowledge Graph

https://openreview.net/forum?id=jJCeMiwHdH

Compressor summary: BioBridge is a framework that connects different unimodal models for biomedical tasks using Knowledge Graphs, improving performance and generalization.

Provably Efficient UCB-type Algorithms For Learning Predictive State Representations

https://openreview.net/forum?id=jId5PXbBbX

Compressor summary: The paper proposes a novel UCB-type approach for learning sequential decision-making problems with predictive state representations, achieving computational efficiency, near-optimal policy, and accurate model estimation.

Lightweight Language Model Calibration for Open-ended Question Answering with Varied Answer Lengths

https://openreview.net/forum?id=jH67LHVOIO

Compressor summary: Litcab is a lightweight calibration mechanism for large language models that improves their accuracy and trustworthiness by adding a linear layer and training only a small fraction of parameters.

IRAD: Implicit Representation-driven Image Resampling against Adversarial Attacks

https://openreview.net/forum?id=jFa5KESW65

Compressor summary: The paper proposes a new image resampling technique to defend against adversarial attacks by creating an implicit representation and using SampleNet to generate pixel-wise shifts.

Compressing Latent Space via Least Volume

https://openreview.net/forum?id=jFJPd9kIiF

Compressor summary: Least Volume is a simple regularization technique that reduces latent dimensions in autoencoders without prior knowledge of the data dimensionality, by using decoder Lipschitz continuity, and has similar PCA-like importance ordering effects on nonlinear models.

Language Models Represent Space and Time

https://openreview.net/forum?id=jE8xbmvFin

Compressor summary: The study shows that large language models learn linear representations of space and time across multiple scales, suggesting they have a coherent world model rather than just memorizing data.

Solving Inverse Problems with Latent Diffusion Models via Hard Data Consistency

https://openreview.net/forum?id=j8hdRqOUhN

Compressor summary: ReSample is an algorithm that uses latent diffusion models to solve inverse problems with data consistency and a novel resampling scheme, achieving better results than current methods.

Feasibility-Guided Safe Offline Reinforcement Learning

https://openreview.net/forum?id=j5JvZCaDM0

Compressor summary: FISOR is a safe offline reinforcement learning method that maximizes rewards and minimizes risks by translating hard safety constraints to identifying the largest feasible region given the offline dataset.

Non-Exchangeable Conformal Risk Control

https://openreview.net/forum?id=j511LaqEeP

Compressor summary: Non-exchangeable conformal risk control is a flexible framework for providing statistical guarantees on any monotone loss function when dealing with non-exchangeable data, such as in change points or time series.

Training Graph Transformers via Curriculum-Enhanced Attention Distillation

https://openreview.net/forum?id=j4VMrwgn1M

Compressor summary: The paper proposes a curriculum-enhanced attention distillation method for semi-supervised node classification using Graph Transformers, which improves performance on seven public graph benchmarks.

ContextRef: Evaluating Referenceless Metrics for Image Description Generation

https://openreview.net/forum?id=j0ZvKSNZiP

Compressor summary: ContextRef is a benchmark for evaluating referenceless metrics that measure image description quality by assessing their alignment with human preferences and considering contextual factors.

Spoken Question Answering and Speech Continuation Using Spectrogram-Powered LLM

https://openreview.net/forum?id=izrOLJov5y

Compressor summary: The text introduces a new way to adapt large language models for question answering and speech continuation using a pre-trained speech encoder and cross-modal training on spectrograms, which improves speaker preservation, semantic coherence, and knowledge retention.

Parameter-Efficient Multi-Task Model Fusion with Partial Linearizeation

https://openreview.net/forum?id=iynRvVVAmH

Compressor summary: The paper proposes a novel method to improve multi-task fusion using partial linearization, which allows for more effective incorporation of different tasks into a single model while maintaining efficiency in fine-tuning and inference.

The Effect of Intrinsic Dataset Properties on Generalization: Unraveling Learning Differences Between Natural and Medical Images

https://openreview.net/forum?id=ixP76Y33y1

Compressor summary: This paper explores why neural networks perform differently when trained on medical versus natural images, proposes a metric for label sharpness, and shows how it affects generalization, representation learning, and adversarial robustness.

Learning Performance-Improving Code Edits

https://openreview.net/forum?id=ix7rLVHXyY

Compressor summary: The authors present a framework that adapts large language models to high-level program optimization using various techniques, achieving better performance than human programmers on a dataset of C++ code edits.

MAPE-PPI: Towards Effective and Efficient Protein-Protein Interaction Prediction via Microenvironment-Aware Protein Embedding

https://openreview.net/forum?id=itGkF993gz

Compressor summary: The paper proposes a method, MPAE-PPI, that uses microenvironment codes to predict protein-protein interactions efficiently and effectively, considering both protein sequence and structure.

DOS: Diverse Outlier Sampling for Out-of-Distribution Detection

https://openreview.net/forum?id=iriEqxFB4y

Compressor summary: Diverse Outlier Sampling (DOS) is a novel technique that selects diverse and informative outliers for OOD detection, improving the model's performance by creating a compact decision boundary between ID and OOD data.

Efficient Modulation for Vision Networks

https://openreview.net/forum?id=ip5LHJs6QX

Compressor summary: The paper introduces EfficientMod, a novel block design for efficient vision networks that improves accuracy-efficiency trade-offs and sets new state-of-the-art performance on several tasks.

Multi-View Representation is What You Need for Point-Cloud Pre-Training

https://openreview.net/forum?id=imZcqOrbig

Compressor summary: The paper proposes a novel method for pre-training 3D point clouds using 2D networks to learn consistent 2D projections and retain 3D features, achieving state-of-the-art results on various downstream tasks.

How Do Transformers Learn In-Context Beyond Simple Functions? A Case Study on Learning with Representations

https://openreview.net/forum?id=ikwEDva1JZ

Compressor summary: This paper studies how transformer-based language models learn complex in-context tasks by analyzing their representations, behaviors, and capabilities in synthetic problems.

Scaling Convex Neural Networks with Burer-Monteiro Factorization

https://openreview.net/forum?id=ikmuHqugN7

Compressor summary: The paper proposes a method to improve training convex neural networks by using a factorization technique that makes it computationally feasible and ensures global optima, while also showing that linear self-attention with many heads is free of spurious local minima.

A Neural Framework for Generalized Causal Sensitivity Analysis

https://openreview.net/forum?id=ikX6D1oM1c

Compressor summary: NeuralCSA is a neural framework for causal sensitivity analysis that can handle various sensitivity models, treatment types, and causal queries with theoretical guarantees.

FreeNoise: Tuning-Free Longer Video Diffusion via Noise Rescheduling

https://openreview.net/forum?id=ijoqFqSC7p

Compressor summary: The paper proposes FreeNoise, a tuning-free and time-efficient paradigm to enhance text-driven video generation models by rescheduling noises and injecting motions for multiple texts.

Graph Metanetworks for Processing Diverse Neural Architectures

https://openreview.net/forum?id=ijK5hyxs0n

Compressor summary: Graph Metanetworks (GMNs) use graph neural networks to process graphs representing input neural networks, generalizing to various neural architectures and accounting for symmetries and geometry of parameter spaces.

USB-NeRF: Unrolling Shutter Bundle Adjusted Neural Radiance Fields

https://openreview.net/forum?id=igfDXfMvm5

Compressor summary: USB-NeRF is a method that corrects rolling shutter distortions and recovers accurate camera motion trajectory for NeRF-based novel view synthesis using a physical image formation model.

First-order ANIL provably learns representations despite overparametrisation

https://openreview.net/forum?id=if2vRbS8Ew

Compressor summary: This paper shows that first-order adaptive network induction (ANIL), a model-agnostic meta-learning method, learns linear shared representations with overparametrized linear two-layer networks and performs well on new tasks after one gradient step.

Dissecting sample hardness: Fine-grained analysis of Hardness Characterization Methods

https://openreview.net/forum?id=icTZCUbtD6

Compressor summary: The paper introduces a fine-grained taxonomy of hardness types and presents H-CAT, a toolkit for comprehensive and quantitative benchmarking of Hardness Characterization Methods (HCMs).

Ultra-sparse network advantage in deep learning via Cannistraci-Hebb brain-inspired training with hyperbolic meta-deep community-layered epitopology

https://openreview.net/forum?id=iayEcORsGd

Compressor summary: The paper proposes a new method called Epitopological Sparse Meta-deep Learning (ESML) that uses brain-inspired learning paradigms to train ultra-sparse artificial neural networks, achieving better performance than fully connected networks on some tasks and reducing the number of neurons during training.

CausalTime: Realistically Generated Time-series for Benchmarking of Causal Discovery

https://openreview.net/forum?id=iad1yyyGme

Compressor summary: CausalTime is a pipeline that generates realistic time-series with ground truth causal graphs for evaluating and predicting TSCD algorithms' performance on real data.

Identifiable Latent Polynomial Causal Models through the Lens of Change

https://openreview.net/forum?id=ia9fKO1Vjq

Compressor summary: The paper explores nonlinear latent causal models and proposes a method to learn them with partial identifiability and consistent representations.

On the Hardness of Online Nonconvex Optimization with Single Oracle Feedback

https://openreview.net/forum?id=iZgECfyHXF

Compressor summary: This paper studies online nonconvex optimization with limited feedback, deriving lower bounds on the local regret and showing that classical algorithms are optimal in certain settings.

Contrastive Preference Learning: Learning from Human Feedback without Reinforcement Learning

https://openreview.net/forum?id=iX1RjVQODj

Compressor summary: CPL is an off-policy algorithm for learning optimal policies from human preferences without using reward functions or reinforcement learning, addressing limitations of previous methods in high-dimensional and sequential RLHF problems.

Gradual Domain Adaptation via Gradient Flow

https://openreview.net/forum?id=iTTZFKrlGV

Compressor summary: The paper introduces GGF, a domain adaptation method that generates intermediate domains with labels using Wasserstein gradient flow and Langevin algorithm, and uses classifier-based and sample-based potentials to ensure quality.

Kernelised Normalising Flows

https://openreview.net/forum?id=iTFdNLHE7k

Compressor summary: Ferumal flow is a new kernel-based normalizing flow model that can estimate and generate densities efficiently and accurately, especially when data is scarce.

Identifying Policy Gradient Subspaces

https://openreview.net/forum?id=iPWxqnt2ke

Compressor summary: This paper shows that policy gradient methods can be more efficient by using low-dimensional subspaces of gradients in reinforcement learning despite changing data distributions.

ECoFLaP: Efficient Coarse-to-Fine Layer-Wise Pruning for Vision-Language Models

https://openreview.net/forum?id=iIT02bAKzv

Compressor summary: ECoFLaP is a two-stage pruning approach for large vision-language models that uses global importance scores to achieve better model compression and improve multimodal tasks.

Disentangling Time Series Representations via Contrastive based $l$-Variational Inference

https://openreview.net/forum?id=iI7hZSczxE

Compressor summary: DisCo learns disentangled representations of home appliance electricity usage, considering attribute correlations, to help users understand and optimize consumption for a reduced carbon footprint.

Poly-View Contrastive Learning

https://openreview.net/forum?id=iHcTLIor0m

Compressor summary: The text discusses poly-view tasks in contrastive learning, where multiple related views are used to improve representation learning with information maximization and sufficient statistics, leading to better performance than SimCLR on ImageNet1k.

Rethinking Backdoor Attacks on Dataset Distillation: A Kernel Method Perspective

https://openreview.net/forum?id=iCNOK45Csv

Compressor summary: The study proposes new trigger pattern generation methods for dataset distillation that can execute effective and resilient backdoor attacks in deep learning.

Neural Contractive Dynamical Systems

https://openreview.net/forum?id=iAYIRHOYy8

Compressor summary: The text introduces a novel approach to learn stable neural networks for autonomous robot control using a variational autoencoder that encodes high-dimensional and full-pose dynamics and beats the current methods.

Negatively Correlated Ensemble Reinforcement Learning for Online Diverse Game Level Generation

https://openreview.net/forum?id=iAW2EQXfwb

Compressor summary: The paper proposes an ensemble reinforcement learning method to generate diverse game levels with a selector model and policy regularization.

Feature emergence via margin maximization: case studies in algebraic tasks

https://openreview.net/forum?id=i9wDX850jR

Compressor summary: The paper investigates why neural networks choose certain strategies for algebraic tasks and shows that margin maximization and Fourier features are key factors in their learning process.

Asymptotically Free Sketched Ridge Ensembles: Risks, Cross-Validation, and Tuning

https://openreview.net/forum?id=i9Vs5NGDpk

Compressor summary: The paper applies random matrix theory to optimize the tuning of regularization and sketching parameters in ridge regression ensembles, enabling efficient prediction risk estimation and construction of prediction intervals.

Scaling Laws for Sparsely-Connected Foundation Models

https://openreview.net/forum?id=i9K2ZWkYIP

Compressor summary: The paper investigates how reducing parameters in Transformers affects their performance on large datasets and identifies optimal sparsity levels that depend on data size and other factors.

Locality Sensitive Sparse Encoding for Learning World Models Online

https://openreview.net/forum?id=i8PjQT3Uig

Compressor summary: The paper proposes a linear regression model with nonlinear random features and a sparse locality sensitive encoding for efficient incremental updates in model-based reinforcement learning, achieving follow-the-leader behavior and outperforming neural networks in continuous control tasks.

EventRPG: Event Data Augmentation with Relevance Propagation Guidance

https://openreview.net/forum?id=i7LCsDMcZ4

Compressor summary: The paper introduces two methods to generate saliency maps for spiking neural networks, improving object and action recognition in event-based vision sensors.

On Bias-Variance Alignment in Deep Models

https://openreview.net/forum?id=i2Phucne30

Compressor summary: The paper shows that deep learning models align bias and variance at a sample level, and provides theoretical explanations for this phenomenon based on calibration and neural collapse theories.

Strategic Preys Make Acute Predators: Enhancing Camouflaged Object Detectors by Generating Camouflaged Objects

https://openreview.net/forum?id=hywpSoHwgX

Compressor summary: The paper proposes a new camouflaged object detection method that uses adversarial training to generate more challenging camouflage and a novel detection algorithm that excavates internal coherence and removes false predictions.

Graph Parsing Networks

https://openreview.net/forum?id=hv3SklibkL

Compressor summary: The Graph Parsing Network (GPN) adapts to each graph's structure to learn an efficient and informative pooling structure for graph and node classification tasks.

Information Bottleneck Analysis of Deep Neural Networks via Lossy Compression

https://openreview.net/forum?id=huGECz8dPp

Compressor summary: The paper introduces a framework to analyze deep neural networks' training process using the Information Bottleneck principle and compresses high-dimensional vectors for mutual information estimation.

InstructDET: Diversifying Referring Object Detection with Generalized Instructions

https://openreview.net/forum?id=hss35aoQ1Y

Compressor summary: InstructDET is a method for referring object detection that uses diverse user instructions and foundation models to train on more data, improving performance on standard REC datasets and object detection tasks.

Causally Aligned Curriculum Learning

https://openreview.net/forum?id=hp4yOjhwTs

Compressor summary: The paper proposes a curriculum learning method for Reinforcement Learning that uses causal reasoning to ensure optimal decision rules are shared across related tasks and the target task.

Conformal Inductive Graph Neural Networks

https://openreview.net/forum?id=homn1jOKI5

Compressor summary: Conformal prediction can create uncertainty sets for transductive node-classification models without sacrificing coverage or efficiency, and it works for both node-exchangeable and edge-exchangeable graphs.

Finetuning Text-to-Image Diffusion Models for Fairness

https://openreview.net/forum?id=hnrB5YHoYu

Compressor summary: The authors present a finetuning method for text-to-image diffusion models that aligns generated images with a target distribution, reducing various biases and allowing flexibility in controlling different attributes.

Polynormer: Polynomial-Expressive Graph Transformer in Linear Time

https://openreview.net/forum?id=hmv1LpNfXa

Compressor summary: Polynormer is a polynomial-expressive graph transformer model with linear complexity that achieves high performance on various graph datasets.

A Differentially Private Clustering Algorithm for Well-Clustered Graphs

https://openreview.net/forum?id=hkSjjs4o5d

Compressor summary: The authors present a differentially private algorithm for clustering well-clustered graphs, which performs nearly as well as non-private algorithms and outperforms pure $\epsilon$-DP algorithms.

Better Neural PDE Solvers Through Data-Free Mesh Movers

https://openreview.net/forum?id=hj9ZuNimRl

Compressor summary: The paper proposes a neural PDE solver with a novel data-free mesh adapter that can create adaptive and dynamic meshes for physical system modeling without optimal mesh data, improving accuracy and efficiency.

A path-norm toolkit for modern networks: consequences, promises and challenges

https://openreview.net/forum?id=hiHZVUIYik

Compressor summary: The paper presents a new toolkit for analyzing general neural networks with improved generalization bounds and ease of computation.

GPAvatar: Generalizable and Precise Head Avatar from Image(s)

https://openreview.net/forum?id=hgehGq2bDv

Compressor summary: GPAvatar is a framework that reconstructs 3D head avatars from images using a dynamic point-based expression field and Multi Tri-planes Attention module for precise expression control and multi-view consistency.

Generating Images in Context with Multimodal Large Language Models

https://openreview.net/forum?id=he6mX9LTyE

Compressor summary: Kosmos-G is a model that uses MLLMs and CLIP to generate images from vision-language inputs, enabling zero-shot multi-entity subject-driven generation without modifying the image decoder.

FedP3: Federated Personalized and Privacy-friendly Network Pruning under Model Heterogeneity

https://openreview.net/forum?id=hbHwZYqk9T

Compressor summary: FedP3 is a flexible federated learning framework that addresses client-side model heterogeneity by privacy-friendly network pruning.

Bongard-OpenWorld: Few-Shot Reasoning for Free-form Visual Concepts in the Real World

https://openreview.net/forum?id=hWS4MueyzC

Compressor summary: The text introduces a new benchmark for testing few-shot visual reasoning called Bongard-OpenWorld, which uses real-world images and open-world concepts to challenge current algorithms and human performance.

Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!

https://openreview.net/forum?id=hTEGyKf0dZ

Compressor summary: The text discusses how fine-tuning large language models can compromise their safety alignment, even with small numbers of adversarial examples or common datasets, and calls for better safety protocols.

Plan-Seq-Learn: Language Model Guided RL for Solving Long Horizon Robotics Tasks

https://openreview.net/forum?id=hQVCCxQrYN

Compressor summary: Plan-Seq-Learn (PSL) is a new method that uses large language models and motion planning to help robots learn high-level policies and fine-grained control actions from raw visual input, achieving over 80% success on various challenging robotics tasks.

Langevin Monte Carlo for strongly log-concave distributions: Randomized midpoint revisited

https://openreview.net/forum?id=hOxgrGM63n

Compressor summary: The paper proposes an improved method to sample from smooth strongly log-concave distributions using randomized midpoint discretization, and provides new guarantees for kinetic Langevin process with Euler discretization.

Communication-Efficient Gradient Descent-Accent Methods for Distributed Variational Inequalities: Unified Analysis and Local Updates

https://openreview.net/forum?id=hORCalGn3Z

Compressor summary: The paper proposes a unified framework for designing and analyzing efficient local training methods for distributed variational inequality problems in machine learning with improved communication complexity.

Piecewise Linear Parametrization of Policies: Towards Interpretable Deep Reinforcement Learning

https://openreview.net/forum?id=hOMVq57Ce0

Compressor summary: The paper proposes HyperCombinator, a piecewise-linear neural architecture for interpretable policies in reinforcement learning, and evaluates it in control and navigation tasks.

Lemur: Harmonizing Natural Language and Code for Language Agents

https://openreview.net/forum?id=hNhwSmtXRh

Compressor summary: Lemur and Lemur-Chat are versatile language agents that excel in natural language and coding tasks, bridging the gap between humans and machines.

Contrastive Learning is Spectral Clustering on Similarity Graph

https://openreview.net/forum?id=hLZQTFGToA

Compressor summary: This paper analyzes contrastive learning's equivalence to spectral clustering and introduces a new loss function called Kernel-InfoNCE for improved performance in CLIP model.

True Knowledge Comes from Practice: Aligning Large Language Models with Embodied Environments via Reinforcement Learning

https://openreview.net/forum?id=hILVmJ4Uvu

Compressor summary: TWOSOME is a framework that uses large language models for efficient decision-making in embodied environments via reinforcement learning, improving sample efficiency and generalization.

Adaptive Window Pruning for Efficient Local Motion Deblurring

https://openreview.net/forum?id=hI18CDyadM

Compressor summary: The paper proposes a local motion deblurring method that adaptively restores high-resolution images using a vision Transformer with adaptive window pruning, achieving better performance and efficiency than existing methods.

Label-free Node Classification on Graphs with Large Language Models (LLMs)

https://openreview.net/forum?id=hESD2NJFg8

Compressor summary: The text introduces LLM-GNN, a label-free node classification pipeline that combines GNNs and LLMs to annotate nodes efficiently and improve performance.

P$^2$OT: Progressive Partial Optimal Transport for Deep Imbalanced Clustering

https://openreview.net/forum?id=hD3sGVqPsr

Compressor summary: The paper proposes a novel framework for deep imbalanced clustering that uses pseudo-label generation based on optimal transport problems to learn from data with class distribution imbalance.

PolyVoice: Language Models for Speech to Speech Translation

https://openreview.net/forum?id=hCrFG9cyuC

Compressor summary: PolyVoice is a language model-based framework for speech-to-speech translation that uses three decoder-only language models to improve translation and audio quality, outperforming the state-of-the-art encoder-decoder model.

PhyloGFN: Phylogenetic inference with generative flow networks

https://openreview.net/forum?id=hB7SlfEmze

Compressor summary: The paper presents a new method, PhyloGFN, that uses generative flow networks to sample from the complex space of phylogenetic trees and evaluate its performance on real data.

Greedy Sequential Execution: Solving Homogeneous and Heterogeneous Cooperative Tasks with a Unified Framework

https://openreview.net/forum?id=hB2hXtxIPH

Compressor summary: Greedy Sequential Execution (GSE) is a new approach to handle cooperative agent tasks by learning optimal policies for both homogeneous and heterogeneous scenarios using individual utility functions and greedy marginal contributions.

FedWon: Triumphing Multi-domain Federated Learning Without Normalization

https://openreview.net/forum?id=hAYHmV1gM8

Compressor summary: FedWon is a novel method for federated learning that improves accuracy by eliminating normalization layers and reparameterizing convolution layers, especially in multi-domain scenarios with diverse feature distributions.

Multi-Source Diffusion Models for Simultaneous Music Generation and Separation

https://openreview.net/forum?id=h922Qhkmx1

Compressor summary: The authors propose a generative model for music creation and separation using diffusion and Dirac likelihood methods, achieving competitive results on a standard dataset.

Neural Network-Based Score Estimation in Diffusion Models: Optimization and Generalization

https://openreview.net/forum?id=h8GeqOxtd4

Compressor summary: The paper proposes a mathematical framework to analyze and improve gradient-based learning of the score function in diffusion models using a novel neural network architecture and kernel regression techniques.

Tractable Probabilistic Graph Representation Learning with Graph-Induced Sum-Product Networks

https://openreview.net/forum?id=h7nOCxFsPg

Compressor summary: GSPNs are a new probabilistic framework for learning graphs that can efficiently answer probabilistic queries, outperforming popular neural models in various scenarios.

Active Retrosynthetic Planning Aware of Route Quality

https://openreview.net/forum?id=h7DGnWGeos

Compressor summary: ARP is a framework that balances query costs and route quality evaluation for retrosynthetic planning, outperforming existing approaches by 6.2% in route quality and reducing query cost by 12.8%.

VQGraph: Rethinking Graph Representation Space for Bridging GNNs and MLPs

https://openreview.net/forum?id=h6Tz85BqRI

Compressor summary: VQGraph uses a structure-aware tokenizer to learn a new graph representation space that enables knowledge transfer from GNN to MLP for efficient graph data analysis.

Self-Supervised Dataset Distillation for Transfer Learning

https://openreview.net/forum?id=h57gkDO2Yg

Compressor summary: The paper proposes a new way to distill an unlabeled dataset into small synthetic samples for efficient self-supervised learning, addressing biases in gradient estimation and minimizing randomness.

