ICLR 2025 Papers — Page 24
International Conference on Learning Representations · 3704 papers
Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification
Hsun-Yu Kuo (Swiss Federal Institute of Technology in Lausanne), Pu-Jen Cheng (National Taiwan University)
ClassificationKnowledge DistillationData-Centric LearningTransformerLarge Language ModelTextFinance Related
🎯 What it does: Two weighted loss functions (IMP-Loss and DIMP-Loss) are proposed, which dynamically or statically weight the synthetic data generated by LLM through a quality checker and a diversity checker to align it with the real data distribution, thereby improving the performance of text classification models.
Not All Prompts Are Made Equal: Prompt-based Pruning of Text-to-Image Diffusion Models
Alireza Ganjdanesh (University of Maryland), Heng Huang (University of Maryland)
GenerationData SynthesisOptimizationKnowledge DistillationTransformerPrompt EngineeringDiffusion modelImageText
🎯 What it does: Adaptive Prompt-Tailored Pruning (APTP) is proposed in text-to-image diffusion models, achieving dynamic pruning based on prompts by learning a prompt router that allocates prompts of varying complexity to adjustable sub-networks.
Not-So-Optimal Transport Flows for 3D Point Cloud Generation
Ka-Hei Hui (Chinese University of Hong Kong), Arash Vahdat (NVIDIA)
GenerationData SynthesisOptimizationDiffusion modelFlow-based ModelPoint Cloud
🎯 What it does: A non-optimal transport flow matching model is proposed for 3D point cloud generation.
Nova: Generative Language Models for Assembly Code with Hierarchical Attention and Contrastive Learning
Nan Jiang (Purdue University), Petr Babkin (J.P. Morgan AI Research)
GenerationAI Code AssistantTransformerLarge Language ModelContrastive LearningText
🎯 What it does: A generative large language model Nova has been developed for x86-64 assembly code, addressing the challenges posed by low information density and optimization diversity in binary analysis.
NovelQA: Benchmarking Question Answering on Documents Exceeding 200K Tokens
Cunxiang Wang (Westlake University), Yue Zhang (Westlake University)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A novel question-answering benchmark, NovelQA, aimed at long texts (averaging over 200k tokens), has been constructed for evaluation using manually annotated real novel content and questions.
NoVo: Norm Voting off Hallucinations with Attention Heads in Large Language Models
Zheng Yi Ho, Dacheng Tao (Nanyang Technological University)
TransformerLarge Language ModelText
🎯 What it does: The Norm Voting (NoVo) method is proposed, which utilizes the L2 norm of attention heads for truth voting in zero-shot multiple-choice questions, significantly reducing hallucinations in large language models and improving factual accuracy.
NRGBoost: Energy-Based Generative Boosted Trees
João Bravo (Feedzai)
GenerationData SynthesisExplainability and InterpretabilityGenerative Adversarial NetworkTabular
🎯 What it does: This paper proposes an energy-based generative model called NRGBoost, which transforms gradient boosting trees into a generative model. It uses second-order approximate likelihood maximization during training and supports conditional inference for arbitrary variables and handling of missing values.
NUDGE: Lightweight Non-Parametric Fine-Tuning of Embeddings for Retrieval
Sepanta Zeighami (University of California Berkeley), Aditya Parameswaran (University of California Berkeley)
RetrievalOptimizationSupervised Fine-TuningImageText
🎯 What it does: A non-parametric embedding fine-tuning method called NUDGE is proposed, which enhances the accuracy of k-NN retrieval by directly modifying the data embedding vectors.
Null Counterfactual Factor Interactions for Goal-Conditioned Reinforcement Learning
Caleb Chuck (University of Texas at Austin), Scott Niekum (University of Texas at Austin)
Graph Neural NetworkReinforcement LearningGraph
🎯 What it does: Two new methods are proposed and implemented: Null Counterfactual Interaction Inference (NCII) for inferring causal interactions between objects, and Hindsight Relabeling using Interactions (HInt), which enhances sample efficiency in goal-conditioned reinforcement learning by filtering trajectories that only contain interactions.
Number Cookbook: Number Understanding of Language Models and How to Improve It
Haotong Yang (Peking University), Muhan Zhang (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: This paper constructs the NUPA Test benchmark, systematically evaluating LLMs on 41 basic numerical understanding and processing tasks under four types of numerical representations: integers, floating-point numbers, fractions, and scientific notation, and conducts zero-shot testing on various large models. It also explores the enhancement effects of tokenizers, positional encodings, numerical format modifications during the pre-training phase, post-training fine-tuning, and the Chain-of-Thought method on NUPA.
NutriBench: A Dataset for Evaluating Large Language Models in Nutrition Estimation from Meal Descriptions
Mehak Preet Dhaliwal (University of California), Yao Qin (University of California)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Developed the NUTRIBENCH benchmark set, which includes 11,857 manually verified natural language meal descriptions, and evaluated the performance of 12 LLMs on carbohydrate estimation tasks.
NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models
Chankyu Lee (NVIDIA), Wei Ping (NVIDIA)
ClassificationRetrievalTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Developed NV-Embed, a general embedding model based on a decoder-only LLM, which employs a latent attention layer for more expressive pooling, removes causal attention masks, and combines retrieval and non-retrieval tasks through two-stage contrastive learning instruction tuning;
NVS-Solver: Video Diffusion Model as Zero-Shot Novel View Synthesizer
Meng YOU, Junhui Hou (City University of Hong Kong)
GenerationData SynthesisDepth EstimationDiffusion modelVideo
🎯 What it does: Using a pre-trained large video diffusion model, a zero-shot novel view synthesis method is proposed, which can generate images from arbitrary viewpoints from a single view, a sequence of views, or a monocular video without additional training.
O(d/T) Convergence Theory for Diffusion Probabilistic Models under Minimal Assumptions
Gen Li (Chinese University of Hong Kong), Yuling Yan (University of Wisconsin-Madison)
Diffusion model
🎯 What it does: A faster convergence speed is proposed in the theoretical convergence study of the DDPM sampler.
OASIS Uncovers: High-Quality T2I Models, Same Old Stereotypes
Sepehr Dehdashtian (Michigan State University), Vishnu Boddeti (Michigan State University)
Image TranslationGenerationTransformerLarge Language ModelPrompt EngineeringImageText
🎯 What it does: The OASIS toolbox is proposed to quantify and explain stereotypes in text-to-image models (T2I). It first provides a sociologically defined measure of stereotypes, then offers two metrics, Stereotype Score and WALS, to evaluate the distribution and spectral differences of generated images. It also employs two methods, StOP and SPI, to