Improved sampling via learned diffusions

https://openreview.net/forum?id=h4pNROsO06

Compressor summary: The paper proposes a unifying framework for deep learning-based methods to sample from unnormalized target densities using the Schrödinger bridge problem, and introduces a new loss function called log-variance loss that improves performance.

DFormer: Rethinking RGBD Representation Learning for Semantic Segmentation

https://openreview.net/forum?id=h1sFUGlI09

Compressor summary: DFormer is a novel framework that pretrains the backbone using image-depth pairs, improves RGB-D encoding with a sequence of tailored blocks, and achieves state-of-the-art performance on RGB-D segmentation tasks with less computational cost.

Understanding Certified Training with Interval Bound Propagation

https://openreview.net/forum?id=h05eQniJsQ

Compressor summary: The authors investigate why interval bound propagation (IBP), an imprecise but effective method for training robust neural networks, works so well and find that it provides strong regularization and exact bounds in certain conditions.

Language Model Beats Diffusion - Tokenizer is key to visual generation

https://openreview.net/forum?id=gzqrANCF4g

Compressor summary: The paper introduces a new video tokenizer that enables large language models to outperform diffusion models in image and video generation tasks, as well as in video compression and action recognition.

Path Choice Matters for Clear Attributions in Path Methods

https://openreview.net/forum?id=gzYgsZgwXa

Compressor summary: The paper introduces a new method, SAMP, to generate more rigorous and clear interpretations of deep neural networks by finding near-optimal manipulation paths based on the Concentration Principle.

Automatic Functional Differentiation in JAX

https://openreview.net/forum?id=gzT61ziSCu

Compressor summary: The paper presents a method to differentiate higher-order functions using JAX, enabling automatic differentiation with a simple syntax and functional gradients as callable python functions.

Learning Polynomial Problems with $SL(2, \mathbb{R})$-Equivariance

https://openreview.net/forum?id=gyfXuRfxW2

Compressor summary: Neural networks can solve positivity optimization and certification problems for polynomials faster and accurately while leveraging the $SL(2,\mathbb{R})$ group equivariance.

Diffeomorphic Mesh Deformation via Efficient Optimal Transport for Cortical Surface Reconstruction

https://openreview.net/forum?id=gxhRR8vUQb

Compressor summary: The paper proposes a novel metric for learning mesh deformation using sliced Wasserstein distance on meshes represented as probability measures, which overcomes limitations of the set-based approach and provides a fast statistical rate for approximating surfaces.

ZeRO++: Extremely Efficient Collective Communication for Large Model Training

https://openreview.net/forum?id=gx2BT0a9MQ

Compressor summary: ZeRO++ improves the efficiency and scalability of large language models by reducing communication volume and overheads in low-bandwidth settings.

Learning Large DAGs is Harder than you Think: Many Losses are Minimal for the Wrong DAG

https://openreview.net/forum?id=gwbQ2YwLhD

Compressor summary: The scale of data affects the performance of structure learning algorithms, especially in non-linear cases, which can have significant implications in fields like medicine and biology.

Dual Associated Encoder for Face Restoration

https://openreview.net/forum?id=gwDuW7Ok5f

Compressor summary: DAEFR is a novel framework that uses a dual-branch approach to restore facial details from low-quality images by extracting domain-specific information and promoting synergy between the branches.

CausalLM is not optimal for in-context learning

https://openreview.net/forum?id=guRNebwZBb

Compressor summary: The paper analyzes how prefix language models (prefixLM) outperform causal language models (causalLM) in in-context learning, showing that prefixLM converges to the optimal solution while causalLM does not.

Generative Judge for Evaluating Alignment

https://openreview.net/forum?id=gtkFw6sZGS

Compressor summary: The paper introduces **Auto-J**, a generative judge model with 13B parameters that evaluates the performance of Large Language Models on tasks aligned with human needs, such as brainstorming and email writing.

Efficient Subgraph GNNs by Learning Effective Selection Policies

https://openreview.net/forum?id=gppLqZLQeY

Compressor summary: The paper proposes a new method called Policy-Learn that learns how to select a small subset of subgraphs for graph neural networks, reducing computational complexity and improving performance.

SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation

https://openreview.net/forum?id=gn0mIhQGNM

Compressor summary: SalUn is a novel machine unlearning method that improves accuracy, stability, and cross-domain applicability by focusing on specific model weights rather than the entire model.

Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory

https://openreview.net/forum?id=gmg7t8b4s0

Compressor summary: The paper proposes CONFAIDE, a benchmark to test privacy reasoning in large language models, and shows that current models reveal private information in inappropriate contexts.

VertiBench: Advancing Feature Distribution Diversity in Vertical Federated Learning Benchmarks

https://openreview.net/forum?id=glwwbaeKm2

Compressor summary: Vertical Federated Learning (VFL) benefits from new evaluation metrics and datasets that consider feature importance and correlation, improving algorithm performance assessment.

Confronting Reward Model Overoptimization with Constrained RLHF

https://openreview.net/forum?id=gkfUvn0fLU

Compressor summary: The paper studies overoptimization in composite reward models for language quality and proposes a constrained reinforcement learning approach to balance component models and improve evaluation performance.

DyVal: Graph-informed Dynamic Evaluation of Large Language Models

https://openreview.net/forum?id=gjfOL9z5Xr

Compressor summary: DyVal is a dynamic evaluation protocol that uses graphs to generate challenging reasoning tasks for large language models, revealing their limitations and potential for improvement.

Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs

https://openreview.net/forum?id=gjeQKFxFpZ

Compressor summary: The text presents a systematic framework for black-box uncertainty estimation in LLMs using prompting, sampling, and aggregation methods. It evaluates the framework on calibration and failure prediction tasks and identifies key insights and challenges.

PRES: Toward Scalable Memory-Based Dynamic Graph Neural Networks

https://openreview.net/forum?id=gjXor87Xfy

Compressor summary: Memory-based Dynamic Graph Neural Networks (MDGNNs) use memory to capture long-term dependencies, but struggle with temporal discontinuity in large temporal batches; the proposed PRES method mitigates this issue and enables faster training without sacrificing performance.

FedImpro: Measuring and Improving Client Update in Federated Learning

https://openreview.net/forum?id=giU9fYGTND

Compressor summary: This paper proposes FedImpro, a method to reduce client drift in federated learning by constructing similar conditional distributions for local training, improving generalization and gradient similarity.

Attention Satisfies: A Constraint-Satisfaction Lens on Factual Errors of Language Models

https://openreview.net/forum?id=gfFVATffPd

Compressor summary: The paper studies how Transformer-based LLMs handle factual constraints when generating text, and proposes SAT Probe, a method that can predict errors and improve reliability.

OmniControl: Control Any Joint at Any Time for Human Motion Generation

https://openreview.net/forum?id=gd0lAEtWso

Compressor summary: OmniControl is a novel method for generating realistic human motion with flexible spatial control signals over different joints using diffusion-based models and analytic guidance.

Feature Collapse

https://openreview.net/forum?id=gctmyMiPHH

Compressor summary: Feature collapse is a phenomenon where similar learning task roles get similar representations, improving generalization and interpretability.

Logical Languages Accepted by Transformer Encoders with Hard Attention

https://openreview.net/forum?id=gbrHZq07mq

Compressor summary: The text studies how two types of self-attention mechanisms in transformer encoders recognize different formal languages, and shows their limitations and capabilities.

Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models

https://openreview.net/forum?id=gU58d5QeGv

Compressor summary: Würstchen is a text-to-image synthesis model that uses a compact semantic image representation to guide the diffusion process, achieving competitive performance with much lower cost and faster inference than existing models.

Safety-Tuned LLaMAs: Lessons From Improving the Safety of Large Language Models that Follow Instructions

https://openreview.net/forum?id=gT5hALch9z

Compressor summary: The paper discusses the importance of safety in instruction-tuned language models and shows that adding a small number of safety examples can improve their safety without compromising their capabilities or helpfulness.

FlashFFTConv: Efficient Convolutions for Long Sequences with Tensor Cores

https://openreview.net/forum?id=gPKTTAfYBp

Compressor summary: FlashFFTConv optimizes FFT convolution for long sequences, speeding up reasoning tasks and enabling larger models with the same compute budget.

A Private Watermark for Large Language Models

https://openreview.net/forum?id=gMLQwKDY3N

Compressor summary: The paper proposes a private watermarking algorithm for large language models that uses different neural networks for generating and detecting watermarks, enhancing security and accuracy while minimizing parameters.

LUT-GEMM: Quantized Matrix Multiplication based on LUTs for Efficient Inference in Large-Scale Generative Language Models

https://openreview.net/forum?id=gLARhFLE0F

Compressor summary: The paper introduces LUT-GEMM, a kernel for quantized matrix multiplication that eliminates the resource-intensive dequantization process and reduces computational costs, achieving significant acceleration in NLP model generation.

A Simple Romance Between Multi-Exit Vision Transformer and Token Reduction

https://openreview.net/forum?id=gJeYtRuguR

Compressor summary: METR introduces multi-exit loss to improve the reliability of ViTs' attention scores for token reduction, making them more efficient and effective.

Neural Optimal Transport with General Cost Functionals

https://openreview.net/forum?id=gIiz7tBtYZ

Compressor summary: The paper presents a new neural network method for computing OT plans that can use extra information, work with diverse costs, and handle high-dimensional data, while providing error guarantees and an example of class-preserving mapping.

Modulated Phase Diffusor: Content-Oriented Feature Synthesis for Detecting Unknown Objects

https://openreview.net/forum?id=gHAr7ZA1OL

Compressor summary: The paper proposes a method to generate out-of-distribution data for training object detectors using phase information from in-distribution data and two branches for content manipulation.

IMPUS: Image Morphing with Perceptually-Uniform Sampling Using Diffusion Models

https://openreview.net/forum?id=gG38EBe2S8

Compressor summary: The paper presents a new method for image morphing called IMPUS, which uses diffusion-based models, latent space interpolation, and perceptually-uniform sampling to produce realistic and smooth results.

Gene Regulatory Network Inference in the Presence of Dropouts: a Causal View

https://openreview.net/forum?id=gFR4QwK53h

Compressor summary: The authors propose a causal graphical model to characterize dropout error in single-cell RNA sequencing data and show that test-wise deletion can be used to integrate existing structure learning methods for gene regulatory network inference without introducing spurious relations.

Efficient Backpropagation with Variance Controlled Adaptive Sampling

https://openreview.net/forum?id=gEwKAZZmSw

Compressor summary: The paper introduces a method called VCAS that reduces computational load in neural network training by sampling unimportant computations, while controlling additional variance to maintain accuracy.

BESA: Pruning Large Language Models with Blockwise Parameter-Efficient Sparsity Allocation

https://openreview.net/forum?id=gC6JTEU3jl

Compressor summary: BESA is a novel technique that prunes large language models by targeting error reduction at the transformer block level and allocating sparsity differently, resulting in less performance degradation and faster computation time.

Evaluating the Zero-shot Robustness of Instruction-tuned Language Models

https://openreview.net/forum?id=g9diuvxN6D

Compressor summary: The paper studies how sensitive instruction-tuned language models are to different phrasings of instructions and proposes a method to improve their robustness using soft prompt embedding parameters.

Adaptive Rational Activations to Boost Deep Reinforcement Learning

https://openreview.net/forum?id=g90ysX1sVs

Compressor summary: The authors propose using adaptable activation functions in reinforcement learning to enhance neural plasticity and improve performance in changing environments, such as Atari games.

On the Parameterization of Second-Order Optimization Effective towards the Infinite Width

https://openreview.net/forum?id=g8sGBSQjYk

Compressor summary: The paper proposes a new parameterization for second-order optimization that improves feature learning in larger deep neural networks and allows sharing hyperparameters across models of different sizes.

Flow Matching on General Geometries

https://openreview.net/forum?id=g7ohDlTITL

Compressor summary: Riemannian Flow Matching (RFM) is a new method for training continuous normalizing flows on manifolds that works without simulation, divergence computation, and uses a simple premetric for target vector fields.

Follow-Up Differential Descriptions: Language Models Resolve Ambiguities for Image Classification

https://openreview.net/forum?id=g6rZtxaXRm

Compressor summary: FuDD improves image classification by generating class descriptions that better differentiate between ambiguous classes using a large language model.

InfoCon: Concept Discovery with Generative and Discriminative Informativeness

https://openreview.net/forum?id=g6eCbercEc

Compressor summary: The paper presents a method for autonomously discovering manipulation concepts from unlabeled robotic demonstrations using a trainable codebook that encodes sub-trajectories based on their informativeness.

A Progressive Training Framework for Spiking Neural Networks with Learnable Multi-hierarchical Model

https://openreview.net/forum?id=g52tgL8jy6

Compressor summary: The paper introduces a new spiking neural network model (LM-H) that improves energy efficiency and information processing by dynamically adjusting membrane-related factors, and proposes a novel training algorithm and framework for it.

Neural Active Learning Beyond Bandits

https://openreview.net/forum?id=g1S72T3FGc

Compressor summary: The paper proposes new neural network algorithms for active learning that reduce the impact of the number of classes on performance and cost, and shows their advantages in theory and practice.

Adaptive Federated Learning with Auto-Tuned Clients

https://openreview.net/forum?id=g0mlwqs8pi

Compressor summary: $\Delta$-SGD is a simple method for adjusting step sizes in federated learning that adapts to local smoothness, improving performance in different scenarios.

Global Optimality for Non-linear Constrained Restoration Problems via Invexity

https://openreview.net/forum?id=fyTPWfXtcc

Compressor summary: The paper introduces invex constrained optimization theory to ensure global optima in non-convex constrained inverse problems while maintaining reconstruction quality.

Crystalformer: Infinitely Connected Attention for Periodic Structure Encoding

https://openreview.net/forum?id=fxQiecl9HB

Compressor summary: Crystalformer is a simple transformer-based encoder architecture that predicts physical properties of materials from their crystal structures more accurately than existing models with fewer parameters and better interpretability.

Improving Generalization of Alignment with Human Preferences through Group Invariant Learning

https://openreview.net/forum?id=fwCoLe3TAX

Compressor summary: The authors propose a method to improve language model-based AI assistants' performance and generalization across various domains by automatically classifying data into groups, optimizing the policy for challenging groups, and adaptively adjusting exploration space.

Improved Probabilistic Image-Text Representations

https://openreview.net/forum?id=ft1mr3WlGM

Compressor summary: The paper proposes a new probabilistic image-text matching method with improved embeddings and optimization techniques that outperforms existing methods on various benchmarks.

Can We Evaluate Domain Adaptation Models Without Target-Domain Labels?

https://openreview.net/forum?id=fszrlQ2DuP

Compressor summary: The paper introduces Transfer Score, a novel metric for unsupervised evaluation of UDA models, which measures spatial uniformity, transferability, and discriminability. The metric helps to choose the best UDA method, tune hyperparameters, and find optimal checkpoints without target labels.

Tensor Trust: Interpretable Prompt Injection Attacks from an Online Game

https://openreview.net/forum?id=fsW7wJGLBd

Compressor summary: Tensor Trust is a dataset with over 100,000 prompt injection attacks and defenses for testing the robustness and security of large language models.

Are Human-generated Demonstrations Necessary for In-context Learning?

https://openreview.net/forum?id=frRDT6EOhg

Compressor summary: This paper proposes self-contemplation prompting strategy (SEC), which eliminates human-generated demonstrations in In-context Learning, and shows that large language models can perform well with only their own decision making capacity.

Sample-efficient Learning of Infinite-horizon Average-reward MDPs with General Function Approximation

https://openreview.net/forum?id=fq1wNrC2ai

Compressor summary: FLOP is a novel algorithmic framework that tackles average-reward Markov decision processes with function approximation using confidence sets and low-switching policy updates, achieving sublinear regret in various complex AMDP models.

Circuit Component Reuse Across Tasks in Transformer Language Models

https://openreview.net/forum?id=fpoAYV6Wsk

Compressor summary: The authors show that insights from mechanistic interpretability can generalize across tasks, demonstrating this by reusing and modifying a circuit for different language tasks with significant improvements and explanations.

Efficient Dynamics Modeling in Interactive Environments with Koopman Theory

https://openreview.net/forum?id=fkrYDQaHOJ

Compressor summary: Using Koopman theory, the authors propose a method to efficiently and accurately model nonlinear dynamics in interactive environments for long-range prediction, planning, and reinforcement learning.

Ito Diffusion Approximation of Universal Ito Chains for Sampling, Optimization and Boosting

https://openreview.net/forum?id=fjpfCOV4ru

Compressor summary: This paper studies a general class of Markov chains derived from stochastic differential equations with arbitrary noise, inexact drift and diffusion coefficients, and improves existing bounds on their distance to the corresponding differential equations.

New Insight of Variance reduce in Zero-Order Hard-Thresholding: Mitigating Gradient Error and Expansivity Contradictions

https://openreview.net/forum?id=fjf3YenThE

Compressor summary: The paper introduces a new algorithm that improves the performance of hard-thresholding for $\ell_0$ constrained optimization using zeroth-order gradients and variance reduction, and shows its usefulness in different applications.

Stochastic Gradient Descent for Gaussian Processes Done Right

https://openreview.net/forum?id=fj2E5OcLFn

Compressor summary: The paper introduces an efficient stochastic dual gradient descent algorithm for optimizing Gaussian process regression and shows its competitive performance against other methods.

GAIA: a benchmark for General AI Assistants

https://openreview.net/forum?id=fibxvahvs3

Compressor summary: GAIA is a benchmark for testing general AI assistants' abilities by asking real-world questions that are simple for humans but challenging for most advanced AIs, aiming to measure their robustness and proficiency in various skills.

The Alignment Problem from a Deep Learning Perspective: A Position Paper

https://openreview.net/forum?id=fh8EYKFKns

Compressor summary: The paper discusses the challenges of aligning artificial general intelligence (AGI) systems with human interests, as they might learn deceptive or power-seeking behaviors from pretrained deep models.

REFACTOR: Learning to Extract Theorems from Proofs

https://openreview.net/forum?id=fgKjiVrm6u

Compressor summary: The paper introduces REFACTOR, a neural network method that learns to extract and use modular theorems from mathematical proofs, improving proof quality and efficiency.

From Posterior Sampling to Meaningful Diversity in Image Restoration

https://openreview.net/forum?id=ff2g30cZxj

Compressor summary: The paper proposes a way to make image restoration methods produce diverse and meaningful outputs by modifying the original method or adding post-processing steps.

Federated Q-Learning: Linear Regret Speedup with Low Communication Cost

https://openreview.net/forum?id=fe6ANBxcKM

Compressor summary: The paper proposes two Q-Learning algorithms for federated reinforcement learning with linear regret speedup and low communication cost, using event-triggered synchronization, novel step size selection, and new concentration inequalities.

AuG-KD: Anchor-Based Mixup Generation for Out-of-Domain Knowledge Distillation

https://openreview.net/forum?id=fcqWJ8JgMR

Compressor summary: AuG-KD is a new method that helps transfer knowledge from large models without access to their training data by aligning student-domain data with teacher domain using uncertainty and generative techniques, improving performance in real-world scenarios.

A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models

https://openreview.net/forum?id=farT6XXntP

Compressor summary: This study proposes ALMA, a novel fine-tuning approach for LLMs that improves translation performance without relying on large parallel data sets, outperforming previous models with only 7B or 13B parameters.

Consistent Multi-Class Classification from Multiple Unlabeled Datasets

https://openreview.net/forum?id=fW7DOHDQvF

Compressor summary: Weakly supervised learning methods aim to train accurate classifiers from unlabeled data and class priors, with RCM being a new approach that improves risk consistency and purifies supervision information.

An interpretable error correction method for enhancing code-to-code translation

https://openreview.net/forum?id=fVxIEHGnVT

Compressor summary: kNN-ECD and kNN-ECS methods improve program translation accuracy by using error correction and interpreting translated codes with k-nearest neighbors.

Nougat: Neural Optical Understanding for Academic Documents

https://openreview.net/forum?id=fUtxNAKpdV

Compressor summary: Nougat is a Visual Transformer model that converts scientific PDFs into markup language for better accessibility and understanding of scientific knowledge.

Align With Purpose: Optimize Desired Properties in CTC Models with a General Plug-and-Play Framework

https://openreview.net/forum?id=fUGhVYPVRM

Compressor summary: AWP is a plug-and-play framework that enhances desired properties in CTC-trained models by complementing the CTC loss with an additional loss term that prioritizes alignments according to the property, improving ASR latency and WER on large-scale data.

Unifying Feature and Cost Aggregation with Transformers for Dense Correspondence

https://openreview.net/forum?id=fQHb1uZzl7

Compressor summary: The paper presents a Transformer-based network that combines feature and cost aggregation for dense matching tasks using self- and cross-attention layers, leading to improved semantic and geometric matching results.

$\mathcal{B}$-Coder: On Value-Based Deep Reinforcement Learning for Program Synthesis

https://openreview.net/forum?id=fLf589bx1f

Compressor summary: The text describes a new approach to program synthesis using value-based reinforcement learning and large language models, which produces accurate code with minimal reward engineering effort.

Convergence of Bayesian Bilevel Optimization

https://openreview.net/forum?id=fLXpXa7iiz

Compressor summary: The paper provides a theoretical guarantee for combining Bayesian optimization with stochastic gradient descent in bilevel frameworks, showing sublinear regret bounds and suggesting optimal configurations for generalization.

Magnitude Invariant Parametrizations Improve Hypernetwork Learning

https://openreview.net/forum?id=fJNnerz6iH

Compressor summary: The text describes a problem with training hypernetworks that affects their convergence and proposes a solution called Magnitude Invariant Parametrizations (MIP) to improve stability and speed up convergence.

CoLiDE: Concomitant Linear DAG Estimation

https://openreview.net/forum?id=fGAIgO75dG

Compressor summary: CoLiDE is a new method for learning directed acyclic graphs from data that estimates scale and noise levels, improving performance and stability over existing methods.

Near-Optimal Solutions of Constrained Learning Problems

https://openreview.net/forum?id=fDaLmkdSKU

Compressor summary: This paper studies how dual ascent methods for constrained machine learning problems affect the optimality and feasibility of primal iterates, and provides a sub-optimality bound based on the problem's parametrization and objective curvature.

Zero-Shot Robustification of Zero-Shot Models

https://openreview.net/forum?id=fCeUoDr9Tq

Compressor summary: RoboShot improves the robustness and performance of pretrained models in zero-shot tasks by using language models to enhance their embeddings without supervision or fine-tuning.

Pre-training LiDAR-based 3D Object Detectors through Colorization

https://openreview.net/forum?id=fB1iiH9xo7

Compressor summary: Grounded Point Colorization (GPC) is a pre-training method that teaches models to colorize LiDAR point clouds, improving 3D object detection for self-driving cars with limited labeled data.

Project and Probe: Sample-Efficient Adaptation by Interpolating Orthogonal Features

https://openreview.net/forum?id=f6CBQYxXvr

Compressor summary: Pro$^2$ is a sample-efficient method that learns diverse features and interpolates them to adapt to distribution shifts using a small target dataset.

UniAdapter: Unified Parameter-Efficient Transfer Learning for Cross-modal Modeling

https://openreview.net/forum?id=f5H8WGLQm5

Compressor summary: UniAdapter is a method to adapt pre-trained vision-language models to various downstream tasks with minimal computational and storage costs, outperforming full fine-tuning in most cases.

LiDAR: Sensing Linear Probing Performance in Joint Embedding SSL Architectures

https://openreview.net/forum?id=f3g5XpL9Kb

Compressor summary: LiDAR is a metric for evaluating joint embedding representations by measuring the rank of the LDA matrix associated with a surrogate self-supervised learning task.

Cycle Consistency Driven Object Discovery

https://openreview.net/forum?id=f1xnBr4WD6

Compressor summary: The paper proposes a method to improve object-centric deep learning models by optimizing the association of objects with distinct slots and applying them to downstream tasks.

Convolution Meets LoRA: Parameter Efficient Finetuning for Segment Anything Model

https://openreview.net/forum?id=ezscMer8L0

Compressor summary: Conv-LoRA is a method that improves the performance of the Segment-Anything Model for image segmentation in specialized domains by adding lightweight convolutional parameters.

AnyText: Multilingual Visual Text Generation and Editing

https://openreview.net/forum?id=ezBH9WE9s2

Compressor summary: AnyText is a diffusion-based multilingual visual text generation and editing model that produces accurate and coherent text in images, achieving state-of-the-art performance on a new large-scale dataset called AnyWord-3M.

Sample-Efficiency in Multi-Batch Reinforcement Learning: The Need for Dimension-Dependent Adaptivity

https://openreview.net/forum?id=ey3GhWXQ97

Compressor summary: The paper investigates how adaptivity affects sample-efficiency in reinforcement learning by analyzing lower bounds on the number of batches needed for policy evaluation and best-policy identification under linear function approximation.

SweetDreamer: Aligning Geometric Priors in 2D diffusion for Consistent Text-to-3D

https://openreview.net/forum?id=extpNXo6hB

Compressor summary: The text proposes a method to improve text-to-3D generation by aligning 2D geometric priors in diffusion models with well-defined 3D shapes, resolving the multi-view inconsistency problem and achieving high generalizability.

LLM-grounded Video Diffusion Models

https://openreview.net/forum?id=exKHibougU

Compressor summary: LLM-grounded Video Diffusion (LVD) uses a large language model to generate dynamic scene layouts based on text inputs, which then guide a video diffusion model for improved spatiotemporal video generation.

Maximum Likelihood Estimation is All You Need for Well-Specified Covariate Shift

https://openreview.net/forum?id=eoTCKKOgIs

Compressor summary: Classical MLE using source data is near-optimal for covariate shift in well-specified models, but MWLE can be better in misspecified settings.

Towards Energy Efficient Spiking Neural Networks: An Unstructured Pruning Framework

https://openreview.net/forum?id=eoSeaK4QJo

Compressor summary: The paper proposes a novel pruning framework that combines unstructured weight and neuron pruning to enhance energy efficiency in deep spiking neural networks without sacrificing much accuracy.

Combining Spatial and Temporal Abstraction in Planning for Better Generalization

https://openreview.net/forum?id=eo9dHwtTFt

Compressor summary: Skipper is a model-based RL agent that learns and decomposes tasks into subtasks using hindsight, enabling better generalization and sparse decision-making.

Consistency Training with Learnable Data Augmentation for Graph Anomaly Detection with Limited Supervision

https://openreview.net/forum?id=elMKXvhhQ9

Compressor summary: ConsisGAD is a new model for graph anomaly detection that uses consistency training, learnable data augmentation, and homophily distribution to handle limited supervision and class imbalance.

Fully Hyperbolic Convolutional Neural Networks for Computer Vision

https://openreview.net/forum?id=ekz1hN5QNh

Compressor summary: The paper introduces HCNN, a novel fully hyperbolic convolutional neural network for computer vision tasks, which leverages the benefits of hyperbolic geometry to better represent intricate structures in images.

Understanding In-Context Learning in Transformers and LLMs by Learning to Learn Discrete Functions

https://openreview.net/forum?id=ekeyCgeRfC

Compressor summary: The study investigates Transformers' and attention-free models' abilities to learn algorithms for various function classes, showing that Transformers adaptively select sample-efficient algorithms while LLMs can generalize beyond their training data.

SAN: Inducing Metrizability of GAN with Discriminative Normalized Linear Layer

https://openreview.net/forum?id=eiF7TU1E8E

Compressor summary: The paper proposes and tests a new GAN training method called Slicing Adversarial Network (SAN) that uses sliced optimal transport to ensure the generator's distribution matches the target distribution more closely.

Turning large language models into cognitive models

https://openreview.net/forum?id=eiC4BKypf1

Compressor summary: The paper shows how finetuning large language models on psychological data allows them to accurately model and predict human behavior in various decision-making tasks.

SEPT: Towards Efficient Scene Representation Learning for Motion Prediction

https://openreview.net/forum?id=efeBC1sQj9

Compressor summary: SEPT is a self-supervised learning framework that helps autonomous vehicles predict motions in complex traffic scenes by capturing kinematics, road structure, and agent interactions using masking-reconstruction tasks.

Multimodal Web Navigation with Instruction-Finetuned Foundation Models

https://openreview.net/forum?id=efFmBWioSc

Compressor summary: WebGUM is a novel web navigation agent that leverages vision-language models and offline training to surpass previous methods and some online agents on multiple tasks.

Enhancing Neural Subset Selection: Integrating Background Information into Set Representations

https://openreview.net/forum?id=eepoE7iLpL

Compressor summary: The paper proposes a new information aggregation module for neural network-based set functions that incorporates superset information to improve performance in tasks like drug discovery.

G$^2$N$^2$ : Weisfeiler and Lehman go grammatical

https://openreview.net/forum?id=eZneJ55mRO

Compressor summary: The paper proposes a method to connect algebraic languages and GNNs using CFGs, reduces redundancies in rules, and shows that their derived G$^2$N$^2$ model performs better than other 3-WL GNNs on various tasks.

Generative Modeling of Regular and Irregular Time Series Data via Koopman VAEs

https://openreview.net/forum?id=eY7sLb0dVF

Compressor summary: Koopman VAE (KVAE) is a new generative framework for realistic time series data generation that leverages Koopman theory and spectral tools to incorporate domain knowledge and improve stability, outperforming existing GAN and VAE methods on various benchmarks.

SaNN: Simple Yet Powerful Simplicial-aware Neural Networks

https://openreview.net/forum?id=eUgS9Ig8JG

Compressor summary: The paper proposes SaNN, a scalable graph representation learning model with constant run-time and memory, which is powerful, equivariant, and achieves state-of-the-art performance.

Annealing Self-Distillation Rectification Improves Adversarial Training

https://openreview.net/forum?id=eT6oLkm1cm

Compressor summary: ADR is a simple method that generates soft labels for adversarial training to improve model robustness without extra computation or pre-trained models.

Diffusion in Diffusion: Cyclic One-Way Diffusion for Text-Vision-Conditioned Generation

https://openreview.net/forum?id=ePOjNlOjLC

Compressor summary: The text proposes a method called Cyclic One-Way Diffusion (COW) to control the direction of diffusion phenomenon in image generation, allowing for more versatile and precise customization without additional training.

Neural Fourier Transform: A General Approach to Equivariant Representation Learning

https://openreview.net/forum?id=eOCvA8iwXH

Compressor summary: Neural Fourier Transform (NFT) is a framework for learning data structure using symmetry without assuming explicit knowledge of the group's action on data, and shows theoretical and practical benefits.

Unknown Domain Inconsistency Minimization for Domain Generalization

https://openreview.net/forum?id=eNoiRal5xi

Compressor summary: UDIM is a new objective for domain generalization that minimizes inconsistency between source and unknown domains by perturbing data and aligning flat minima.

On Trajectory Augmentations for Off-Policy Evaluation

https://openreview.net/forum?id=eMNN0wIyVw

Compressor summary: The paper proposes an offline trajectory augmentation approach to enhance off-policy evaluation in human-involved scenarios using sub-trajectory mining and diversifying state-action space coverage.

Combinatorial Bandits for Maximum Value Reward Function under Value-Index Feedback

https://openreview.net/forum?id=eMHn77ZKOp

Compressor summary: The paper proposes an algorithm for a new combinatorial multi-armed bandit problem with index feedback and provides a regret bound that compares favorably to existing methods.

DIFFTACTILE: A Physics-based Differentiable Tactile Simulator for Contact-rich Robotic Manipulation

https://openreview.net/forum?id=eJHnSg783t

Compressor summary: DIFFTACTILE is a high-fidelity physics-based tactile simulation system for robotic manipulation with differentiable components and contact modeling.

Blending Imitation and Reinforcement Learning for Robust Policy Improvement

https://openreview.net/forum?id=eJ0dzPJq1F

Compressor summary: RPI is a novel algorithm that interleaves imitation learning and reinforcement learning to improve policy robustness and sample efficiency, adapting to different oracles and states.

Merge, Then Compress: Demystify Efficient SMoE with Hints from Its Routing Policy

https://openreview.net/forum?id=eFWG9Cy3WK

Compressor summary: The paper proposes M-SMoE and MC-SMoE, two methods that improve the scalability of neural networks using SMoE by merging and compressing experts based on their activation frequency and routing statistics.

Reward-Free Curricula for Training Robust World Models

https://openreview.net/forum?id=eCGpNGDeNu

Compressor summary: The paper introduces WAKER, a method to train robust world models from reward-free exploration by selecting environments with high world model error, leading to better adaptation, efficiency, and generality.

PORF: POSE RESIDUAL FIELD FOR ACCURATE NEURAL SURFACE RECONSTRUCTION

https://openreview.net/forum?id=eBeECjacpw

Compressor summary: The paper introduces a novel implicit representation, PoRF, which uses an MLP to regress pose updates and enhances pose supervision using epipolar geometry loss, leading to improved camera pose accuracy and neural surface reconstruction in challenging real-world scenarios.

PixArt-$\alpha$: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis

https://openreview.net/forum?id=eAKmQPe3m1

Compressor summary: PixArt-α is a fast and high-quality text-to-image model that uses less training data and reduces CO2 emissions compared to existing models.

Lion Secretly Solves a Constrained Optimization: As Lyapunov Predicts

https://openreview.net/forum?id=e4xS9ZarDr

Compressor summary: Lion, a new AI optimizer, combines elements from existing algorithms and has promising results, but lacks theoretical foundation; a study provides a Lyapunov function to analyze its dynamics and improve it.

The Unreasonable Effectiveness of Linear Prediction as a Perceptual Metric

https://openreview.net/forum?id=e4FG5PJ9uC

Compressor summary: LASI is a perceptual similarity metric based on pixel-level weighted least squares that performs well on image quality assessment tasks without training data or deep features.

ArchLock: Locking DNN Transferability at the Architecture Level with a Zero-Cost Binary Predictor

https://openreview.net/forum?id=e2YOVTenU9

Compressor summary: ArchLock is a novel algorithm that uses neural architecture search and zero-cost proxies to create DNN models with low transferability, reducing the vulnerability of current defense strategies.

Symmetric Single Index Learning

https://openreview.net/forum?id=e1vqloonRy

Compressor summary: The paper studies how gradient flow can learn single-index models in symmetric neural networks and prove its efficiency using an adapted information exponent.

Weight Selection for Model Initialization

https://openreview.net/forum?id=dyrGMhicMw

Compressor summary: Weight selection is a method that initializes smaller models with a subset of weights from a larger pretrained model, improving their performance and reducing training time.

DiffusionNAG: Predictor-guided Neural Architecture Generation with Diffusion Models

https://openreview.net/forum?id=dyG2oLJYyX

Compressor summary: DiffusionNAG is a novel conditional Neural Architecture Generation framework based on diffusion models that generates task-optimal architectures efficiently for diverse tasks by sampling from a region more likely to satisfy desired properties.

On the Scalability and Memory Efficiency of Semidefinite Programs for Lipschitz Constant Estimation of Neural Networks

https://openreview.net/forum?id=dwzLn78jq7

Compressor summary: The paper proposes a new method for estimating Lipschitz constants in deep learning using eigenvalue optimization, which is more scalable and efficient than standard SDP algorithms.

HyperHuman: Hyper-Realistic Human Generation with Latent Structural Diffusion

https://openreview.net/forum?id=duyA42HlCK

Compressor summary: The authors propose HyperHuman, a model that generates realistic and diverse human images by capturing correlations between appearance and structure at multiple scales using a large-scale dataset and a unified framework.

BadEdit: Backdooring Large Language Models by Model Editing

https://openreview.net/forum?id=duZANm2ABX

Compressor summary: The BadEdit attack framework injects backdoors into large language models using a lightweight editing technique, requiring minimal data and time, while preserving model performance and robustness.

Teaching Arithmetic to Small Transformers

https://openreview.net/forum?id=dsUB4bst9S

Compressor summary: Small transformers can efficiently learn arithmetic operations using the next-token prediction objective when trained on high-quality, instructive data with intermediate step results.

Object-Aware Inversion and Reassembly for Image Editing

https://openreview.net/forum?id=dpcVXiMlcv

Compressor summary: OIR is a new image editing paradigm that adapts the number of inversion steps for each editing pair to achieve optimal results and fine-grained editing of objects in images.

Minimum width for universal approximation using ReLU networks on compact domain

https://openreview.net/forum?id=dpDw5U04SU

Compressor summary: The minimum width needed for deep neural networks to approximate any function using ReLU-like activations is $\max\{d_x,d_y,2\}$, where $d_x$ and $d_y$ are the input and output dimensions.

Periodicity Decoupling Framework for Long-term Series Forecasting

https://openreview.net/forum?id=dp27P5HBBt

Compressor summary: The paper proposes a novel framework called PDF that models 2D temporal variations for long-term time series forecasting by decoupling, capturing short- and long-term variations, and aggregating them.

SemiReward: A General Reward Model for Semi-supervised Learning

https://openreview.net/forum?id=dnqPvUjyRI

Compressor summary: SemiReward is a framework that uses reward scores to evaluate and filter high-quality pseudo labels in semi-supervised learning, improving performance and convergence speed.

Towards Optimal Feature-Shaping Methods for Out-of-Distribution Detection

https://openreview.net/forum?id=dm8e7gsH0d

Compressor summary: The paper proposes an optimization framework to improve out-of-distribution detection in deep learning models by optimizing feature shaping functions using only in-distribution data.

Noise-free Score Distillation

https://openreview.net/forum?id=dlIMcmlAdk

Compressor summary: NFSD improves text-to-content generation by reducing noise and over-smoothing using a smaller CFG scale, while maintaining realism and prompt compliance.

Retro-fallback: retrosynthetic planning in an uncertain world

https://openreview.net/forum?id=dl0u4ODCuW

Compressor summary: The paper introduces a new algorithm, retro-fallback, for retrosynthesis that considers uncertainty and improves the chances of creating a viable synthesis plan.

Neural Snowflakes: Universal Latent Graph Inference via Trainable Latent Geometries

https://openreview.net/forum?id=djM3WzpOmK

Compressor summary: The paper introduces a new deep learning architecture called neural snowflake that improves latent graph inference by adaptively implementing fractal-like metrics on high-dimensional data, without the curse of dimensionality and random search.

SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

https://openreview.net/forum?id=di52zR8xgf

Compressor summary: SDXL is a text-to-image model with a larger backbone, new conditioning schemes, and a refinement model that improves its quality and competes with state-of-the-art image generators.

Decoding Natural Images from EEG for Object Recognition

https://openreview.net/forum?id=dhLIno8FmH

Compressor summary: The paper presents a self-supervised method to learn image representations from EEG signals and achieve above-chance results in object recognition tasks, while exploring the biological plausibility of the approach.

Relaxing the Additivity Constraints in Decentralized No-Regret High-Dimensional Bayesian Optimization

https://openreview.net/forum?id=de1218PoEl

Compressor summary: DumBO is a new decentralized Bayesian Optimization algorithm that relaxes assumptions on the function to optimize and reduces over-exploration, making it suitable for high-dimensional problems with complex structure.

Reward Model Ensembles Help Mitigate Overoptimization

https://openreview.net/forum?id=dcjtMYkpXx

Compressor summary: Ensemble-based conservative optimization methods can reduce overoptimization and improve the performance of large language models fine-tuned with human feedback.

Stable Anisotropic Regularization

https://openreview.net/forum?id=dbQH9AOVd5

Compressor summary: The paper introduces I-STAR, a novel regularization method for LLMs that accurately measures and adjusts isotropy in embedding space during training to improve model performance.

Generative Human Motion Stylization in Latent Space

https://openreview.net/forum?id=daEqXJ0yZo

Compressor summary: The paper presents a novel generative model for human motion stylization that uses pretrained autoencoders' latent space to learn probabilistic style space from motions, enabling users to select or sample style cues while preserving content.

Human Motion Diffusion as a Generative Prior

https://openreview.net/forum?id=dTpbEdN9kr

Compressor summary: The paper presents three composition methods based on diffusion priors for generating human motion with more control and flexibility, addressing challenges like long sequences, two-person interactions, and detailed editing.

Neural Polynomial Gabor Fields for Macro Motion Analysis

https://openreview.net/forum?id=dTlKCQuuxP

Compressor summary: The paper introduces Phase-based neural polynomial Gabor fields (Phase-PGF), a method to analyze large motions in 2D and 3D dynamic scenes using low-dimensional time-varying phases.

Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition

https://openreview.net/forum?id=dQVtTdsvZH

Compressor summary: The content-motion latent diffusion model (CMD) is an efficient video generation method that combines a pretrained image diffusion model and a lightweight motion model, achieving high quality and fast speed.

Fast, Expressive $\mathrm{SE}(n)$ Equivariant Networks through Weight-Sharing in Position-Orientation Space

https://openreview.net/forum?id=dPHLbUqGbr

Compressor summary: The paper presents a geometric method to optimize edge attributes for convolutional networks using homogeneous spaces, leading to an efficient 3D group convolutional network that achieves state-of-the-art results on various benchmarks.

Closing the Curious Case of Neural Text Degeneration

https://openreview.net/forum?id=dONpC9GL1o

Compressor summary: The paper explains why truncation sampling heuristics like nucleus sampling work well for language generation and proposes a new strategy based on softmax bottleneck to improve expressiveness without relying on thresholds.

Efficacy of Dual-Encoders for Extreme Multi-label Classification

https://openreview.net/forum?id=dNe1T0Ahby

Compressor summary: The paper explores the performance of dual-encoder models on extreme multi-label classification tasks and proposes a new loss function to optimize for Recall@k metrics.

TUVF: Learning Generalizable Texture UV Radiance Fields

https://openreview.net/forum?id=dN4vpVTvWX

Compressor summary: The paper proposes Texture UV Radiance Fields, a method for generating and controlling realistic textures on 3D models independent of their shapes.

Topological data analysis on noisy quantum computers

https://openreview.net/forum?id=dLrhRIMVmB

Compressor summary: NISQ-TDA is a quantum machine learning algorithm for high-dimensional data analysis with proven speedup and low noise sensitivity.

Robustness of AI-Image Detectors: Fundamental Limits and Practical Attacks

https://openreview.net/forum?id=dLoAdIKENc

Compressor summary: The paper analyzes the strengths and weaknesses of AI-image detectors, especially watermarking methods, and their vulnerabilities to attacks that can falsely flag real images as fake or vice versa.

Prompt Learning with Quaternion Networks

https://openreview.net/forum?id=dKlxDx2SoS

Compressor summary: The paper proposes QNet, a method that uses quaternion networks to enhance multimodal pre-trained models for better semantic alignment across diverse modalities.

Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models

https://openreview.net/forum?id=dKl6lMwbCy

Compressor summary: The choice of feedback protocol, either ratings or rankings, significantly affects how well large language models are aligned with human values and intents.

ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs

https://openreview.net/forum?id=dHng2O0Jjr

Compressor summary: The authors introduce ToolLLM, a framework for enhancing large language models' tool-use capabilities by constructing a dataset (ToolBench) and fine-tuning LLaMA with a neural API retriever.

GenCorres: Consistent Shape Matching via Coupled Implicit-Explicit Shape Generative Models

https://openreview.net/forum?id=dGH4kHFKFj

Compressor summary: GenCorres is a new unsupervised method for joint shape matching that learns a mesh generator to fit a large synthetic shape collection, preserving local geometric structures and outperforming existing techniques.

PlaSma: Procedural Knowledge Models for Language-based Planning and Re-Planning

https://openreview.net/forum?id=dFcXJgnrGB

Compressor summary: The paper proposes PlaSma, a two-pronged approach to enhance small language models with procedural knowledge and planning capabilities, and shows they can outperform larger models in planning and replanning tasks.

Regularized Robust MDPs and Risk-Sensitive MDPs: Equivalence, Policy Gradient, and Sample Complexity

https://openreview.net/forum?id=dEz3ge8QSo

Compressor summary: The paper introduces a new risk-sensitive MDP formulation, shows its equivalence with regularized robust MDPs, derives policy gradient theorems, proves global convergence, and proposes an efficient offline learning algorithm for specific problems.

Learning to Reject with a Fixed Predictor: Application to Decontextualization

https://openreview.net/forum?id=dCHbFDsCZz

Compressor summary: The paper proposes a new problem formulation and algorithm for classification with reject option, shows strong consistency guarantees, and evaluates it on decontextualization task achieving high coverage.

Tool-Augmented Reward Modeling

https://openreview.net/forum?id=d94x0gWTUX

Compressor summary: Themis is a tool-augmented preference modeling approach that improves large language models' functionality by enabling them to use external tools for tasks like arithmetic, code execution, and factual lookup.

Small-scale proxies for large-scale Transformer training instabilities

https://openreview.net/forum?id=d8w0pmvXbZ

Compressor summary: The paper investigates training instability in Transformer-based models at different scales and suggests interventions to improve stability and loss performance across learning rate variations.

Energy-guided Entropic Neural Optimal Transport

https://openreview.net/forum?id=d6tUsZeVs7

Compressor summary: The paper introduces a novel method that combines energy-based models and entropy-regularized optimal transport to improve generative modeling and provides theoretical and empirical results.

Spurious Feature Diversification Improves Out-of-distribution Generalization

https://openreview.net/forum?id=d6H4RBi7RH

Compressor summary: The study examines how weight space ensembles like WiSE-FT use diverse spurious features for better out-of-distribution generalization and proposes BANG to address overconfidence issues in such methods.

Dynamic Layer Tying for Parameter-Efficient Transformers

https://openreview.net/forum?id=d4uL2MSe0z

Compressor summary: The paper proposes a method to reduce the number of trainable parameters in deep transformer networks using Reinforcement Learning for layer selection and weight sharing.

LLaMA-Adapter: Efficient Fine-tuning of Large Language Models with Zero-initialized Attention

https://openreview.net/forum?id=d4UiXAHN2W

Compressor summary: The paper introduces LLaMA-Adapter, a method to efficiently fine-tune LLaMA for instruction following using self-attention layers with learnable zero gating, achieving similar performance to Alpaca with fewer parameters and extending to multi-modal tasks.

Bounds on Representation-Induced Confounding Bias for Treatment Effect Estimation

https://openreview.net/forum?id=d3xKPQVjSc

Compressor summary: The paper proposes a new method to estimate how much dimensionality reduction in CATE estimation can introduce bias due to confounding.

Procedural Fairness Through Decoupling Objectionable Data Generating Components

https://openreview.net/forum?id=cxfPefbu1s

Compressor summary: The paper addresses the problem of disguised procedural unfairness in automated decision-making and proposes a framework to separate objectionable from neutral data generating components.

On Differentially Private Federated Linear Contextual Bandits

https://openreview.net/forum?id=cuAxSHcsSX

Compressor summary: The paper proposes a new algorithmic framework for collaborative federated linear contextual bandits under differential privacy that addresses privacy, regret, and communication cost issues.

Generalized Policy Iteration using Tensor Approximation for Hybrid Control

https://openreview.net/forum?id=csukJcpYDe

Compressor summary: The paper introduces TTPI, an ADP algorithm that uses low-rank tensor approximation to efficiently control hybrid systems in robotics.

End-to-End (Instance)-Image Goal Navigation through Correspondence as an Emergent Phenomenon

https://openreview.net/forum?id=cphhnHjCvC

Compressor summary: The paper presents a method to improve goal-oriented visual navigation in complex scenes using pretext tasks that help learn correspondence and goal detection from image pairs.

Decongestion by Representation: Learning to Improve Economic Welfare in Marketplaces

https://openreview.net/forum?id=coIaBY8EVF

Compressor summary: The study investigates how platforms can learn to display information in a way that reduces congestion and improves social welfare in online marketplaces by proposing a differentiable proxy of welfare based on consumer choices.

Learning the greatest common divisor: explaining transformer predictions

https://openreview.net/forum?id=cmcD05NPKa

Compressor summary: The paper trains small transformers to find the greatest common divisor (GCD) of two numbers using lists of divisors and shows that their predictions are explainable, but performance depends on the training distribution of GCD.

Large Language Models are Efficient Learners of Noise-Robust Speech Recognition

https://openreview.net/forum?id=ceATjGPTUD

Compressor summary: The authors propose a method to improve automatic speech recognition by using a large language model that learns to correct errors and remove noise from speech transcriptions.

Robust NAS benchmark under adversarial training: assessment, theory, and beyond

https://openreview.net/forum?id=cdUpf6t6LZ

Compressor summary: This paper introduces a new data set for evaluating robustness against adversarial attacks and provides a generalization theory for searching architectures that balance clean and robust accuracy.

Can LLM-Generated Misinformation Be Detected?

https://openreview.net/forum?id=ccxD4mtkTU

Compressor summary: LLM-generated misinformation is harder to detect than human-written misinformation, which may increase its potential harm.

Approximating Nash Equilibria in Normal-Form Games via Stochastic Optimization

https://openreview.net/forum?id=cc8h3I3V4E

Compressor summary: The paper introduces a new loss function for finding approximate Nash equilibria in normal-form games using non-convex stochastic optimization techniques, which perform better than existing methods.

Diffusion Models for Multi-Task Generative Modeling

https://openreview.net/forum?id=cbv0sBIZh9

Compressor summary: The paper proposes a unified multi-task diffusion model that can generate different types of data simultaneously and improve generalization by aggregating information from multiple tasks.

Beyond Accuracy: Evaluating Self-Consistency of Code LLMs

https://openreview.net/forum?id=caW7LdAALh

Compressor summary: The paper introduces IdentityChain, a framework to evaluate and improve the self-consistency of code large language models (LLMs) across different tasks, as it affects their trustworthiness.

Uncertainty Quantification via Stable Distribution Propagation

https://openreview.net/forum?id=cZttUMTiPL

Compressor summary: The text introduces a new method for using stable probability distributions in neural networks, which improves uncertainty quantification and outperforms existing methods like moment matching.

Latent 3D Graph Diffusion

https://openreview.net/forum?id=cXbnGtO0NZ

Compressor summary: The paper introduces latent 3D graph diffusion, a fast and effective way to generate 3D graphs with symmetry and quality in a low-dimensional latent space, and applies it to conditional generation of molecules with better properties and validity.

Few-Shot Detection of Machine-Generated Text using Style Representations

https://openreview.net/forum?id=cWiEN1plhJ

Compressor summary: The paper proposes a new method to detect machine-written text using style features from human authors, which works even against state-of-the-art language models like Llama 2 and GPT-4.

State Representation Learning Using an Unbalanced Atlas

https://openreview.net/forum?id=cWdAYDLmPa

Compressor summary: The paper presents a new self-supervised learning method (DIM-UA) that leverages an unbalanced atlas to improve dimensionality reduction and outperforms existing approaches on the AtariARI benchmark.

Improved Analysis of Sparse Linear Regression in Local Differential Privacy Model

https://openreview.net/forum?id=cVUOnF7iVp

Compressor summary: The paper proposes a novel non-interactive differential privacy algorithm for sparse linear regression with lower bounds and upper bounds on estimation error.

GlucoBench: Curated List of Continuous Glucose Monitoring Datasets with Prediction Benchmarks

https://openreview.net/forum?id=cUSNs8nGaV

Compressor summary: CGM-based glucose trajectory prediction can improve diabetes management, but current evaluation on small datasets hinders reproducibility and adoption.

Chain-of-Knowledge: Grounding Large Language Models via Dynamic Knowledge Adapting over Heterogeneous Sources

https://openreview.net/forum?id=cPgh4gWZlz

Compressor summary: Chain-of-knowledge (CoK) is a novel framework that enhances large language models by dynamically incorporating grounding information from heterogeneous sources, resulting in more factual and rational answers for knowledge-intensive questions.

Towards Faithful XAI Evaluation via Generalization-Limited Backdoor Watermark

https://openreview.net/forum?id=cObFETcoeW

Compressor summary: The paper proposes a new backdoor-based method for evaluating saliency-based representation visualization methods in XAI, by minimizing the effects of a generalization-limited backdoor watermark on benign and poisoned samples.

ADDP: Learning General Representations for Image Recognition and Generation with Alternating Denoising Diffusion Process

https://openreview.net/forum?id=cMPm8YFXZe

Compressor summary: The paper proposes an ADDP framework that combines pixels and VQ tokens for representation learning, achieving competitive performance on image generation and recognition tasks.

Exploring Effective Stimulus Encoding via Vision System Modeling for Visual Prostheses

https://openreview.net/forum?id=cKAUvMePUN

Compressor summary: The paper proposes an optimization framework for generating stimulation patterns for visual prostheses using a retinal network, phosphene model, and primary vision system network, which can simulate the whole process of visual signals processing from external scenes to visual perception in the cortex.

Deep Orthogonal Hypersphere Compression for Anomaly Detection

https://openreview.net/forum?id=cJs4oE4m9Q

Compressor summary: The paper proposes a novel deep anomaly detection model that improves hypersphere learning using an orthogonal projection layer, and extends it to graph-level anomaly detection with better performance than existing methods.

Free Lunches in Auxiliary Learning: Exploiting Auxiliary Labels with Negligibly Extra Inference Cost

https://openreview.net/forum?id=cINwAhrgLf

Compressor summary: The paper proposes a method to use auxiliary labels from an independent task to improve primary task performance while keeping single task inference cost, using a flexible asymmetric network structure that can be removed during inference.

Output-Domain Focused Inductive Bias on Latent Feature Clusters in Visual Classification

https://openreview.net/forum?id=cH3oufN8Pl

Compressor summary: The paper proposes Output-Domain focused Biasing (ODB), a training strategy that improves image classification by incorporating human label knowledge into neural networks, without external resources.

Enhancing Neural Training via a Correlated Dynamics Model

https://openreview.net/forum?id=c9xsaASm9L

Compressor summary: CMD is a technique that groups correlated parameters in neural networks to efficiently represent and enhance their training dynamics.

BadChain: Backdoor Chain-of-Thought Prompting for Large Language Models

https://openreview.net/forum?id=c93SBwz1Ma

Compressor summary: BadChain is a novel backdoor attack against large language models using chain-of-thought prompting that alters the model's output when triggered by a specific query.

Solving Challenging Math Word Problems Using GPT-4 Code Interpreter with Code-based Self-Verification

https://openreview.net/forum?id=c8McWs4Av0

Compressor summary: The paper explores how code usage frequency affects GPT-4 Code Interpreter's math reasoning skills and proposes a new prompting method, CSV, to enhance its self-verification abilities and accuracy on MATH dataset.

Decoupling Weighing and Selecting for Integrating Multiple Graph Pre-training Tasks

https://openreview.net/forum?id=c85tdYOOju

Compressor summary: The paper proposes a novel framework called Weigh And Select (WAS) for integrating multiple graph pre-training tasks using decoupled siamese networks.

Democratizing Fine-grained Visual Recognition with Large Language Models

https://openreview.net/forum?id=c7DND1iIgb

Compressor summary: FineR uses large language models to reason about fine-grained visual categories without needing expert annotations, by translating part-level image attributes to text and leveraging the LLM's world knowledge.

METRA: Scalable Unsupervised RL with Metric-Aware Abstraction

https://openreview.net/forum?id=c5pwL0Soay

Compressor summary: METRA is a novel unsupervised reinforcement learning objective that learns diverse behaviors by exploring a compact latent space connected to the state space, enabling scalability in high-dimensional environments.

GAFormer: Enhancing Timeseries Transformers Through Group-Aware Embeddings

https://openreview.net/forum?id=c56TWtYp0W

Compressor summary: The paper introduces GAFormer, a transformer-based model that learns spatiotemporal structure from multivariate time series data and achieves state-of-the-art performance on various tasks.

Denoising Diffusion Step-aware Models

https://openreview.net/forum?id=c43FGk8Pcg

Compressor summary: DDSM is a novel framework that adapts neural network sizes to the importance of each generative step in denoising diffusion models, improving efficiency and preserving quality.

The All-Seeing Project: Towards Panoptic Visual Recognition and Understanding of the Open World

https://openreview.net/forum?id=c2R7ajodcI

Compressor summary: The All-Seeing (AS) project is a dataset and model for recognizing and understanding a wide range of real-world concepts using human feedback and efficient models, aiming to advance vision-language artificial general intelligence research.

Zero-Shot Robotic Manipulation with Pre-Trained Image-Editing Diffusion Models

https://openreview.net/forum?id=c0chJTSbci

Compressor summary: SuSIE is a method that uses an image editing diffusion model to help generalist robots recognize and reason about novel objects and scenarios, achieving better generalization than conventional language-conditioned policies.

Prioritized Soft Q-Decomposition for Lexicographic Reinforcement Learning

https://openreview.net/forum?id=c0MyyXyGfn

Compressor summary: The paper proposes a novel algorithm, PSQD, that solves continuous space lexicographic multi-objective RL problems by learning and adapting subtask solutions under priorities, enabling reuse and offline learning.

AlignDiff: Aligning Diverse Human Preferences via Behavior-Customisable Diffusion Model

https://openreview.net/forum?id=bxfKIYfHyx

Compressor summary: AlignDiff is a framework that uses reinforcement learning and human feedback to customize agent behaviors according to diverse and changing human preferences in locomotion tasks.

Sharpness-Aware Data Poisoning Attack

https://openreview.net/forum?id=bxITGFPVWh

Compressor summary: The paper introduces Sharpness-Aware Data Poisoning Attack (SAPA), which exploits the loss landscape sharpness to optimize data poisoning for maximum impact on trained models.

Structural Fairness-aware Active Learning for Graph Neural Networks

https://openreview.net/forum?id=bvjcMvMn7B

Compressor summary: The paper proposes SCARCE, a framework to improve GNNs' performance and fairness in semi-supervised node classification by incorporating graph structure and mitigating structural bias in active learning.

AnomalyCLIP: Object-agnostic Prompt Learning for Zero-shot Anomaly Detection

https://openreview.net/forum?id=buC4E91xZE

Compressor summary: AnomalyCLIP adapts CLIP for accurate zero-shot anomaly detection by learning object-agnostic text prompts that capture generic normality and abnormality in images, enabling generalized recognition across different domains.

Efficient Planning with Latent Diffusion

https://openreview.net/forum?id=btpgDo4u4j

Compressor summary: The paper proposes LatentDiffuser, a framework that learns and plans in continuous latent action spaces using score-based diffusion models for efficient offline reinforcement learning.

RAPPER: Reinforced Rationale-Prompted Paradigm for Natural Language Explanation in Visual Question Answering

https://openreview.net/forum?id=bshfchPM9H

Compressor summary: The text introduces a new method, called Rapper, for improving visual question answering by generating explanations that are informative, plausible, and grounded in both language and vision facts.

Mediator Interpretation and Faster Learning Algorithms for Linear Correlated Equilibria in General Sequential Games

https://openreview.net/forum?id=bsKMPAFHO7

Compressor summary: The paper connects linear deviations in extensive-form games to untimed communication deviations and develops faster algorithms for computing linear correlated equilibria.

Stabilizing Backpropagation Through Time to Learn Complex Physics

https://openreview.net/forum?id=bozbTTWcaw

Compressor summary: The paper proposes a new vector field for physics simulations that improves optimization by avoiding the issues caused by the rotational gradient field used in recurrent learning setups.

Adaptive Stochastic Gradient Algorithm for Black-box Multi-Objective Learning

https://openreview.net/forum?id=bm1JVsVZVu

Compressor summary: The paper proposes a new algorithm, ASMG, for black-box multi-objective optimization that uses stochastic gradient approximation and adaptive weights to optimize multiple objectives efficiently and proves its convergence rate in both convex and non-convex scenarios.

A Simple Interpretable Transformer for Fine-Grained Image Classification and Analysis

https://openreview.net/forum?id=bkdWThqE6q

Compressor summary: The paper introduces INTR, a Transformer-based image classification method that generates class-specific queries to help each class find its patterns in an image, enabling interpretable predictions.

MBR and QE Finetuning: Training-time Distillation of the Best and Most Expensive Decoding Methods

https://openreview.net/forum?id=bkNx3O0sND

Compressor summary: The authors propose efficient methods to finetune NMT models using external data and show significant quality improvements over the base model.

Multilinear Operator Networks

https://openreview.net/forum?id=bbCL5aRjUx

Compressor summary: MONet is a new model for image recognition that uses only multilinear operators, without activation functions, and achieves performance comparable to modern neural networks.

R-MAE: Regions Meet Masked Autoencoders

https://openreview.net/forum?id=ba84RDHFnz

Compressor summary: The paper proposes a new method called masked region autoencoding to learn from groups of pixels in images and improve self-supervised learning for tasks like detection and segmentation.

Exploring Diffusion Time-steps for Unsupervised Representation Learning

https://openreview.net/forum?id=bWzxhtl1HP

Compressor summary: The text explores using Denoising Diffusion Probabilistic Model to learn modular attributes from data by adding noise and learning features to compensate for the lost attributes at each time-step.

FITS: Modeling Time Series with $10k$ Parameters

https://openreview.net/forum?id=bWcnvZ3qMb

Compressor summary: FITS is a lightweight model that uses complex frequency domain interpolation to perform well in time series analysis tasks, and it can run on edge devices.

Transferring Learning Trajectories of Neural Networks

https://openreview.net/forum?id=bWNJFD1l8M

Compressor summary: The paper proposes a method to transfer learning trajectories between DNNs to speed up training and improve performance.

Latent Trajectory Learning for Limited Timestamps under Distribution Shift over Time

https://openreview.net/forum?id=bTMMNT7IdW

Compressor summary: SDE-EDG is a novel method that collects infinitely fined-grid evolving trajectories of data distributions using stochastic differential equations to capture distribution shifts in time-varying systems, outperforming existing approaches.

Future Language Modeling from Temporal Document History

https://openreview.net/forum?id=bRLed9prWC

Compressor summary: The authors introduce the task of future language modeling, which predicts texts in the future based on a temporal history of texts, and show that it improves upon non-temporal language models.

Detecting Generated Text via Rewriting

https://openreview.net/forum?id=bQWE2UqXmf

Compressor summary: Large language models (LLMs) can be used to detect AI-generated text by rewriting it and measuring the editing distance, improving detection scores across various domains.

Phenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinement

https://openreview.net/forum?id=bNt7oajl2a

Compressor summary: The text studies how well language models do inductive reasoning using a technique called iterative hypothesis refinement, which shows they are good at proposing rules but struggle to apply them, revealing differences between human and model reasoning.

Boundary Denoising for Video Activity Localization

https://openreview.net/forum?id=bLpUtGyf9g

Compressor summary: DenosieLoc is an encoder-decoder model that learns to localize video activities by denoising action spans with controlled noise, leading to better understanding and faster convergence.

UC-NERF: Neural Radiance Field for under-calibrated multi-view cameras

https://openreview.net/forum?id=bLKcCe7hYh

Compressor summary: UC-NeRF is a new method for creating realistic images from multiple cameras by correcting color inconsistencies and improving pose calibration.

Facing the Elephant in the Room: Visual Prompt Tuning or Full finetuning?

https://openreview.net/forum?id=bJx4iOIOxn

Compressor summary: The paper analyzes Visual Prompt Tuning (VPT) as a transfer learning technique and identifies scenarios where it outperforms full-finetuning, as well as the underlying reasons for its success.

Meta Inverse Constrained Reinforcement Learning: Convergence Guarantee and Generalization Analysis

https://openreview.net/forum?id=bJ3gFiwRgi

Compressor summary: The paper proposes a method for learning reward functions and constraints of an expert from few demonstrations by using meta-learning over related tasks, and shows theoretical and experimental results.

Latent Representation and Simulation of Markov Processes via Time-Lagged Information Bottleneck

https://openreview.net/forum?id=bH6T0Jjw5y

Compressor summary: This paper introduces Time-lagged Information Bottleneck (T-IB), a method to simplify complex systems for accurate simulation using information theory principles.

Understanding In-Context Learning from Repetitions

https://openreview.net/forum?id=bGGYcvw8mp

Compressor summary: The paper investigates how surface features and token co-occurrence reinforcement affect in-context learning in LLMs, revealing its mechanisms and limitations.

COSA: Concatenated Sample Pretrained Vision-Language Foundation Model

https://openreview.net/forum?id=bDkisS75zy

Compressor summary: COSA is a vision-language model that pretrains on image-text pairs to jointly model visual contents and event-level temporal cues, achieving state-of-the-art results on multiple semantic tasks.

Learning model uncertainty as variance-minimizing instance weights

https://openreview.net/forum?id=bDWXhzZT40

Compressor summary: The paper proposes a method to capture and use predictive uncertainty for various tasks using an auxiliary network trained with a meta-objective that minimizes dropout variance, resulting in significant performance gains across different datasets and scenarios.

Supervised Knowledge Makes Large Language Models Better In-context Learners

https://openreview.net/forum?id=bAMPOUF227

Compressor summary: The text proposes a framework to improve large language models' generalizability and factuality by using task-specific fine-tuned models during inference, and demonstrates its effectiveness with experiments and resources.

Pareto Deep Long-Tailed Recognition: A Conflict-Averse Solution

https://openreview.net/forum?id=b66P1u0k15

Compressor summary: The paper proposes a Pareto deep long-tailed recognition method (PLOT) that uses a multi-objective optimization framework and dynamic re-balancing to address gradient conflicts among categories and improve overall performance.

Representation Deficiency in Masked Language Modeling

https://openreview.net/forum?id=b3l0piOrGU

Compressor summary: MAE-LM is a new pretraining method that excludes $ exttt{[MASK]}$ tokens from the encoder and improves text encoder's performance on downstream tasks.

KITAB: Evaluating LLMs on Constraint Satisfaction for Information Retrieval

https://openreview.net/forum?id=b3kDP3IytM

Compressor summary: The paper introduces KITAB, a new dataset to test language models' ability to answer book-related queries with constraints, and analyzes their limitations in different scenarios.

Meta-Learning Priors Using Unrolled Proximal Neural Networks

https://openreview.net/forum?id=b3Cu426njo

Compressor summary: Meta-learning improves by using a novel prior representation model that leverages algorithm unrolling for more expressive and interpretable priors.

Retrieval-augmented Vision-Language Representation for Fine-grained Recognition

https://openreview.net/forum?id=b2UlHeyyC0

Compressor summary: The paper introduces a method to enhance CLIP's performance on fine-grained tasks by using external memory to refine its embeddings, instead of expanding the pre-training data.

The Hidden Language of Diffusion Models

https://openreview.net/forum?id=awWpHnEJDw

Compressor summary: Conceptor is a method that reveals how diffusion models represent textual concepts in their images by decomposing them into interpretable elements, discovering structures and connections between concepts, and linking them to visual impacts.

Expressivity of ReLU-Networks under Convex Relaxations

https://openreview.net/forum?id=awHTL3Hpto

Compressor summary: The study explores how different convex relaxations affect the ability of ReLU neural networks to encode and analyze various classes of convex functions, revealing limitations in some cases.

Implicit Maximum a Posteriori Filtering via Adaptive Optimization

https://openreview.net/forum?id=auUngos7eR

Compressor summary: The paper proposes a new way of doing Bayesian filtering using optimization instead of matrix operations, which works well for high-dimensional systems and is easier to implement.

Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts

https://openreview.net/forum?id=auKAUJZMO6

Compressor summary: The study investigates how large language models respond to external information that conflicts with their internal knowledge, finding that they can be receptive to contradicting evidence but also show confirmation bias when presented with consistent information.

Adaptive Regularization of Representation Rank as an Implicit Constraint of Bellman Equation

https://openreview.net/forum?id=apXtolxDaJ

Compressor summary: The paper introduces BEER, a regularizer that adaptively controls the representation rank of value networks in Deep Reinforcement Learning, improving agent performance and Q-value approximation.

A Topological Perspective on Demystifying GNN-Based Link Prediction Performance

https://openreview.net/forum?id=apA6SSXx2e

Compressor summary: The paper explores how Topological Concentration (TC), a new metric based on local subgraphs, can improve link prediction performance in graph neural networks (GNNs) by identifying and addressing topological distribution shifts.

Unveiling and Manipulating Prompt Influence in Large Language Models

https://openreview.net/forum?id=ap1ByuwQrX

Compressor summary: Token Distribution Dynamics (TDD) is a method to analyze and manipulate prompts for Large Language Models, using the LM head's interpreting capabilities to estimate input saliency based on distribution dynamics over the vocabulary.

Making Pre-trained Language Models Great on Tabular Prediction

https://openreview.net/forum?id=anzIzGZuLi

Compressor summary: TP-BERTa is a pre-trained language model that can handle numerical features in tables and achieve high performance in tabular data prediction tasks.

Self-Supervised Speech Quality Estimation and Enhancement Using Only Clean Speech

https://openreview.net/forum?id=ale56Ya59q

Compressor summary: The paper proposes VQScore, a self-supervised metric for evaluating speech quality using vector-quantized-variational autoencoder (VQ-VAE) and domain knowledge, which can also be used for self-supervised speech enhancement with improved robustness.

Predictive auxiliary objectives in deep RL mimic learning in the brain

https://openreview.net/forum?id=agPpmEgf8C

Compressor summary: The study shows how predictive objectives in deep reinforcement learning improve and stabilize representation learning, mimic brain activity changes, and suggest a new role for the hippocampus as an auxiliary learning system.

Circumventing Concept Erasure Methods For Text-To-Image Generative Models

https://openreview.net/forum?id=ag3o2T51Ht

Compressor summary: The authors examine seven methods to remove sensitive content from text-to-image models and show that none of them fully succeed, as they can still retrieve the erased concepts using special learned word embeddings.

Entropy Coding of Unordered Data Structures

https://openreview.net/forum?id=afQuNt3Ruh

Compressor summary: Shuffle coding is a new method for compressing unordered sequences of objects using bits-back coding, which works well for various data structures like graphs and molecular data.

Decision ConvFormer: Local Filtering in MetaFormer is Sufficient for Decision Making

https://openreview.net/forum?id=af2c8EaKl8

Compressor summary: The authors propose Decision ConvFormer, a new action sequence predictor for offline reinforcement learning, which improves on Decision Transformer by using local convolution filtering to capture local dependencies in trajectories.

On the Posterior Distribution in Denoising: Application to Uncertainty Quantification

https://openreview.net/forum?id=adSGeugiuj

Compressor summary: The paper derives a relation between higher-order moments of the posterior distribution and derivatives of the posterior mean, enabling efficient uncertainty quantification for pre-trained denoisers without training or fine tuning.

SCHEMA: State CHangEs MAtter for Procedure Planning in Instructional Videos

https://openreview.net/forum?id=abL5LJNZ49

Compressor summary: The paper proposes a method to represent and track state changes in instructional videos for better procedure planning using large language models and cross-modal contrastive learning.

Soft Mixture Denoising: Beyond the Expressive Bottleneck of Diffusion Models

https://openreview.net/forum?id=aaBnFAyW9O

Compressor summary: The paper analyzes the limitations of current diffusion models in denoising tasks, introduces a new model called soft mixture denoising (SMD) that improves their performance, and provides experimental results supporting its effectiveness.

Multi-Task Reinforcement Learning with Mixture of Orthogonal Experts

https://openreview.net/forum?id=aZH1dM3GOX

Compressor summary: MOORE is a novel method for representation learning in multi-task reinforcement learning that uses orthogonal subspaces to promote diversity and generate task-specific representations.

Horizon-free Reinforcement Learning in Adversarial Linear Mixture MDPs

https://openreview.net/forum?id=aPNwsJgnZJ

Compressor summary: The paper proposes a new algorithm for adversarial RL that achieves sub-optimal regret using variance-aware weighted least squares and occupancy measure-based policy search.

Symbol as Points: Panoptic Symbol Spotting via Point-based Representation

https://openreview.net/forum?id=aOnUe8ah7j

Compressor summary: SymPoint uses point cloud segmentation methods to spot and parse objects and stuff from CAD drawings, outperforming recent state-of-the-art method by a significant margin.

The mechanistic basis of data dependence and abrupt learning in an in-context classification task

https://openreview.net/forum?id=aN4Jf6Cx69

Compressor summary: In-context learning in transformer models depends on distributional properties and a specific sequence of multi-layer operations that enable an induction head with three nested logits.

Approximately Piecewise E(3) Equivariant Point Networks

https://openreview.net/forum?id=aKJEHWmBEf

Compressor summary: APEN is a framework that improves point cloud neural networks by approximating their equivariance to local Euclidean transformations, using a compositional design for partition prediction with guaranteed bounds on approximation errors.

LLMCarbon: Modeling the End-to-End Carbon Footprint of Large Language Models

https://openreview.net/forum?id=aIok3ZD9to

Compressor summary: LLMCarbon is a tool that accurately predicts the carbon emissions of different large language models before training, overcoming limitations of previous methods like mlco2.

InsertNeRF: Instilling Generalizability into NeRF with HyperNet Modules

https://openreview.net/forum?id=aHmNpLlUlb

Compressor summary: InsertNeRF is a method that improves NeRF's ability to generalize to new scenes by using HyperNet modules to adapt the model's weights and features.

Multi-modal Gaussian Process Variational Autoencoders for Neural and Behavioral Data

https://openreview.net/forum?id=aGH43rjoe4

Compressor summary: The paper proposes an unsupervised model that extracts shared and independent latent structures from different experimental modalities using Gaussian Process Factor Analysis and Gaussian Process Variational Autoencoders.

Interpretable Sparse System Identification: Beyond Recent Deep Learning Techniques on Time-Series Prediction

https://openreview.net/forum?id=aFWUY3E7ws

Compressor summary: The paper introduces a novel interpretable method for time series prediction that uses sparse optimization, Fourier basis, and trend-fluctuation analysis, and beats deep learning without needing training or computational expenses.

Object-Centric Learning with Slot Mixture Module

https://openreview.net/forum?id=aBUidW4Nkd

Compressor summary: The paper proposes a learnable clustering method for object-centric architectures that uses Gaussian Mixture Models to represent slots and improve performance on set property prediction tasks.

Gaining Wisdom from Setbacks: Aligning Large Language Models via Mistake Analysis

https://openreview.net/forum?id=aA33A70IO6

Compressor summary: The study suggests a new way to train language models to avoid generating harmful and toxic responses by analyzing their mistakes and using them as feedback for improvement.

Unveiling Options with Neural Network Decomposition

https://openreview.net/forum?id=a8VETFwcVR

Compressor summary: The paper presents an algorithm to create reusable sub-policies from neural networks for Markov Decision Processes, which are used as options to speed up learning on related tasks.

Understanding prompt engineering may not require rethinking generalization

https://openreview.net/forum?id=a745RnSFLT

Compressor summary: The paper explains how classical PAC-Bayes bounds can justify the success of zero-shot learning with prompts, despite their potential for overfitting, by showing that prompts' discrete nature and language model prior lead to tight generalization bounds.

StructComp: Substituting propagation with Structural Compression in Training Graph Contrastive Learning

https://openreview.net/forum?id=a4DBEeGfQq

Compressor summary: Structural Compression is a framework that trains graph contrastive learning models faster and more efficiently by compressing nodes, reducing computation and memory requirements, and improving performance.

iGraphMix: Input Graph Mixup Method for Node Classification

https://openreview.net/forum?id=a2ljjXeDcE

Compressor summary: iGraphMix is a novel node classification method that interpolates input features and labels to generate virtual nodes and graphs, improving generalization performance in graph neural networks.

Kalman Filter Online Learning from non-Stationary Data

https://openreview.net/forum?id=ZzmKEpze8e

Compressor summary: The paper proposes a probabilistic Bayesian online learning model that adapts to non-stationary data and quantifies predictive uncertainty using a neural representation and a state space model over linear weights.

Clifford Group Equivariant Simplicial Message Passing Networks

https://openreview.net/forum?id=Zz594UBNOH

Compressor summary: Clifford Group Equivariant Simplicial Message Passing Networks combine Clifford group-equivariant layers with shared simplicial message passing for steerable, topologically intricate geometric analysis.

Learning dynamic representations of the functional connectome in neurobiological networks

https://openreview.net/forum?id=ZwhHSOHMTM

Compressor summary: The text describes a method to learn dynamic neural networks in live animals by analyzing calcium activity and non-negative tensor factorization.

Diffusion Model for Dense Matching

https://openreview.net/forum?id=Zsfiqpft6K

Compressor summary: The paper proposes DiffMatch, a conditional diffusion-based framework for dense matching of paired images, addressing ambiguities by explicitly modeling data and prior terms with a cascaded pipeline.

Retrieval-based Disentangled Representation Learning with Natural Language Supervision

https://openreview.net/forum?id=ZlQRiFmq7Y

Compressor summary: VDR uses natural language to help learn disentangled representations from data by leveraging linguistic equivalents as proxies for the underlying variation.

INSIDE: LLMs' Internal States Retain the Power of Hallucination Detection

https://openreview.net/forum?id=Zj12nzlQbz

Compressor summary: The paper proposes a new method to detect when large language models make wrong answers by analyzing their internal states and using a simple metric called EigenScore.

From Zero to Turbulence: Generative Modeling for 3D Flow Simulation

https://openreview.net/forum?id=ZhlwoC1XaN

Compressor summary: The paper presents a novel generative model for 3D turbulent flows that learns the distribution of all possible flow states and can produce realistic samples without relying on an initial condition.

Towards the Fundamental Limits of Knowledge Transfer over Finite Domains

https://openreview.net/forum?id=Zh2iqiOtMt

Compressor summary: The paper studies how different types of information from a teacher can improve the efficiency of transferring knowledge to a probabilistic student classifier.

Towards Foundational Models for Molecular Learning on Large-Scale Multi-Task Datasets

https://openreview.net/forum?id=Zc2aIcucwc

Compressor summary: The authors introduce seven new molecular machine learning datasets with diverse labels, a graph library for easy modeling, and baseline results that show potential in multi-task and multi-level training for foundation models.

BENO: Boundary-embedded Neural Operators for Elliptic PDEs

https://openreview.net/forum?id=ZZTkLDRmkg

Compressor summary: Boundary-Embedded Neural Operators (BENO) is a new technique that solves elliptic PDEs with complex geometries and boundary conditions more efficiently using neural networks and Transformer encoder.

LipVoicer: Generating Speech from Silent Videos Guided by Lip Reading

https://openreview.net/forum?id=ZZCPSC5OgD

Compressor summary: LipVoicer is a novel lip-to-speech method that generates high-quality, intelligible speech from silent videos by incorporating the text modality and outperforming multiple baselines on in-the-wild datasets.

Bayesian Bi-clustering of Neural Spiking Activity with Latent Structures

https://openreview.net/forum?id=ZYm1Ql6udy

Compressor summary: The paper presents a bi-clustering method to analyze neural spiking activity in multiple brain regions, considering both latent trajectories and spatial-temporal clustering structures, using a non-parametric model and an MCMC algorithm.

Effective Data Augmentation With Diffusion Models

https://openreview.net/forum?id=ZWzUA9zeAg

Compressor summary: Our method uses text-to-image diffusion models to semantically edit images for data augmentation, improving performance on few-shot classification and weed recognition.

Functional Bayesian Tucker Decomposition for Continuous-indexed Tensor Data

https://openreview.net/forum?id=ZWyZeqE928

Compressor summary: FunBaT is a novel tensor model that extends Tucker decomposition to handle continuous-indexed data by using functional priors and efficient inference algorithms.

DMBP: Diffusion model based predictor for robust offline reinforcement learning against state observation perturbations

https://openreview.net/forum?id=ZULjcYLWKe

Compressor summary: The paper proposes DMBP, a framework that uses conditional diffusion models to recover actual states in offline RL and improve robustness against state observation perturbations.

How to Fine-Tune Vision Models with SGD

https://openreview.net/forum?id=ZTssMmhC2X

Compressor summary: SGD outperforms AdamW for fine-tuning large neural networks in computer vision when the embedding layer is frozen and memory usage is reduced.

Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps

https://openreview.net/forum?id=ZSD3MloKe6

Compressor summary: The paper proposes Time-Shift Sampler, a novel method to reduce exposure bias in Diffusion Probabilistic Models without retraining, by selecting better time steps for inference.

Making Retrieval-Augmented Language Models Robust to Irrelevant Context

https://openreview.net/forum?id=ZS4m74kZpH

Compressor summary: The authors analyze cases when retrieval-augmented language models reduce accuracy and propose two methods to improve their performance by filtering out irrelevant passages and fine-tuning the model with a mix of relevant and irrelevant contexts.

Mean Field Theory in Deep Metric Learning

https://openreview.net/forum?id=ZPdZLlNXSm

Compressor summary: The paper introduces mean field theory from statistical physics to create new loss functions for deep metric learning, reducing training complexity and improving image retrieval performance.

Efficient Distributed Training with Full Communication-Computation Overlap

https://openreview.net/forum?id=ZO5cn4IfaN

Compressor summary: CO2 is a new approach for large language model training that uses local updating and asynchronous communication to achieve 100% scalability on clusters with limited bandwidth, and it works well on various tasks in computer vision and natural language processing.

PanoDiffusion: 360-degree Panorama Outpainting via Diffusion

https://openreview.net/forum?id=ZNzDXDFZ0B

Compressor summary: PanoDiffusion is a new model that uses latent diffusion and progressive camera rotations to generate high-quality 360-degree RGB and depth panoramas from narrow field of view images.

Learning semilinear neural operators: A unified recursive framework for prediction and data assimilation.

https://openreview.net/forum?id=ZMv6zKYYUs

Compressor summary: The paper proposes a learning-based method for fast and accurate predictions and data assimilation of spatio-temporal PDEs using Neural Operators and nonlinear observers in function spaces.

A Stochastic Centering Framework for Improving Calibration in Graph Neural Networks

https://openreview.net/forum?id=ZL6yd6N1S2

Compressor summary: The paper proposes G-$\Delta$UQ, a framework to improve intrinsic uncertainty estimates of graph neural networks (GNNs) using graph anchoring strategies under distribution shifts.

A Lightweight Method for Tackling Unknown Participation Statistics in Federated Averaging

https://openreview.net/forum?id=ZKEuFKfCKA

Compressor summary: The paper proposes FedAU, a lightweight method for federated learning that adapts aggregation weights based on client participation history to handle diverse participation statistics and improve performance.

Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning

https://openreview.net/forum?id=ZGNWW7xZ6Q

Compressor summary: The paper proposes a method called reasoning on graphs (RoG) that combines large language models and knowledge graphs for faithful and interpretable reasoning, using relation paths as plans to retrieve valid reasoning paths from the graphs.

Connecting Large Language Models with Evolutionary Algorithms Yields Powerful Prompt Optimizers

https://openreview.net/forum?id=ZG3RaNIsO8

Compressor summary: EvoPrompt is a novel framework for discrete prompt optimization using evolutionary algorithms and large language models to improve task performance on various natural language processing datasets.

Skill or Luck? Return Decomposition via Advantage Functions

https://openreview.net/forum?id=ZFMiHfZwIf

Compressor summary: The paper proposes Off-policy DAE, a method that learns from off-policy data by decomposing return into skill and luck components, without using importance sampling or truncation.

Det-CGD: Compressed Gradient Descent with Matrix Stepsizes for Non-Convex Optimization

https://openreview.net/forum?id=ZEZ0CPmoSI

Compressor summary: The paper presents new algorithms that use a matrix stepsize to optimize non-convex problems and achieve faster convergence than scalar methods, along with theoretical and empirical results.

Improving Language Models with Advantage-based Offline Policy Gradients

https://openreview.net/forum?id=ZDGKPbF0VQ

Compressor summary: A-LoL is an offline RL method for finetuning LMs that uses sequence-level rewards and value estimates to filter low-quality data and improve diversity, safety, and helpfulness of generated texts.

Neuro-Inspired Information-Theoretic Hierarchical Perception for Multimodal Learning

https://openreview.net/forum?id=Z9AZsU1Tju

Compressor summary: The ITHP model uses information bottleneck to create a compact information flow in multimodal learning, outperforming state-of-the-art benchmarks and achieving human-level performance on sentiment classification.

Addressing Signal Delay in Deep Reinforcement Learning

https://openreview.net/forum?id=Z8UfDs4J46

Compressor summary: The paper explores the impact of signal delay in deep reinforcement learning, formalizes delayed-observation Markov decision processes, and proposes effective strategies to overcome related challenges for continuous robotic control tasks.

Trajeglish: Learning the Language of Driving Scenarios

https://openreview.net/forum?id=Z59Rb5bPPP

Compressor summary: The authors present a language modeling approach to generate realistic driving scenarios with partial autonomy and intra-timestep interaction, and evaluate it on two datasets.

Posterior Sampling Based on Gradient Flows of the MMD with Negative Distance Kernel

https://openreview.net/forum?id=YrXHEb2qMb

Compressor summary: The text proposes using MMD with negative distance kernel for posterior sampling and conditional generative modelling, showing its advantages in computation and error bound, and demonstrating its effectiveness in various numerical examples.

An Agnostic View on the Cost of Overfitting in (Kernel) Ridge Regression

https://openreview.net/forum?id=YrTI2Zu0dd

Compressor summary: The paper analyzes how different sample sizes affect the cost of overfitting in noisy kernel ridge regression and classifies overfitting scenarios as benign, tempered, or catastrophic.

Emerging Pixel-level Semantic Knowledge in Diffusion Models

https://openreview.net/forum?id=YqyTXmF8Y2

Compressor summary: The text describes a method to build an image segmentor using semantic knowledge from Stable Diffusion without additional supervision.

AUC-CL: A Batchsize-Robust Framework for Self-Supervised Contrastive Representation Learning

https://openreview.net/forum?id=YgMdDQB09U

Compressor summary: The paper proposes a new contrastive learning approach that performs well with smaller batch sizes and is robust to stochastic optimization, using AUC-maximization for embedding space arrangements.

Is ImageNet worth 1 video? Learning strong image encoders from 1 long unlabelled video

https://openreview.net/forum?id=Yen1lGns2o

Compressor summary: The authors present a self-supervised video pretraining method that leverages tracking and attention to learn from unannotated first-person videos, such as the new "Walking Tours" dataset.

Adversarial Causal Bayesian Optimization

https://openreview.net/forum?id=YcW8i9VCf5

Compressor summary: The paper introduces ACBO, a generalization of CBO that considers multiple agents or external events intervening on the system, and proposes CBO-MW, an algorithm with bounded regret that combines online learning and causal modeling.

BayesDiff: Estimating Pixel-wise Uncertainty in Diffusion via Bayesian Inference

https://openreview.net/forum?id=YcM6ofShwY

Compressor summary: BayesDiff is a method that uses Bayesian inference to estimate pixel-wise uncertainty in diffusion models, which can help filter low-quality images and improve image generation.

Improving Intrinsic Exploration by Creating Stationary Objectives

https://openreview.net/forum?id=YbZxT0SON4

Compressor summary: The paper proposes a framework to transform non-stationary intrinsic rewards in reinforcement learning into stationary ones using state augmentations, improving agent performance in various exploration problems.

Composed Image Retrieval with Text Feedback via Multi-grained Uncertainty Regularization

https://openreview.net/forum?id=Yb5KvPkKQg

Compressor summary: The text introduces a unified learning approach for composed image retrieval with text feedback that considers coarse- and fine-grained retrieval by modeling multi-grained uncertainty, improving the recall rate on three datasets.

Sample Efficient Myopic Exploration Through Multitask Reinforcement Learning with Diverse Tasks

https://openreview.net/forum?id=YZrg56G0JV

Compressor summary: This paper shows how multitask reinforcement learning (MTRL) with myopic exploration can be more efficient when tasks are diverse and related, and validates this theory with experiments on synthetic robotics.

Tuning LayerNorm in Attention: Towards Efficient Multi-Modal LLM Finetuning

https://openreview.net/forum?id=YR3ETaElNK

Compressor summary: The paper presents an efficient method to make large language models handle multiple modalities by tuning LayerNorm layers, which leads to significant performance improvements and reduced resource usage.

In-Context Learning Learns Label Relationships but Is Not Conventional Learning

https://openreview.net/forum?id=YPIA7bgd5y

Compressor summary: This paper investigates how Large Language Models use label information when learning new tasks in context, revealing their capabilities, limitations, and biases.

Conditional Variational Diffusion Models

https://openreview.net/forum?id=YOKnEkIuoi

Compressor summary: The authors propose a new method for learning the variance schedule of diffusion models during training, which improves the performance and adaptability of inverse problems and avoids the need for fine-tuning by experimentation.

Towards Poisoning Fair Representations

https://openreview.net/forum?id=YLJs4mKJCF

Compressor summary: The text introduces a new data poisoning attack on fair representation learning, which can produce unfair representations that contain demographic information by injecting crafted samples into training data.

Symmetric Mean-field Langevin Dynamics for Distributional Minimax Problems

https://openreview.net/forum?id=YItWKZci78

Compressor summary: The paper presents two algorithms for minimax optimization over probability distributions, MFL-AG and MFL-ABR, which use mean-field Langevin dynamics and have symmetric and convergent updates.

Parsing neural dynamics with infinite recurrent switching linear dynamical systems

https://openreview.net/forum?id=YIls9HEa52

Compressor summary: The irSLDS model extends the rSLDS model to analyze trial-varying neural activity data by incorporating latent geometry, flexible state cardinality, and dynamical sufficient statistics.

Active Test-Time Adaptation: Theoretical Analyses and An Algorithm

https://openreview.net/forum?id=YHUGlwTzFB

Compressor summary: ATTA is a novel test-time adaptation method that uses active learning to improve performance under distribution shifts, with theoretical guarantees and efficient sample selection techniques.

TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting

https://openreview.net/forum?id=YH5w12OUuU

Compressor summary: TEMPO uses GPT-like architecture to learn time series representations, improving accuracy and adapting to non-stationary data.

Differentially Private Synthetic Data via Foundation Model APIs 1: Images

https://openreview.net/forum?id=YEhQs8POIo

Compressor summary: The paper proposes Private Evolution (PE), an API-based framework that generates differentially private synthetic data without training models, and shows its effectiveness on synthetic images and large foundation models.

Implicit Gaussian process representation of vector fields over arbitrary latent manifolds

https://openreview.net/forum?id=YEPlTU5mZC

Compressor summary: RVGP is a novel method for learning vector signals over latent Riemannian manifolds, which can super-resolve and inpaint vector fields while preserving singularities and reconstruct high-density neural dynamics from low-density EEG recordings.

Training Diffusion Models with Reinforcement Learning

https://openreview.net/forum?id=YCWjhGrJFD

Compressor summary: The paper proposes reinforcement learning methods to optimize diffusion models for downstream objectives like image quality and drug effectiveness, outperforming reward-weighted likelihood approaches.

Leveraging augmented-Lagrangian techniques for differentiating over infeasible quadratic programs in machine learning

https://openreview.net/forum?id=YCPDFfmkFr

Compressor summary: The paper proposes a unified approach to differentiate both feasible and infeasible convex quadratic programming layers in neural networks, enabling faster computations and improved performance in some tasks.

Sparse Weight Averaging with Multiple Particles for Iterative Magnitude Pruning

https://openreview.net/forum?id=Y9t7MqZtCR

Compressor summary: SWAMP is a simple modification of IMP that achieves better performance by training multiple sparse models and weight averaging them, resulting in faster inference speeds and less memory demand for modern neural networks.

GoLLIE: Annotation Guidelines improve Zero-Shot Information-Extraction

https://openreview.net/forum?id=Y3wpuxd7u9

Compressor summary: GoLLIE is a model that follows annotation guidelines to improve its performance on unseen information extraction tasks.

DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization

https://openreview.net/forum?id=Y3BbxvAQS9

Compressor summary: DecompOpt is a new structure-based method that combines optimization and conditional diffusion models to generate novel ligands with desired properties, such as high binding affinity and easy synthesis, by decomposing them into substructures for fine-grained control.

Balancing Act: Sparse Models with Constrained Disparate Impact

https://openreview.net/forum?id=Xz13DtbOVW

Compressor summary: Our method constrains the accuracy change for each sub-group when pruning models, ensuring fairness and interpretability while scaling well.

Quantifying and Enhancing Multi-modal Robustness with Modality Preference

https://openreview.net/forum?id=XyrB1Ay44j

Compressor summary: The paper explores the limitations of multi-modal models' robustness, identifies essential components for achieving higher robustness, and proposes a new training procedure called CRMT that significantly improves performance and robustness.

Smooth ECE: Principled Reliability Diagrams via Kernel Smoothing

https://openreview.net/forum?id=XwiA1nDahv

Compressor summary: SmoothECE is a well-behaved calibration measure that fixes common flaws in binning and ECE by smoothing observations using an RBF kernel before computing ECE, and provides a Python package for measuring and plotting calibration.

Continual Learning on a Diet: Learning from Sparsely Labeled Streams Under Constrained Computation

https://openreview.net/forum?id=Xvfz8NHmCj

Compressor summary: The paper introduces a realistic Continual Learning setting with limited resources and proposes a simple but effective method called DietCL that uses both labeled and unlabeled data efficiently to outperform existing methods.

Dichotomy of Early and Late Phase Implicit Biases Can Provably Induce Grokking

https://openreview.net/forum?id=XsHqr9dEGH

Compressor summary: The paper investigates the grokking phenomenon in neural networks and shows that it is caused by a shift from early to late phase implicit biases during training.

Spatio-Temporal Approximation: A Training-Free SNN Conversion for Transformers

https://openreview.net/forum?id=XrunSYwoLr

Compressor summary: The paper proposes a method to convert pretrained Transformers into spiking neural networks by approximating their non-linear operations in both time and space dimensions.

Adding 3D Geometry Control to Diffusion Models

https://openreview.net/forum?id=XlkN11Xj6J

Compressor summary: The paper proposes a method to incorporate 3D geometry control into diffusion models by using visual prompts from rendered images, enabling better performance in various vision tasks.

DNA-GPT: Divergent N-Gram Analysis for Training-Free Detection of GPT-Generated Text

https://openreview.net/forum?id=Xlayxj2fWp

Compressor summary: The authors propose DNA-GPT, a training-free detection strategy that analyzes the differences between original and regenerated text by LLMs to distinguish machine-generated from human-written text with high performance and explainability.

The importance of feature preprocessing for differentially private linear optimization

https://openreview.net/forum?id=XlTDBZFXWp

Compressor summary: DPSGD is not sufficient for differentially private optimization without feature preprocessing, and DPSGD-F combines it with feature preprocessing to reduce privacy errors for classification tasks.

Stabilizing Contrastive RL: Techniques for Robotic Goal Reaching from Offline Data

https://openreview.net/forum?id=Xkf2EBj4w3

Compressor summary: The paper proposes a self-supervised RL method for robotic systems that uses contrastive learning and shows its effectiveness on simulated and real-world image-based manipulation tasks.

MogaNet: Multi-order Gated Aggregation Network

https://openreview.net/forum?id=XhYWgjqCrV

Compressor summary: MogaNet is a new family of modern ConvNets that efficiently gathers and contextualizes discriminative features for improved computer vision tasks with better complexity-performance trade-offs.

Stack Attention: Improving the Ability of Transformers to Model Hierarchical Patterns

https://openreview.net/forum?id=XVhm3X8Fum

Compressor summary: Stack attention, an attention operator incorporating stacks, enhances transformers' ability to recognize hierarchical patterns of arbitrary nesting depth in both natural language and context-free languages without syntactic supervision.

Quantifying the Plausibility of Context Reliance in Neural Machine Translation

https://openreview.net/forum?id=XTHfNGI3zT

Compressor summary: PECoRe is an interpretability framework that quantifies how language models use context for translation, comparing their rationales with human annotations and identifying context-mediated predictions.

Consistent algorithms for multi-label classification with macro-at-$k$ metrics

https://openreview.net/forum?id=XOnya9gSdF

Compressor summary: The paper studies how to optimize complex metrics in multi-label classification under population utility framework and proposes a practical learning algorithm based on Frank-Wolfe method.

Abstractors and relational cross-attention: An inductive bias for explicit relational reasoning in Transformers

https://openreview.net/forum?id=XNa6r6ZjoB

Compressor summary: The text proposes an extension of Transformers called the Abstractor that enables relational reasoning through a novel module with relational cross-attention, improving abstraction, generalization, and sample efficiency on various tasks.

COLEP: Certifiably Robust Learning-Reasoning Conformal Prediction via Probabilistic Circuits

https://openreview.net/forum?id=XN6ZPINdSg

Compressor summary: COLEP is a robust conformal prediction framework using probabilistic circuits that provides certified coverage guarantees against adversarial perturbations and outperforms single models in various datasets.

Adversarial Training Should Be Cast as a Non-Zero-Sum Game

https://openreview.net/forum?id=XJ9vjEAqbx

Compressor summary: The paper proposes a new approach to train neural networks that are more robust against adversarial attacks by using a non-zero-sum bilevel formulation of adversarial training.

InstructPix2NeRF: Instructed 3D Portrait Editing from a Single Image

https://openreview.net/forum?id=XIxhINXtQk

Compressor summary: The paper proposes a novel end-to-end framework for instructed 3D-aware portrait editing from a single open-world image using conditional latent 3D diffusion and identity consistency module.

Idempotent Generative Network

https://openreview.net/forum?id=XIaS66XkNA

Compressor summary: The new approach trains a neural network to be idempotent, mapping source noise to target images, and can generate output in one step, refine with sequential applications, and handle corrupted inputs.

Debiasing Algorithm through Model Adaptation

https://openreview.net/forum?id=XIZEFyVGC9

Compressor summary: The paper proposes a new method (DAMA) to detect and reduce gender bias in large language models by modifying certain layers based on causal analysis.

GRANDE: Gradient-Based Decision Tree Ensembles

https://openreview.net/forum?id=XEFWBxi075

Compressor summary: The paper introduces GRANDE, a novel gradient-based method for learning hard decision tree ensembles on tabular data, which achieves better performance than existing methods.

Interpretable Diffusion via Information Decomposition

https://openreview.net/forum?id=X6tNkN6ate

Compressor summary: Denoising diffusion models can learn complex relationships between words and pixels in an image, and these relationships can be analyzed using information decomposition methods.

Training-free Multi-objective Diffusion Model for 3D Molecule Generation

https://openreview.net/forum?id=X41c4uB4k0

Compressor summary: The text presents a method to generate 3D molecules with multiple desired properties using off-the-shelf models and probabilistic graphs, outperforming trained diffusion models.

Visual Data-Type Understanding does not emerge from scaling Vision-Language Models

https://openreview.net/forum?id=WyEdX2R4er

Compressor summary: This paper introduces a new vision-language model task called Visual Data-Type Identification and evaluates 39 models on it, revealing their weaknesses in recognizing simple image manipulations.

Theoretical Understanding of Learning from Adversarial Perturbations

https://openreview.net/forum?id=Ww9rWUAcdo

Compressor summary: The study shows how adversarial examples contain hidden data features that help neural networks learn and generalize, even when trained on mislabeled samples.

QA-LoRA: Quantization-Aware Low-Rank Adaptation of Large Language Models

https://openreview.net/forum?id=WvFoJccpo8

Compressor summary: The paper introduces QA-LoRA, an algorithm that makes large language models more efficient, accurate, and scalable by balancing quantization and adaptation.

Harnessing Overlap in Blockwise Transformers for Near-Infinite Context

https://openreview.net/forum?id=WsRHpHH4s0

Compressor summary: Ring Attention is a novel method that allows Transformers to handle long sequences by distributing them across multiple devices and overlapping communication through blockwise attention computation, enabling efficient training and inference on sequences longer than 100 million tokens.

A Statistical Analysis of Wasserstein Autoencoders for Intrinsically Low-dimensional Data

https://openreview.net/forum?id=WjRPZsfeBO

Compressor summary: The paper analyzes Wasserstein Autoencoders' statistical guarantees and shows they can learn data distributions when network architectures are well designed, focusing on convergence rates that depend only on the intrinsic dimension of the data.

Adaptive Self-training Framework for Fine-grained Scene Graph Generation

https://openreview.net/forum?id=WipsLtH77t

Compressor summary: The paper proposes a self-training framework for scene graph generation that uses class-specific adaptive thresholding with momentum to assign pseudo-labels to unannotated triplets and improves performance on fine-grained predicate classes.

Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting

https://openreview.net/forum?id=WhgB5sispV

Compressor summary: The paper presents a novel 4D primitive-based approach to synthesize diverse views of dynamic scenes from 2D images, with improved simplicity, flexibility, efficiency, and quality.

Light Schrödinger Bridge

https://openreview.net/forum?id=WhZoCLRWYJ

Compressor summary: The paper proposes a fast and simple Schrodinger Bridges solver using sum-exp quadratic functions and log-Schrodinger potentials as energy functions, making it suitable for solving moderate-dimensional problems on CPU without complex optimization or hyperparameter selection.

Uni-RLHF: Universal Platform and Benchmark Suite for Reinforcement Learning with Diverse Human Feedback

https://openreview.net/forum?id=WesY0H9ghM

Compressor summary: Uni-RLHF is a system for reinforcement learning with human feedback that provides a complete workflow, large-scale datasets, and modular baselines to facilitate progress in practical problems.

Quasi-Monte Carlo for 3D Sliced Wasserstein

https://openreview.net/forum?id=Wd47f7HEXg

Compressor summary: Quasi-Monte Carlo methods improve the Sliced Wasserstein distance approximation in 3D tasks by providing a better class of empirical SW and unbiased estimators.

Guiding Masked Representation Learning to Capture Spatio-Temporal Relationship of Electrocardiogram

https://openreview.net/forum?id=WcOohbsF4H

Compressor summary: The paper introduces ST-MEM, a self-supervised learning method for ECG data that reconstructs masked 12-lead ECG data and learns spatio-temporal features, improving arrhythmia classification performance.

Knowledge Card: Filling LLMs' Knowledge Gaps with Plug-in Specialized Language Models

https://openreview.net/forum?id=WbWtOYIzIK

Compressor summary: Knowledge Card is a framework to update large language models with domain-specific facts, improving their performance on various tasks.

Latent Intuitive Physics: Learning to Transfer Hidden Physics from a 3D Video

https://openreview.net/forum?id=WZu4gUGN13

Compressor summary: The text introduces latent intuitive physics, a method to simulate fluids from 3D videos without knowing their physical properties, using a learned neural renderer and a parametrized prior learner.

An operator preconditioning perspective on training in physics-informed machine learning

https://openreview.net/forum?id=WWlxFtR5sV

Compressor summary: The paper explores how the difficulty of training physics-informed machine learning models depends on the conditioning of a specific differential operator, and proposes preconditioning methods to improve training.

Learning to Embed Time Series Patches Independently

https://openreview.net/forum?id=WS7GuBDFa2

Compressor summary: The paper proposes a simpler and more efficient method for time series representation learning that uses patch reconstruction, independent embedding of patches, and contrastive learning, outperforming masked time series modeling with Transformers.

Modeling state-dependent communication between brain regions with switching nonlinear dynamical systems

https://openreview.net/forum?id=WQwV7Y8qwa

Compressor summary: MR-SDS is a nonlinear state space model that decomposes neural dynamics into local and cross-communication components in the latent space, accounting for sensory inputs, history effects, and heterogeneity, and can accurately recover latent trajectories, vector fields, and cross-region communication profiles in multi-region neural datasets.

Mind Your Augmentation: The Key to Decoupling Dense Self-Supervised Learning

https://openreview.net/forum?id=WQYHbr36Fo

Compressor summary: The paper proposes a novel augmentation method (RCC) and decoupling branch to address the coupling phenomenon in dense self-supervised learning, improving feature disentanglement among objects.

Fine-Tuning Language Models for Factuality

https://openreview.net/forum?id=WPZ2yPag4K

Compressor summary: The authors propose a method to improve the factual accuracy of language models by fine-tuning them using automated criteria or preference rankings, reducing errors in open-ended text generation settings.

TextField3D: Towards Enhancing Open-Vocabulary 3D Generation with Noisy Text Fields

https://openreview.net/forum?id=WOiOzHG2zD

Compressor summary: The TextField3D model introduces dynamic noise to latent space of text prompts to improve 3D representation learning and enable open-vocabulary generation with large vocabulary, text consistency, and low latency.

Improved Techniques for Training Consistency Models

https://openreview.net/forum?id=WNzy9bRDvG

Compressor summary: The paper proposes improved techniques for training consistency models that achieve higher quality samples without distillation or bias, leading to significant improvements over prior approaches.

Lipschitz Singularities in Diffusion Models

https://openreview.net/forum?id=WNkW0cOwiz

Compressor summary: The paper studies a problem in diffusion models where they have infinite Lipschitz near zero timesteps, which affects their stability and accuracy, and proposes a novel approach called E-TSDM to mitigate it.

AUGCAL: Improving Sim2Real Adaptation by Uncertainty Calibration on Augmented Synthetic Images

https://openreview.net/forum?id=WNQjN5HzXt

Compressor summary: AUGCAL is a method that improves the performance of models trained on synthetic data by reducing miscalibration and overconfidence in real-world applications.

On the Role of Discrete Tokenization in Visual Representation Learning

https://openreview.net/forum?id=WNLAkjUm19

Compressor summary: This study investigates how discrete visual tokens enhance masked image modeling and proposes a new metric and tokenizer design to improve generalization performance.

Learning Optimal Contracts: How to Exploit Small Action Spaces

https://openreview.net/forum?id=WKuimaBj4I

Compressor summary: The paper proposes an algorithm for designing optimal payment schemes in multi-round principal-agent problems with small agent action spaces, achieving better regret bounds than existing methods.

GNNBoundary: Towards Explaining Graph Neural Networks through the Lens of Decision Boundaries

https://openreview.net/forum?id=WIzzXCVYiH

Compressor summary: The paper proposes a method called GNNBoundary to understand the decision boundaries of graph neural networks (GNNs) by generating near-boundary graphs for adjacent class pairs and analyzing them.

Motion Guidance: Diffusion-Based Image Editing with Differentiable Motion Estimators

https://openreview.net/forum?id=WIAO4vbnNV

Compressor summary: Motion guidance is a technique that uses optical flow networks to steer diffusion models for precise editing of images.

Neuroformer: Multimodal and Multitask Generative Pretraining for Brain Data

https://openreview.net/forum?id=W8S8SxS9Ng

Compressor summary: Neuroformer is a pre-trained transformer model for analyzing large-scale neuronal spiking data in systems neuroscience, which can predict behavior and infer neural circuit connectivity from simulated datasets.

Modeling Boundedly Rational Agents with Latent Inference Budgets

https://openreview.net/forum?id=W3VsHuga3j

Compressor summary: The latent inference budget model (L-IBM) simulates computational constraints in agent populations by using a latent variable that controls the runtime of an iterative inference algorithm, allowing for better learning of suboptimal decision-making and inference of goals.

A Flexible Generative Model for Heterogeneous Tabular EHR with Missing Modality

https://openreview.net/forum?id=W2tCmRrj7H

Compressor summary: FLEXGEN-EHR is a diffusion model for synthetic EHR generation that handles missing modalities and preserves fidelity and utility better than existing methods.

A unified sampling framework for solver searching of Diffusion Probabilistic Models

https://openreview.net/forum?id=W2d3LZbhhI

Compressor summary: The paper proposes a unified sampling framework that improves sample quality by adjusting solver strategies at different timesteps, and introduces $S^3$, a method to automatically optimize the solver schedule for efficient sampling from diffusion probabilistic models.

Tree Cross Attention

https://openreview.net/forum?id=Vw24wtSddM

Compressor summary: Tree Cross Attention (TCA) is a method that retrieves information from a logarithmic number of tokens for inference in a tree structure, making it more token-efficient than Cross Attention and Perceiver IO.

MetaGPT: Meta Programming for Multi-Agent Collaborative Framework

https://openreview.net/forum?id=VtmBAGCN7o

Compressor summary: MetaGPT is a meta-programming framework that uses human workflows and Standardized Operating Procedures to improve the collaboration of language models in solving complex tasks.

Understanding Catastrophic Forgetting in Language Models via Implicit Inference

https://openreview.net/forum?id=VrHiF2hsrm

Compressor summary: The text discusses the drawbacks of fine-tuning language models and proposes Conjugate Prompting to improve their performance on various tasks.

Understanding Reconstruction Attacks with the Neural Tangent Kernel and Dataset Distillation

https://openreview.net/forum?id=VoLDkQ6yR3

Compressor summary: This work studies a data reconstruction attack on neural networks, showing how it can recover the entire training set in the infinite width regime and revealing its dependence on network properties and image characteristics.

Influencer Backdoor Attack on Semantic Segmentation

https://openreview.net/forum?id=VmGRoNDQgJ

Compressor summary: The text presents Influencer Backdoor Attack, a new way to manipulate semantic segmentation models by injecting triggers on non-victim pixels, and two trigger injection methods. It also demonstrates that current segmentation models are susceptible to IBA in real-world applications.

RECOMBINER: Robust and Enhanced Compression with Bayesian Implicit Neural Representations

https://openreview.net/forum?id=VkWbxFrCC8

Compressor summary: RECOMBINER is a improved data compression method that uses hierarchical priors, learnable positional encodings, and patch-based encoding to overcome the limitations of previous INR-based approaches.

Towards Green AI in Fine-tuning Large Language Models via Adaptive Backpropagation

https://openreview.net/forum?id=Vja3ecieXY

Compressor summary: GreenTrainer is a new technique that reduces the environmental impact of fine-tuning large language models by adaptively selecting important tensors for training, achieving up to 64% FLOPs reduction without sacrificing accuracy.

T-MARS: Improving Visual Representations by Circumventing Text Feature Learning

https://openreview.net/forum?id=ViPtjIVzUw

Compressor summary: The paper proposes T-MARS, a data filtering approach for large multimodal datasets that removes image-caption pairs with overlapping text dominating visual features, improving computer vision tasks.

Deep Temporal Graph Clustering

https://openreview.net/forum?id=ViNe1fjGME

Compressor summary: The paper introduces TGC, a general framework for deep Temporal Graph Clustering that preserves dynamic information and reduces computational consumption by adjusting deep clustering techniques to suit temporal graphs.

WildFusion: Learning 3D-Aware Latent Diffusion Models in View Space

https://openreview.net/forum?id=VdwVOREDZM

Compressor summary: WildFusion is a new method that uses latent diffusion models to generate high-quality, 3D-consistent images from in-the-wild datasets without relying on posed images or canonical camera coordinates.

Generative Sliced MMD Flows with Riesz Kernels

https://openreview.net/forum?id=VdkGRV1vcf

Compressor summary: The paper shows that Riesz kernels in maximum mean discrepancy flows have efficient computation properties, allowing their use in training generative models for images.

Provable Out-of-Distribution Generalization in Hypersphere

https://openreview.net/forum?id=VXak3CZZGC

Compressor summary: The paper presents a novel OOD generalization framework that learns invariant features in a hyperspherical space based on class alignment and separation, and shows its advantages theoretically and experimentally.

Order-Preserving GFlowNets

https://openreview.net/forum?id=VXDPXuq4oG

Compressor summary: Order-Preserving GFlowNets use a learned reward function to sample candidates with probabilities proportional to their order, enabling efficient exploration and exploitation in single-objective and multi-objective tasks.

An LLM can Fool Itself: A Prompt-Based Adversarial Attack

https://openreview.net/forum?id=VVgGbB9TNV

Compressor summary: PromptAttack is an efficient tool to audit large language models' adversarial robustness by converting attacks into a prompt that fools the model without changing its semantic meaning.

ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms

https://openreview.net/forum?id=VTYg5ykEGS

Compressor summary: The paper introduces ImageNet-OOD, a dataset that separates semantic and covariate shifts in out-of-distribution detection, and shows that existing detectors are more sensitive to covariate shift than semantic shift.

SWE-bench: Can Language Models Resolve Real-world Github Issues?

https://openreview.net/forum?id=VTF8yNQM66

Compressor summary: SWE-bench is a new benchmark that assesses language models' ability to fix real-world software bugs by editing code and passing tests.

Tangent Transformers for Composition,Privacy and Removal

https://openreview.net/forum?id=VLFhbOCz5D

Compressor summary: TAFT is a method for fine-tuning linearized transformers using first-order Taylor expansion, which has lower cost, comparable performance, and several advantages over non-linear fine-tuning.

GnnX-Bench: Unravelling the Utility of Perturbation-based GNN Explainers through In-depth Benchmarking

https://openreview.net/forum?id=VJvbOSXRUq

Compressor summary: The paper evaluates and compares various perturbation-based explainability methods for GNNs, identifying Pareto-optimal techniques and highlighting stability and feasibility issues.

Don't Trust: Verify -- Grounding LLM Quantitative Reasoning with Autoformalization

https://openreview.net/forum?id=V5tdi14ple

Compressor summary: The paper proposes using formal mathematics examples in training LLMs to improve their quantitative reasoning skills and prevent logical errors.

FedLoGe: Joint Local and Generic Federated Learning under Long-tailed Data

https://openreview.net/forum?id=V3j5d0GQgH

Compressor summary: FedLoGe improves local and global model performance in Federated Long-Tailed Learning by integrating representation learning and classifier alignment using a shared backbone and individualized classifiers.

Exploring the Promise and Limits of Real-Time Recurrent Learning

https://openreview.net/forum?id=V2cBKtdC3a

Compressor summary: RTRL is a sequence-processing RNN method that has conceptual advantages over BPTT but suffers from high time and space complexity; by combining it with policy gradients, researchers can achieve competitive results on challenging environments without approximation.

Neural structure learning with stochastic differential equations

https://openreview.net/forum?id=V1GM9xDvIY

Compressor summary: The paper introduces SCOTCH, a novel method that combines neural stochastic differential equations with variational inference to learn structures from continuous-time stochastic processes observed at any time points, improving upon existing approaches.

DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation

https://openreview.net/forum?id=UyNXMqnN3c

Compressor summary: DreamGaussian is a fast and efficient 3D content generation framework that uses Gaussians and UV space techniques to create high-quality textured meshes from single-view images.

Meaning Representations from Trajectories in Autoregressive Models

https://openreview.net/forum?id=UyGWafcopT

Compressor summary: The authors propose a novel way to extract meaning representations from autoregressive language models without prompts or fine-tuning, which can handle complex semantic tasks and multimodal data.

A representation-learning game for classes of prediction tasks

https://openreview.net/forum?id=Uw8xvFqVAE

Compressor summary: The paper proposes a game-based method for learning dimensionality-reducing representations using prior knowledge on prediction tasks, with theoretical and practical implications.

Fourier Transporter: Bi-Equivariant Robotic Manipulation in 3D

https://openreview.net/forum?id=UulwvAU1W0

Compressor summary: FourTran is an open-loop behavior cloning method that uses symmetry to improve sample efficiency and achieve state-of-the-art results in pick and place tasks.

Accelerating Data Generation for Neural Operators via Krylov Subspace Recycling

https://openreview.net/forum?id=UpgRVWexaD

Compressor summary: SKR is a novel method to efficiently generate labeled data for training neural operators by sorting and recycling similar PDE systems, achieving up to 13.9x speedup.

Query-Policy Misalignment in Preference-Based Reinforcement Learning

https://openreview.net/forum?id=UoBymIwPJR

Compressor summary: The paper proposes a policy-aligned query and hybrid experience replay to improve preference-based reinforcement learning by addressing the issue of query-policy misalignment.

Time-LLM: Time Series Forecasting by Reprogramming Large Language Models

https://openreview.net/forum?id=Unb5CVPtae

Compressor summary: Time-LLM is a framework to repurpose large language models for time series forecasting using text prototypes and Prompt-as-Prefix, achieving state-of-the-art results and few-shot/zero-shot learning capabilities.

WizardCoder: Empowering Code Large Language Models with Evol-Instruct

https://openreview.net/forum?id=UnUwSIgK5W

Compressor summary: Code Evol-Instruct is a novel method to enhance Code LLMs by adapting the Evol-Instruct technique, leading to superior code generation performance and surpassing both open-source and closed-source LLMs on various benchmarks.

VDT: General-purpose Video Diffusion Transformers via Mask Modeling

https://openreview.net/forum?id=Un0rgm9f04

Compressor summary: Video Diffusion Transformer (VDT) is a transformer-based video generation model that captures temporal dependencies, handles diverse conditioning information, and can perform various tasks such as unconditional generation, video prediction, interpolation, animation, and completion.

EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models

https://openreview.net/forum?id=UmMa3UNDAz

Compressor summary: EfficientDM is a data-free framework that combines quantization-aware fine-tuning and post-training quantization to achieve high performance in low-bit diffusion models while maintaining time and data efficiency.

The Discovery of Binding Modes Requires Rethinking Docking Generalization

https://openreview.net/forum?id=UfBIxpTK10

Compressor summary: The text discusses a new benchmark called DockGen for assessing the generalization abilities of machine learning-based docking methods in finding ligand-binding proteins, and proposes Confidence Bootstrapping, a novel training paradigm that leverages diffusion and confidence models to improve generalizability.

Let 2D Diffusion Model Know 3D-Consistency for Robust Text-to-3D Generation

https://openreview.net/forum?id=UbxWjq0UO2

Compressor summary: 3DFuse is a novel framework that improves text-to-3D generation by incorporating 3D awareness into pretrained 2D diffusion models using a consistency injection module and semantic coding techniques, achieving state-of-the-art performance in geometric robustness and fidelity.

Decoupling regularization from the action space

https://openreview.net/forum?id=UaMgmoKEBj

Compressor summary: This paper proposes methods to improve regularized reinforcement learning by maintaining consistent regularization across different action spaces and temperatures, leading to better performance on various tasks.

Separate and Diffuse: Using a Pretrained Diffusion Model for Better Source Separation

https://openreview.net/forum?id=UXALv0lJZS

Compressor summary: The text describes a new speech separation method that uses generative models and achieves state-of-the-art results, even surpassing the previous upper bound for two speakers.

Source-Free and Image-Only Unsupervised Domain Adaptation for Category Level Object Pose Estimation

https://openreview.net/forum?id=UPvufoBAIs

Compressor summary: The paper presents a novel unsupervised 3D pose estimation method that uses cuboid meshes and generative neural features to adapt to the target domain without 3D data or annotations, leveraging stable object subparts and simulating fine-tuning.

Faithful Explanations of Black-box NLP Models Using LLM-generated Counterfactuals

https://openreview.net/forum?id=UMfcdRIotC

Compressor summary: The paper proposes two model-agnostic methods for counterfactual approximation, using a large language model either at inference-time or training-time, to improve the explanations of NLP system predictions effectively and efficiently.

A Precise Characterization of SGD Stability Using Loss Surface Geometry

https://openreview.net/forum?id=UMOlFJzLfL

Compressor summary: The paper investigates the relationship between linear stability and sharpness of SGD in overparameterized neural networks and introduces a new coherence measure for the loss Hessian to identify linear instability.

To Grok or not to Grok: Disentangling Generalization and Memorization on Corrupted Algorithmic Datasets

https://openreview.net/forum?id=UHjE5v5MB7

Compressor summary: The authors study how two-layer neural networks generalize on modular arithmetic tasks with corrupted labels and show that regularization methods improve performance by forcing the network to learn interpretable representations.

CrossLoco: Human Motion Driven Control of Legged Robots via Guided Unsupervised Reinforcement Learning

https://openreview.net/forum?id=UCfz492fM8

Compressor summary: CrossLoco is a guided unsupervised reinforcement learning framework that learns robot skills and their correspondence to human motions, achieving better accuracy, diversity, and user preference.

CivRealm: A Learning and Reasoning Odyssey for Decision-Making Agents

https://openreview.net/forum?id=UBVNwD3hPN

Compressor summary: CivRealm is an environment that tests decision-making agents' abilities in learning from past experiences and reasoning in novel contexts inspired by the Civilization game.

CALICO: Self-Supervised Camera-LiDAR Contrastive Pre-training for BEV Perception

https://openreview.net/forum?id=U7iiF79kI3

Compressor summary: CALICO is a novel pretraining framework for multimodal bird's eye view perception in autonomous driving systems that uses contrastive learning to improve 3D object detection and map segmentation performance, as well as robustness against adversarial attacks and corruption.

Respect the model: Fine-grained and Robust Explanation with Sharing Ratio Decomposition

https://openreview.net/forum?id=U7VW3KBm34

Compressor summary: SRD is a novel XAI method that enhances robustness by reflecting the model's inference process with a vector perspective and considering both active and inactive neurons.

3D-Aware Hypothesis & Verification for Generalizable Relative Object Pose Estimation

https://openreview.net/forum?id=U6hEOZlDf5

Compressor summary: The paper proposes a novel method for relative object pose estimation using a hypothesis-and-verification framework with 3D transformations, which achieves high accuracy and robustness.

Provable Benefits of Multi-task RL under Non-Markovian Decision Making Processes

https://openreview.net/forum?id=U6Qulbv2qT

Compressor summary: The paper investigates whether multi-task reinforcement learning can improve sample efficiency in more general sequential decision making problems, such as partially observable MDPs and predictive state representations, by using a joint model class for tasks and measuring its complexity with the $\eta$-bracketing number.

Sin3DM: Learning a Diffusion Model from a Single 3D Textured Shape

https://openreview.net/forum?id=U0IOMStUQ8

Compressor summary: Sin3DM is a diffusion model that learns from low-dimensional latent space to create realistic 3D shapes with different textures and geometry, and supports various editing tasks.

ODEFormer: Symbolic Regression of Dynamical Systems with Transformers

https://openreview.net/forum?id=TzoHLiGVMo

Compressor summary: ODEFormer is a new transformer that can learn ODE systems from single trajectory observations, beating existing methods in accuracy and robustness, and being faster to infer.

Scaling Laws for Associative Memories

https://openreview.net/forum?id=Tzh6xAJSll

Compressor summary: The paper studies associative memory mechanisms in a model using high-dimensional matrices of embeddings, deriving scaling laws and evaluating estimators with numerical experiments and visualizations.

FFB: A Fair Fairness Benchmark for In-Processing Group Fairness Methods

https://openreview.net/forum?id=TzAJbTClAz

Compressor summary: The paper presents FFB, an open-source framework for evaluating group fairness methods in machine learning using standardized benchmarks and extensive analysis.

Safe RLHF: Safe Reinforcement Learning from Human Feedback

https://openreview.net/forum?id=TyFrPOKYXw

Compressor summary: The text proposes a novel algorithm called Safe Reinforcement Learning from Human Feedback (Safe RLHF) for training large language models (LLMs) that balance between being helpful and harmless, using human feedback to adjust the trade-off during fine-tuning.

Multi-task Learning with 3D-Aware Regularization

https://openreview.net/forum?id=TwBY17Hgiy

Compressor summary: The paper proposes a 3D-aware regularizer that improves the performance of deep neural networks on multiple computer vision tasks by projecting features into a shared space and decoding them through differentiable rendering, reducing noisy cross-task correlations.

Towards Reliable and Efficient Backdoor Trigger Inversion via Decoupling Benign Features

https://openreview.net/forum?id=Tw9wemV6cb

Compressor summary: The paper presents a novel BTI method that decouples benign features from backdoor features, which improves efficiency and robustness against backdoor threats. The paper also uses the new BTI method to develop effective backdoor-removal and pre-processing defenses.

Defending Against Transfer Attacks From Public Models

https://openreview.net/forum?id=Tvwf4Vsi5F

Compressor summary: The paper proposes a realistic threat model for adversarial attacks using surrogate models and introduces PubDef, a game-theoretic defense method that significantly improves robustness against transfer attacks.

TEST: Text Prototype Aligned Embedding to Activate LLM's Ability for Time Series

https://openreview.net/forum?id=Tuh4nZVb0g

Compressor summary: The paper proposes TEST, a method to convert time-series (TS) data into a format that large language models (LLMs) can handle, and shows that it can achieve competitive results on various TS tasks without sacrificing LLM's language abilities.

Training Bayesian Neural Networks with Sparse Subspace Variational Inference

https://openreview.net/forum?id=TskzCtpMEO

Compressor summary: SSVI is a method for creating sparse Bayesian neural networks that maintains sparsity throughout training and inference, achieving significant compression and robustness while preserving performance.

Spatially-Aware Transformers for Embodied Agents

https://openreview.net/forum?id=Ts95eXsPBc

Compressor summary: The paper introduces spatially-aware transformers for episodic memory in AI that consider both temporal and spatial dimensions, improving efficiency and accuracy in various tasks.

Large Content And Behavior Models To Understand, Simulate, And Optimize Content And Behavior

https://openreview.net/forum?id=TrKq4Wlwcz

Compressor summary: The paper introduces Large Content and Behavior Models (LCBMs) that incorporate "behavior tokens" in LLM training to improve communication effectiveness by predicting and optimizing receiver behavior.

Sample-Efficient Linear Representation Learning from Non-IID Non-Isotropic Data

https://openreview.net/forum?id=Tr3fZocrI6

Compressor summary: The text proposes a method called De-bias & Feature-Whiten to improve representation learning from heterogeneous data sources by reducing biases and noise dependencies, leading to better generalization and computational efficiency.

Plug-and-Play: An Efficient Post-training Pruning Method for Large Language Models

https://openreview.net/forum?id=Tr0lPx9woF

Compressor summary: The paper introduces a new post-training pruning method for large language models that considers both weight and activations, preserves important weights, and improves performance and efficiency.

LILO: Learning Interpretable Libraries by Compressing and Documenting Code

https://openreview.net/forum?id=TqYbAWKMIe

Compressor summary: LILO is a neurosymbolic framework that creates and documents reusable code libraries using LLM-guided synthesis, Stitch's refactoring, and AutoDoc's natural language documentation.

GeoLLM: Extracting Geospatial Knowledge from Large Language Models

https://openreview.net/forum?id=TqL2xBwXP3

Compressor summary: GeoLLM is a new method that leverages language models' spatial information and map data to perform better in geospatial prediction tasks than existing methods.

Meta Continual Learning Revisited: Implicitly Enhancing Online Hessian Approximation via Variance Reduction

https://openreview.net/forum?id=TpD2aG1h0D

Compressor summary: Variance Reduced Meta-Continual Learning (VR-MCL) combines regularization-based and meta-learning approaches to improve knowledge transfer and reduce forgetting in continual learning.

Generative Adversarial Equilibrium Solvers

https://openreview.net/forum?id=TlyiaPXaVN

Compressor summary: The authors propose GAES, a method using generative adversarial networks to compute general game-theoretic equilibria from samples, and apply it to various settings including normal-form games, competitive economies, and environmental policy.

Understanding and Mitigating the Label Noise in Pre-training on Downstream Tasks

https://openreview.net/forum?id=TjhUtloBZU

Compressor summary: This paper studies how noise in pre-training datasets affects model generalization and proposes a method (NMTune) to improve both in-domain and out-of-domain performance on vision and language models.

NOLA: Networks as Linear Combination of Low Rank Random Basis

https://openreview.net/forum?id=TjfXcDgvzk

Compressor summary: NOLA is a method to efficiently adapt large language models for specific tasks by reducing parameters using linear combinations of random matrices and optimizing their coefficients.

Doubly Robust Proximal Causal Learning for Continuous Treatments

https://openreview.net/forum?id=TjGJFkU3xL

Compressor summary: The paper proposes a kernel-based doubly robust estimator for continuous treatments in proximal causal learning, showing its consistency, efficiency, and performance.

Learning Hierarchical World Models with Adaptive Temporal Abstractions from Discrete Latent Dynamics

https://openreview.net/forum?id=TjCDNssXKU

Compressor summary: THICK is an algorithm that learns a hierarchical world model for better planning and reinforcement learning by updating lower levels sparsely and higher levels when contexts change.

Probabilistically Rewired Message-Passing Neural Networks

https://openreview.net/forum?id=Tj6Wcx7gVk

Compressor summary: PR-MPNNs learn to add and remove edges in graphs based on their relevance to the prediction task, improving expressiveness and performance over random methods and existing graph neural networks.

On the Foundations of Shortcut Learning

https://openreview.net/forum?id=Tj3xLVuE9f

Compressor summary: The paper investigates how predictivity and availability affect feature use in deep-learning models, finding that nonlinear models tend to over-rely on less predictive but more available features.

Waxing-and-Waning: a Generic Similarity-based Framework for Efficient Self-Supervised Learning

https://openreview.net/forum?id=TilcG5C8bN

Compressor summary: SimWnW is a similarity-based efficient self-supervised learning framework that reduces computation costs by removing less important regions in augmented images and feature maps without compromising accuracy.

Overthinking the Truth: Understanding how Language Models Process False Demonstrations

https://openreview.net/forum?id=Tigr1kMDZy

Compressor summary: The text discusses the risks of harmful imitation by language models through two phenomena called overthinking and false induction heads, which can be studied by examining their internal representations.

DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models

https://openreview.net/forum?id=Th6NyL07na

Compressor summary: The DoLa approach reduces hallucinations in large language models by contrasting different layers to obtain a better next-token distribution and surface factual knowledge.

Patched Denoising Diffusion Models For High-Resolution Image Synthesis

https://openreview.net/forum?id=TgSRPRz8cI

Compressor summary: Patch-DM is a denoising diffusion model that generates high-resolution images by combining features from neighboring patches and outperforms previous methods with less memory complexity.

Momentum Benefits Non-iid Federated Learning Simply and Provably

https://openreview.net/forum?id=TdhkAcXkRi

Compressor summary: This paper explores how adding momentum to federated learning algorithms FedAvg and SCAFFOLD can improve their performance and convergence without making restrictive assumptions or adjustments.

Towards Transparent Time Series Forecasting

https://openreview.net/forum?id=TYXtXLYHpR

Compressor summary: The paper proposes a top-down framework called bi-level transparency for time series forecasting with machine learning models to ensure interpretability and trustworthiness in decision-making systems.

Learning Semantic Proxies from Visual Prompts for Parameter-Efficient Fine-Tuning in Deep Metric Learning

https://openreview.net/forum?id=TWVMVPx2wO

Compressor summary: The paper proposes a framework for fine-tuning pre-trained models for deep metric learning tasks using visual prompts, which improves performance with fewer parameters and less computation.

Enhancing Human-AI Collaboration Through Logic-Guided Reasoning

https://openreview.net/forum?id=TWC4gLoAxY

Compressor summary: The paper proposes a framework to use logical rules derived from data to understand human goals and guide human-like agents in collaboration tasks.

Towards Seamless Adaptation of Pre-trained Models for Visual Place Recognition

https://openreview.net/forum?id=TVg6hlfsKa

Compressor summary: The paper proposes a novel method to adapt pre-trained vision models for visual place recognition, using hybrid adaptation and mutual nearest neighbor local feature loss, achieving better performance with less data and time.

Zeroth-Order Optimization Meets Human Feedback: Provable Learning via Ranking Oracles

https://openreview.net/forum?id=TVDUVpgu9s

Compressor summary: ZO-RankSGD is a novel algorithm that optimizes black-box functions with ranking feedback and improves image quality in diffusion generative models using human guidance.

Distributionally Robust Optimization with Bias & Variance Reduced Gradients

https://openreview.net/forum?id=TTrzgEZt9s

Compressor summary: Prospect is a fast and easy-to-tune algorithm for distributionally robust optimization with spectral risk-based uncertainty sets and $f$-divergence penalty, achieving linear convergence and outperforming existing methods on various benchmarks.

Fast Hyperboloid Decision Tree Algorithms

https://openreview.net/forum?id=TTonmgTT9X

Compressor summary: Hyperbolic geometry helps machine learning capture hierarchical structures in data, but its computational challenges are overcome by HyperDT, a decision tree algorithm that adapts to hyperbolic space without Riemannian optimization.

Feature-aligned N-BEATS with Sinkhorn divergence

https://openreview.net/forum?id=TS8HoIWAPQ

Compressor summary: Feature-aligned N-BEATS is a time series forecasting model that learns invariant features across multiple domains while preserving interpretability and performance.

Test-time Adaption against Multi-modal Reliability Bias

https://openreview.net/forum?id=TPZRq4FALB

Compressor summary: This paper introduces a new challenge in multi-modal test-time adaptation called reliability bias, proposes a novel method RFRA that adaptively modulates attention between modalities and uses a robust objective function for multi-modal adaption, and provides two new benchmarks to evaluate the performance of methods under this challenge.

Discovering Failure Modes of Text-guided Diffusion Models via Adversarial Search

https://openreview.net/forum?id=TOWdQQgMJY

Compressor summary: SAGE is a method that systematically explores and discovers failures in text-guided diffusion models for image generation by using adversarial search, image classifiers, and human inspections.

TiC-CLIP: Continual Training of CLIP Models

https://openreview.net/forum?id=TLADT8Wrhn

Compressor summary: The paper introduces TiC benchmarks for continual learning on vision-language models, evaluates temporal robustness of existing models, and proposes a rehearsal-based approach to reduce compute costs.

Modulate Your Spectrum in Self-Supervised Learning

https://openreview.net/forum?id=TKqMmKlmA7

Compressor summary: Spectral Transformation (ST) is a framework that modulates the spectrum of embedding in self-supervised learning to prevent dimensional collapse and improve representation learning, with IterNorm with trace loss being a novel effective instance of ST.

Structural Inference with Dynamics Encoding and Partial Correlation Coefficients

https://openreview.net/forum?id=TKnzPdyeJu

Compressor summary: The paper presents a new structural inference method using variational dynamics encoder and partial correlation coefficients, which is scalable, accurate, and robust for one- and multi-dimensional feature spaces.

Querying Easily Flip-flopped Samples for Deep Active Learning

https://openreview.net/forum?id=THUBTfSAS2

Compressor summary: The paper proposes a new active learning strategy based on predictive uncertainty measured by the least disagree metric, which outperforms existing methods on various datasets and deep neural networks.

Harnessing Density Ratios for Online Reinforcement Learning

https://openreview.net/forum?id=THJEa8adBn

Compressor summary: The paper proposes GLOW and HYGLOW algorithms for online and hybrid reinforcement learning that use density ratios and exploration under coverability condition.

Proximal Policy Gradient Arborescence for Quality Diversity Reinforcement Learning

https://openreview.net/forum?id=TFKIfhvdmZ

Compressor summary: The text introduces PPGA, a new algorithm that combines QD and RL to improve exploration and skill discovery for robot learning.

Neural Fine-Tuning Search for Few-Shot Learning

https://openreview.net/forum?id=T7YV5UZKBc

Compressor summary: The paper proposes a neural architecture search method to find the best adaptation strategy for few-shot recognition using adapters and freezing layers.

Neur2RO: Neural Two-Stage Robust Optimization

https://openreview.net/forum?id=T5Xb0iGCCv

Compressor summary: Neur2RO is a fast and efficient machine learning algorithm for solving two-stage robust optimization problems, especially on large-scale instances.

Emergent Communication with Conversational Repair

https://openreview.net/forum?id=Sy8upuD6Bw

Compressor summary: The study shows that conversational repair improves language performance in signaling games with feedback and noise, challenging the role of compositionality for generalization.

Variational Bayesian Last Layers

https://openreview.net/forum?id=Sx7BIiPzys

Compressor summary: The paper proposes a deterministic variational approach to train Bayesian last layer neural networks that improves uncertainty estimation and predictive accuracy.

CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing

https://openreview.net/forum?id=Sx038qxjek

Compressor summary: CRITIC is a framework that helps large language models improve their outputs by using external tools for validation and revision, similar to how humans use search engines and code interpreters.

Sliced Wasserstein Estimation with Control Variates

https://openreview.net/forum?id=StYc4hQAEi

Compressor summary: The paper proposes efficient control variates to reduce variance in Monte Carlo estimation of sliced Wasserstein distances for comparing probability measures in various applications.

Provable Robust Watermarking for AI-Generated Text

https://openreview.net/forum?id=SsmT8aO45L

Compressor summary: The paper proposes a watermarking method for large language models that ensures high-quality generated text, correct detection, and robustness against text manipulation.

IDEAL: Influence-Driven Selective Annotations Empower In-Context Learners in Large Language Models

https://openreview.net/forum?id=Spp2i1hKwV

Compressor summary: The paper proposes a selective annotation method for in-context learning that minimizes costs and improves quality by using a directed graph and a greedy algorithm to choose influential unlabeled data subsets.

Generalized Schrödinger Bridge Matching

https://openreview.net/forum?id=SoismgeX7z

Compressor summary: The paper introduces GSBM, a generalized distribution matching algorithm for training diffusion or flow models that can handle nonlinear state costs and preserves feasible transport maps throughout training, leading to improved scalability and convergence.

R-EDL: Relaxing Nonessential Settings of Evidential Deep Learning

https://openreview.net/forum?id=Si3YFA641c

Compressor summary: The paper proposes R-EDL, a modified Evidential Deep Learning method that adjusts prior weight and optimizes Dirichlet PDF to improve predictive uncertainty estimation.

Self-Consuming Generative Models Go MAD

https://openreview.net/forum?id=ShjMHfmPs0

Compressor summary: The quality or diversity of future generative models will decline without enough fresh real data in each generation of an autophagous loop, which is called Model Autophagy Disorder (MAD).

Accurate Forgetting for Heterogeneous Federated Continual Learning

https://openreview.net/forum?id=ShQrnAsbPI

Compressor summary: The paper proposes accurate forgetting (AF) and a generative-replay method for federated continual learning, which selectively uses previous knowledge in federated networks based on a normalizing flow model.

OpenNerf: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views

https://openreview.net/forum?id=SgjAojPKb3

Compressor summary: OpenNeRF is a method that uses visual-language models to segment arbitrary concepts from 3D scenes in images without relying on predefined training sets or point cloud representations.

Dropout-Based Rashomon Set Exploration for Efficient Predictive Multiplicity Estimation

https://openreview.net/forum?id=Sf2A2PUXO3

Compressor summary: The text proposes a framework that uses dropout techniques to explore and measure the Rashomon set, which consists of multiple competing models with almost-equally-optimal performance but conflicting outputs, and use this information to mitigate predictive multiplicity in classification tasks.

Incremental Randomized Smoothing Certification

https://openreview.net/forum?id=SdeAPV1irk

Compressor summary: IRS is a method that reduces computational cost and maintains strong robustness guarantees for certifying modified deep neural networks against adversarial attacks using incremental robustness certification for randomized smoothing.

Horizon-Free Regret for Linear Markov Decision Processes

https://openreview.net/forum?id=SdBApv9iT4

Compressor summary: The paper proposes a new horizon-free regret bound for linear MDPs, where it estimates value functions and confidence sets directly, instead of estimating transition models and inhomogeneous value functions.

Efficient local linearity regularization to overcome catastrophic overfitting

https://openreview.net/forum?id=SZzQz8ikwg

Compressor summary: The paper proposes ELLE, a regularization term for single-step adversarial training that reduces catastrophic overfitting and is more efficient than previous methods.

OWL: A Large Language Model for IT Operations

https://openreview.net/forum?id=SZOQ9RKYJu

Compressor summary: The paper introduces OWL, a large language model for IT operations, and proposes HMCE, a method to extend homogeneous Markov context, improving performance on IT tasks.

When Semantic Segmentation Meets Frequency Aliasing

https://openreview.net/forum?id=SYBdkHcXXK

Compressor summary: The paper analyzes hard pixel errors in semantic segmentation, identifies aliasing as a major cause, and proposes new methods to improve segmentation by removing or adjusting high-frequency components.

VCR-Graphormer: A Mini-batch Graph Transformer via Virtual Connections

https://openreview.net/forum?id=SUUrkC3STJ

Compressor summary: The text proposes a new graph learning method called VCR-Graphormer, which uses personalized PageRank tokenization and virtual connections to enable efficient mini-batch training and improve performance on node classification tasks.

When can transformers reason with abstract symbols?

https://openreview.net/forum?id=STUGfUz8ob

Compressor summary: Transformer models struggle with generalizing to new symbols when trained on abstract tasks, requiring large amounts of data or modifications to improve performance.

CABINET: Content Relevance-based Noise Reduction for Table Question Answering

https://openreview.net/forum?id=SQrHpTllXa

Compressor summary: CABINET improves LLMs' question-answering over tables by filtering out irrelevant information using an unsupervised relevance scorer and a weakly supervised parsing module.

Social-Transmotion: Promptable Human Trajectory Prediction

https://openreview.net/forum?id=SQpnEfv9WH

Compressor summary: Social-Transmotion is a transformer-based model that leverages diverse visual cues to enhance human trajectory prediction by using prompts inspired by NLP.

The Devil is in the Neurons: Interpreting and Mitigating Social Biases in Language Models

https://openreview.net/forum?id=SQGUDc9tC8

Compressor summary: The authors introduce Social Bias Neurons to identify and suppress harmful information in pre-trained language models, using a novel technique called Integrated Gap Gradients (IG$^2$).

Neuron Activation Coverage: Rethinking Out-of-distribution Detection and Generalization

https://openreview.net/forum?id=SNGXbZtK6Q

Compressor summary: The paper proposes neuron activation coverage (NAC) as a simple and effective measure for detecting out-of-distribution inputs in neural networks and evaluating model robustness.

Controlled Text Generation via Language Model Arithmetic

https://openreview.net/forum?id=SLw9fp4yI6

Compressor summary: Model arithmetic is a new way to combine and adjust large language models for better text generation with less retraining and more control.

Rethinking the symmetry-preserving circuits for constrained variational quantum algorithms

https://openreview.net/forum?id=SL7djdVpde

Compressor summary: The paper explores how to use a quantum algorithm called Hamming Weight preserving ansatz to incorporate symmetries and constraints in physical systems, which could lead to better performance in problems like ground state energy estimation and feature selection.

Interventional Fairness on Partially Known Causal Graphs: A Constrained Optimization Approach

https://openreview.net/forum?id=SKulT2VX9p

Compressor summary: The paper proposes a framework for achieving causal fairness in machine learning using partially known causal graphs and interventions, and shows its effectiveness on various datasets.

Label-Agnostic Forgetting: A Supervision-Free Unlearning in Deep Models

https://openreview.net/forum?id=SIZWiya7FE

Compressor summary: The text introduces Label-Agnostic Forgetting (LAF), a supervision-free method for removing information from forgotten data in a trained model, which performs well without labels and excels in semi-supervised scenarios.

Mask-based modeling for Neural Radiance Fields

https://openreview.net/forum?id=SEiuSzlD1d

Compressor summary: The paper proposes a new method called MRVM-NeRF that improves 3D representation learning for NeRF models by using masks to guide ray and view modeling, leading to better generalization and detail capture across diverse scenes.

Prediction without Preclusion: Recourse Verification with Reachable Sets

https://openreview.net/forum?id=SCQfYpdoGE

Compressor summary: The paper proposes a method to check if machine learning models can offer solutions when they make decisions that affect people's access to resources like loans or jobs, and shows how this can help improve fairness in these models.

THOUGHT PROPAGATION: AN ANALOGICAL APPROACH TO COMPLEX REASONING WITH LARGE LANGUAGE MODELS

https://openreview.net/forum?id=SBoRhRCzM3

Compressor summary: Thought Propagation (TP) improves complex reasoning in large language models by solving analogous problems and reusing their solutions to help with multi-step reasoning tasks.

Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition

https://openreview.net/forum?id=SBj2Qdhgew

Compressor summary: The paper proposes an information-theoretic approach to analyze fairness trade-offs in federated learning with respect to sensitive attributes, identifying three sources of unfairness and a convex optimization problem that defines the theoretical limits of accuracy and fairness trade-offs.

A Study of Bayesian Neural Network Surrogates for Bayesian Optimization

https://openreview.net/forum?id=SA19ijj44B

Compressor summary: Bayesian neural networks can optimize expensive objective functions and handle non-stationarity better than standard Gaussian process surrogates, but their performance depends on the problem and the inference method used.

Understanding Transferable Representation Learning and Zero-shot Transfer in CLIP

https://openreview.net/forum?id=S5yOuNfSA0

Compressor summary: The paper studies the theoretical aspects of CLIP, a multi-modal learning method that uses contrastive pretraining to learn joint image and text representations, and proposes a new approach based on its insights.

Complex priors and flexible inference in recurrent circuits with dendritic nonlinearities

https://openreview.net/forum?id=S5aUhpuyap

Compressor summary: The text describes a recurrent network model with dendritic nonlinearities and stochastic somatic integration that can implicitly represent priors over latent variables and encode task-specific posteriors for flexible inference.

Do Generated Data Always Help Contrastive Learning?

https://openreview.net/forum?id=S5EqslEHnz

Compressor summary: Adaptive Inflation (AdaInf) is a data-centric strategy that improves contrastive learning by adjusting the balance between data inflation and augmentation, achieving state-of-the-art results on benchmark datasets without extra computation cost.

A Variational Framework for Estimating Continuous Treatment Effects with Measurement Error

https://openreview.net/forum?id=S46Knicu56

Compressor summary: The paper proposes a general variational framework to estimate the average dose-response function with error-contaminated treatment data, using neural networks to handle unobserved treatments and confounding variables.

SmartPlay : A Benchmark for LLMs as Intelligent Agents

https://openreview.net/forum?id=S2oTVrlcp3

Compressor summary: SmartPlay is a benchmark and methodology to evaluate large language models' abilities as intelligent agents by testing them on 6 different games that challenge various capabilities of an LLM agent.

Guiding Instruction-based Image Editing via Multimodal Large Language Models

https://openreview.net/forum?id=S1RKWSyZ2Y

Compressor summary: MGIE uses multimodal large language models to generate expressive image editing instructions and improve controllability and flexibility of image manipulation.

Kill Two Birds with One Stone: Rethinking Data Augmentation for Deep Long-tailed Learning

https://openreview.net/forum?id=RzY9qQHUXy

Compressor summary: The paper proposes an adaptive data augmentation method called DODA that addresses the challenges of long-tailed learning by allowing each class to choose appropriate augmentation methods based on its distribution, reducing the imbalance and improving performance.

$t^3$-Variational Autoencoder: Learning Heavy-tailed Data with Student's t and Power Divergence

https://openreview.net/forum?id=RzNlECeoOB

Compressor summary: The paper introduces $t^3$VAE, a modified VAE framework that uses Student's t-distributions for the prior, encoder, and decoder to better capture rare features in heavy-tailed data.

A Stable, Fast, and Fully Automatic Learning Algorithm for Predictive Coding Networks

https://openreview.net/forum?id=RyUvzda8GH

Compressor summary: Incremental predictive coding (iPC) is an improved algorithm for training neuroscience-inspired models that is more efficient, stable, biologically plausible, and performs better on image and language tasks.

BrainLM: A foundation model for brain activity recordings

https://openreview.net/forum?id=RwI7ZEfR27

Compressor summary: The Brain Language Model is a large-scale model that can predict clinical variables, forecast future brain states, identify functional networks, and generate interpretable latent representations from fMRI data, enabling novel insights into human brain activity dynamics.

What's In My Big Data?

https://openreview.net/forum?id=RvfPnOkPV4

Compressor summary: WIMBD is a platform that analyzes large text corpora, revealing their content quality and diversity, and finds issues like duplicate and low-quality data, as well as contamination in benchmark datasets.

A Benchmark on Robust Semi-Supervised Learning in Open Environments

https://openreview.net/forum?id=RvUVMjfp8i

Compressor summary: SSL can benefit from unlabeled data but needs to be improved for consistent and diverse data sources in open environments.

Meta-Evolve: Continuous Robot Evolution for One-to-many Policy Transfer

https://openreview.net/forum?id=RthOl4jHw5

Compressor summary: Meta-Evolve is a method that uses continuous robot evolution to efficiently transfer policies from a source robot to multiple target robots, achieving significant improvements in simulation cost.

Simplifying Transformer Blocks

https://openreview.net/forum?id=RtDok9eS3s

Compressor summary: The authors propose a simplified design for deep Transformers by removing some components without losing speed or performance.

On Error Propagation of Diffusion Models

https://openreview.net/forum?id=RtAct1E2zS

Compressor summary: The paper develops a theory of error propagation in diffusion models and proposes a regularization method based on cumulative error to improve their performance.

A Primal-Dual Approach to Solving Variational Inequalities with General Constraints

https://openreview.net/forum?id=RsztjXcvUf

Compressor summary: The paper proposes a warm-starting technique for solving variational inequalities and shows its faster convergence than existing methods, while also relaxing smoothness assumptions.

Optimal Sketching for Residual Error Estimation for Matrix and Vector Norms

https://openreview.net/forum?id=RsJwmWvE6Q

Compressor summary: The text studies how to estimate residual errors for matrix and vector norms using linear sketches, improving previous bounds and allowing for faster algorithms with sparse matrices.

COMPARATOR: Reference-free machine translation evaluation by inter-system comparison

https://openreview.net/forum?id=Rry1SeSOQL

Compressor summary: The text proposes a new approach to Machine Translation Evaluation that compares translations without references using natural language inference and synthetic data, achieving state-of-the-art results on various benchmarks.

Achieving Human Parity in Content-Grounded Datasets Generation

https://openreview.net/forum?id=RjYKTQ0L0W

Compressor summary: The paper proposes a novel method for generating high-quality content-grounded data for tasks like LFQA and summarization, which performs well in human evaluation and outperforms existing models in faithfulness.

Learning to make adherence-aware advice

https://openreview.net/forum?id=RgELE1dQXx

Compressor summary: The paper proposes a model for human-AI interactions that considers human adherence and selective advice, and presents learning algorithms for optimal policy and timing of advice.

Hierarchical Data-efficient Representation Learning for Tertiary Structure-based RNA Design

https://openreview.net/forum?id=RemfXx7ebP

Compressor summary: The study presents a data-driven pipeline for RNA sequence design based on tertiary structure using a large curated dataset, structural modeling, and contrastive learning with secondary structures as prior knowledge.

NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers

https://openreview.net/forum?id=Rc7dAwVL3v

Compressor summary: NaturalSpeech 2 is a text-to-speech system that uses a neural audio codec and a diffusion model to generate high-quality, diverse speech from large datasets, even with zero-shot learning.

Tree Search-Based Policy Optimization under Stochastic Execution Delay

https://openreview.net/forum?id=RaqZX9LSGA

Compressor summary: The paper introduces a new formalism for Markov decision processes with random delays, called stochastic delayed execution MDPs, and proposes a model-based algorithm, Delayed EfficientZero, that achieves optimal performance by searching over the class of Markov policies.

Zero-Mean Regularized Spectral Contrastive Learning

https://openreview.net/forum?id=RZBy8oHTz4

Compressor summary: The paper proposes a simple modification to the spectral contrastive loss that improves performance and robustness in self-supervised learning by introducing zero-mean regularization for negative pairs.

Harnessing Explanations: LLM-to-LM Interpreter for Enhanced Text-Attributed Graph Representation Learning

https://openreview.net/forum?id=RXFVcynVe1

Compressor summary: The paper introduces a novel technique to improve graph learning with text by using language model explanations as features, which boost performance and speed on various text-attributed graph datasets.

Sufficient conditions for offline reactivation in recurrent neural networks

https://openreview.net/forum?id=RVrINT6MT7

Compressor summary: This study develops a mathematical framework to explain how noisy neural networks can autonomously reactivate the same network states during periods of quiescence, such as sleep, based on their task optimization.

TAIL: Task-specific Adapters for Imitation Learning with Large Pretrained Models

https://openreview.net/forum?id=RRayv1ZPN3

Compressor summary: TAIL is a framework for adapting large pretrained models to new control tasks efficiently using parameter-efficient fine-tuning techniques like LoRA, with minimal data and without forgetting previous knowledge.

Ensemble Distillation for Unsupervised Constituency Parsing

https://openreview.net/forum?id=RR8y0WKrFv

Compressor summary: The paper proposes a novel ensemble method for unsupervised constituency parsing using tree averaging and multi-teacher distillation, which outperforms previous methods.

Efficient Instance-Optimal Finite-Sum Minimization

https://openreview.net/forum?id=RR70yWYenC

Compressor summary: The paper proposes a new method, SIOPT-Grad, for minimizing finite-sums with instance-optimal guarantees, which requires fewer gradient evaluations than existing methods.

Weaker MVI Condition: Extragradient Methods with Multi-Step Exploration

https://openreview.net/forum?id=RNGUbTYSjk

Compressor summary: The paper presents a new framework of algorithms based on extragradient for min-max optimization in structured problems, with adaptive exploration to avoid limit cycles.

Offline Data Enhanced On-Policy Policy Gradient with Provable Guarantees

https://openreview.net/forum?id=RMgqvQGTwH

Compressor summary: The paper proposes a new hybrid RL algorithm that combines an on-policy actor-critic method with offline data, achieving both theoretical and empirical benefits.

Generalized Neural Sorting Networks with Error-Free Differentiable Swap Functions

https://openreview.net/forum?id=RLSWbk9kPw

Compressor summary: The paper proposes a neural sorting network for abstract inputs, improves the swap function's differentiability, and uses a Transformer network with multi-head attention to capture dependencies and leverage model capacity.

Weakly Supervised Virus Capsid Detection with Image-Level Annotations in Electron Microscopy Images

https://openreview.net/forum?id=RJDjSXNuAZ

Compressor summary: The paper proposes a weakly supervised object detection method that uses image-level annotations and distills knowledge from a pre-trained model to detect viruses more efficiently than existing methods.

ConR: Contrastive Regularizer for Deep Imbalanced Regression

https://openreview.net/forum?id=RIuevDSK5V

Compressor summary: The text introduces ConR, a contrastive regularizer that improves deep imbalanced regression by modeling global and local label similarities in feature space and penalizing incorrect proximities between samples.

Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting

https://openreview.net/forum?id=RIu5lyNXjT

Compressor summary: This work studies how different prompt formatting affects large language models' few-shot performance, and proposes FormatSpread to quickly evaluate a range of prompt formats for a task.

Don't Judge by the Look: A Motion Coherent Augmentation for Video Recognition

https://openreview.net/forum?id=RIcYTbpO38

Compressor summary: Motion Coherent Augmentation is a data augmentation method for video recognition that uses Hue Jittering to introduce appearance variation and improve motion pattern detection.

Graph Generation with $K^2$-trees

https://openreview.net/forum?id=RIEW6M9YoV

Compressor summary: The paper proposes a new method for generating graphs using a $K^2$ representation that captures the hierarchical structure of graphs and uses a Transformer-based architecture with a tree positional encoding scheme.

Demystifying Linear MDPs and Novel Dynamics Aggregation Framework

https://openreview.net/forum?id=RDSj6S8WJe

Compressor summary: This paper challenges the common assumption of linear MDPs and proposes a new hierarchical reinforcement learning framework called dynamics aggregation, which improves performance guarantees.

Scaling for Training Time and Post-hoc Out-of-distribution Detection Enhancement

https://openreview.net/forum?id=RDSTjtnqCg

Compressor summary: The paper discusses OOD detection methods, proposing SCALE and ISH to improve OOD/ID accuracy without compromising performance.

PAE: Reinforcement Learning from External Knowledge for Efficient Exploration

https://openreview.net/forum?id=R7rZUSGOPD

Compressor summary: PAE is a framework that teaches AI agents to learn from external knowledge, improving their exploration efficiency and task performance.

Memory-Consistent Neural Networks for Imitation Learning

https://openreview.net/forum?id=R3Tf7LDdX4

Compressor summary: Memory-consistent neural networks (MCNNs) improve imitation learning by limiting policy errors using memory regions from expert demonstrations.

MetaTool Benchmark: Deciding Whether to Use Tools and Which to Use

https://openreview.net/forum?id=R0c2qtalgG

Compressor summary: The paper introduces MetaTool, a benchmark to evaluate if large language models can choose appropriate tools for various tasks, and finds that most LLMs struggle with this task but have room for improvement.

Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI

https://openreview.net/forum?id=QzTpTRVtrP

Compressor summary: Large Brain Model (LaBraM) is a unified foundation model for EEG that enables cross-dataset learning and achieves state-of-the-art performance on various downstream tasks.

A Generative Pre-Training Framework for Spatio-Temporal Graph Transfer Learning

https://openreview.net/forum?id=QyFm3D3Tzi

Compressor summary: GPDiff is a new method that pre-trains STG models with a diffusion model and a transformer network to adapt to different cities' data and achieve better performance on various smart city tasks.

TorchRL: A data-driven decision-making library for PyTorch

https://openreview.net/forum?id=QxItoEAVMb

Compressor summary: TorchRL is a PyTorch library that simplifies algorithm development and improves efficiency in decision and control tasks, with TensorDict as a new flexible primitive.

Exploring the Relationship Between Model Architecture and In-Context Learning Ability

https://openreview.net/forum?id=Qwq4cpLtoX

Compressor summary: This study evaluates fifteen model architectures on synthetic in-context learning tasks and finds that contemporary architectures perform best as task complexity increases, while state-space models are more robust than transformers.

Is Generalized Dynamic Novel View Synthesis from Monocular Videos Possible Today?

https://openreview.net/forum?id=QuVlUn4T2G

Compressor summary: The paper explores whether a generalized method for synthesizing novel views from dynamic scenes in monocular videos is possible, and finds that it is partially feasible without scene-specific appearance optimization but requires consistent depth estimates.

Algorithms for Caching and MTS with reduced number of predictions

https://openreview.net/forum?id=QuIiLSktO4

Compressor summary: The text introduces parsimonious algorithms for caching and multi-item single-server settings that use predictions efficiently and maintain good performance with varying prediction error.

PRIME: Prioritizing Interpretability in Failure Mode Extraction

https://openreview.net/forum?id=QrEHs9w5UF

Compressor summary: The paper proposes a new method to explain why image classification models fail by using human-understandable tags of images instead of clustering in feature space, resulting in better and simpler descriptions.

It's Never Too Late: Fusing Acoustic Information into Large Language Models for Automatic Speech Recognition

https://openreview.net/forum?id=QqjFHyQwtF

Compressor summary: The authors propose a new method, Uncertainty-Aware Dynamic Fusion (UADF), to improve generative error correction by incorporating acoustic information into the language model's output, leading to better speech recognition results.

Learning Multi-Agent Communication from Graph Modeling Perspective

https://openreview.net/forum?id=Qox9rO0kN0

Compressor summary: The paper proposes CommFormer, a method that learns a communication graph for multiple intelligent agents to improve collaboration in AI applications.

Localizing and Editing Knowledge In Text-to-Image Generative Models

https://openreview.net/forum?id=Qmw9ne6SOQ

Compressor summary: Text-to-image diffusion models use multiple components in the conditional UNet to store knowledge about different visual attributes, while only one causal state in the text-encoder is shared across them, enabling fast data-free model editing with DiffQuickFix.

LanguageBind: Extending Video-Language Pretraining to N-modality by Language-based Semantic Alignment

https://openreview.net/forum?id=QmZKc7UZCy

Compressor summary: LanguageBind extends video-language pretraining to multiple modalities by using language as a bind, leading to improved performance in various tasks.

Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMs

https://openreview.net/forum?id=QmYNBVukex

Compressor summary: This paper proposes a data selection method that improves pre-fine-tuning of pre-trained language models by aligning their distributions and achieves better performance, faster than existing techniques.

Towards Category Unification of 3D Single Object Tracking on Point Clouds

https://openreview.net/forum?id=QlqdXrzzD1

Compressor summary: The paper introduces unified models for 3D object tracking that adapt to different categories by encoding shape and size attributes using a transformer-based network called AdaFormer.

Early Neuron Alignment in Two-layer ReLU Networks with Small Initialization

https://openreview.net/forum?id=QibPzdVrRu

Compressor summary: The paper analyzes how a two-layer ReLU network with gradient flow and small initialization learns from well-separated input data, showing an upper bound on the time for neurons to align and convergence rate of the loss.

Efficient Heterogeneous Meta-Learning via Channel Shuffling Modulation

https://openreview.net/forum?id=QiJuMJl0QS

Compressor summary: The paper proposes a new heterogeneous meta-learning strategy that captures the multi-modality of task distribution using a neural permutation layer with sub-quadratic parameter complexity.

Learning Hierarchical Polynomials with Three-Layer Neural Networks

https://openreview.net/forum?id=QgwAYFrh9t

Compressor summary: The paper shows how three-layer neural networks can efficiently learn hierarchical polynomials over Gaussian distribution by feature learning, outperforming kernel methods and two-layer networks.

Scalable and Effective Implicit Graph Neural Networks on Large Graphs

https://openreview.net/forum?id=QcMdPYBwTu

Compressor summary: SEIGNN is a new efficient implicit GNN that uses mini-batch training and a stochastic solver to handle large graphs, while maintaining prediction accuracy.

Scaling Supervised Local Learning with Augmented Auxiliary Networks

https://openreview.net/forum?id=Qbf1hy8b7m

Compressor summary: AugLocal is a local learning method that enhances the synergy between layers and reduces memory usage in deep neural networks by selectively incorporating some layers into auxiliary networks and adjusting their depth.

OpenTab: Advancing Large Language Models as Open-domain Table Reasoners

https://openreview.net/forum?id=Qa0ULgosc9

Compressor summary: OpenTab is a table reasoning framework using LLMs that retrieves relevant tables, generates SQL programs, and conducts grounded inference for accurate responses on structured data.

MOFI: Learning Image Representations from Noisy Entity Annotated Images

https://openreview.net/forum?id=QQYpgReSRk

Compressor summary: MOFI is a new vision foundation model that learns from noisy image-text pairs using entity labels and achieves state-of-the-art performance on various tasks.

MovingParts: Motion-based 3D Part Discovery in Dynamic Radiance Field

https://openreview.net/forum?id=QQ6RgKYiQq

Compressor summary: MovingParts is a NeRF-based method for reconstructing dynamic scenes and discovering parts by tracking particle trajectories under both Eulerian and Lagrangian views.

Decodable and Sample Invariance Continuous Object Encoder

https://openreview.net/forum?id=QLoepRnoue

Compressor summary: HDFE is a vector representation method for continuous objects that works with neural networks and improves performance on tasks like function-to-function mapping and point cloud surface normal estimation.

Guaranteed Approximation Bounds for Mixed-Precision Neural Operators

https://openreview.net/forum?id=QJGj07PD9C

Compressor summary: The paper proposes a method to train Fourier Neural Operators (FNO) in mixed precision, reducing memory usage and increasing training speed while maintaining accuracy.

Less is More: One-shot Subgraph Reasoning on Large-scale Knowledge Graphs

https://openreview.net/forum?id=QHROe7Mfcb

Compressor summary: The paper proposes a one-shot subgraph reasoning method for efficient and adaptive knowledge graph reasoning using Personalized PageRank heuristics and optimal configuration searching.

Adaptive Sharpness-Aware Pruning for Robust Sparse Networks

https://openreview.net/forum?id=QFYVVwiAM8

Compressor summary: AdaSAP is a method that creates sparse deep learning models with better robustness to unseen input variations by optimizing the loss landscape through weight perturbations during pruning.

Masked Autoencoders with Multi-Window Local-Global Attention Are Better Audio Learners

https://openreview.net/forum?id=Q53QLftNkA

Compressor summary: The MW-MAE model uses a novel attention module to learn better general-purpose audio representations and improve performance on various downstream tasks, by capturing local and global information independently in each decoder block.

Lemur: Integrating Large Language Models in Automated Program Verification

https://openreview.net/forum?id=Q3YaCghZNt

Compressor summary: The text proposes a method to use LLMs with automated reasoners for program verification and shows its effectiveness on benchmarks.

MixSup: Mixed-grained Supervision for Label-efficient LiDAR-based 3D Object Detection

https://openreview.net/forum?id=Q1vkAhdI6j

Compressor summary: MixSup is a 3D object detection method that uses cheap cluster labels and some box labels to learn semantics and geometry, improving performance on various datasets.

SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression

https://openreview.net/forum?id=Q1u25ahSuy

Compressor summary: SpQR is a new compression technique that allows near-lossless LLMs on memory-limited devices by isolating outlier weights and compressing others to 3-4 bits, achieving up to 15% speedup with minimal accuracy loss.

A Simple and Effective Pruning Approach for Large Language Models

https://openreview.net/forum?id=PxoFut3dWW

Compressor summary: Wanda is a novel, straightforward, and effective pruning method for Large Language Models that does not require retraining or weight update.

Masked Completion via Structured Diffusion with White-Box Transformers

https://openreview.net/forum?id=PvyOYleymy

Compressor summary: The authors propose a new white-box architecture, CRATE-MAE, for large-scale unsupervised representation learning of images that is interpretable, structured, and efficient.

Quadratic models for understanding neural network dynamics

https://openreview.net/forum?id=PvJnX3dwsD

Compressor summary: Neural Quadratic Models can mimic the "catapult phase" and help analyze neural networks' generalization behavior.

Geographic Location Encoding with Spherical Harmonics and Sinusoidal Representation Networks

https://openreview.net/forum?id=PudduufFLa

Compressor summary: This paper introduces a new way to represent geographical data in machine learning models using a combination of spherical harmonic functions and sinusoidal networks, which improves performance and avoids artifacts.

Learning Unsupervised World Models for Autonomous Driving via Discrete Diffusion

https://openreview.net/forum?id=Psl75UCoZM

Compressor summary: The authors propose a novel world modeling approach that combines VQVAE and discrete diffusion to handle complex robotic observations and achieve better predictions than existing methods.

Learning to Solve Bilevel Programs with Binary Tender

https://openreview.net/forum?id=PsDFgTosqb

Compressor summary: The paper proposes deep learning methods for solving bilevel programs with discrete decision variables, comparing two types of neural networks and introducing an enhanced sampling technique.

A Linear Algebraic Framework for Counterfactual Generation

https://openreview.net/forum?id=PoDkdFQIu3

Compressor summary: The paper proposes a method to generate synthetic counterfactual data for estimating time-varying causal effects in real-world settings and shows its superior performance over existing methods.

Deep Geodesic Canonical Correlation Analysis for Covariance-Based Neuroimaging Data

https://openreview.net/forum?id=PnR1MNen7u

Compressor summary: DeepGeoCCA is a novel deep learning framework for multi-view and multi-modal neuroimaging data that enhances the correlation of latent representations while preserving geometric structures.

Fusing Models with Complementary Expertise

https://openreview.net/forum?id=PhMrGCMIRL

Compressor summary: The paper proposes a supervised learning method for fusing outputs of different expert models to improve generalization across tasks and domains in various AI applications, including text classification and summarization.

From Molecules to Materials: Pre-training Large Generalizable Models for Atomic Property Prediction

https://openreview.net/forum?id=PfPnugdxup

Compressor summary: This paper introduces Joint Multi-domain Pre-training (JMP), a method that uses data from multiple chemical domains to generate transferable atomic representations for various molecular tasks with high accuracy.

Making RL with Preference-based Feedback Efficient via Randomization

https://openreview.net/forum?id=Pe2lo3QOvo

Compressor summary: Our algorithm for RLHF using preferences over trajectories minimizes regret bound and query complexity by combining randomization, active learning, and Bayesian supervised learning oracles in linear MDPs and general nonlinear function approximation settings.

Never Train from Scratch: Fair Comparison of Long-Sequence Models Requires Data-Driven Priors

https://openreview.net/forum?id=PdaPky8MUn

Compressor summary: Pretraining with denoising objectives improves long-range dependency modeling across various architectures and reduces the gap between Transformers and state space models.

Cross$Q$: Batch Normalization in Deep Reinforcement Learning for Greater Sample Efficiency and Simplicity

https://openreview.net/forum?id=PczQtTsTIX

Compressor summary: CrossQ is a deep reinforcement learning algorithm that achieves high sample efficiency with low computational cost by using Batch Normalization and removing target networks, without relying on advanced bias-reduction schemes.

Pre-training with Synthetic Data Helps Offline Reinforcement Learning

https://openreview.net/forum?id=PcxQgtHGj2

Compressor summary: This paper shows that pre-training for offline deep reinforcement learning can be done with simple synthetic data, instead of language pre-training, to improve performance.

Synapse: Trajectory-as-Exemplar Prompting with Memory for Computer Control

https://openreview.net/forum?id=Pc8AU1aF5e

Compressor summary: Synapse is an agent that uses trajectory-as-exemplar prompting and memory to improve performance, efficiency, and generalizability in computer control tasks with large language models.

From Matching to Mixing: A Graph Interpolation Approach for SAT Instance Generation

https://openreview.net/forum?id=PXXuLvIH5r

Compressor summary: The paper proposes a graph interpolation method to modify SAT instances by exchanging clause pairs based on node correspondence learned from a matching network, with the option of adding Gumbel noise for hardness preservation.

SOHES: Self-supervised Open-world Hierarchical Entity Segmentation

https://openreview.net/forum?id=PXNrncg2DF

Compressor summary: SOHES is a novel self-supervised approach for open-world entity segmentation that requires no human annotations and captures entities' parts hierarchically.

Understanding the Effects of RLHF on LLM Generalisation and Diversity

https://openreview.net/forum?id=PXD3FAVHJT

Compressor summary: RLHF improves OOD generalization but reduces output diversity in LLMs, highlighting a tradeoff between these properties.

Integrating Planning and Deep Reinforcement Learning via Automatic Induction of Task Substructures

https://openreview.net/forum?id=PR6RMsxuW7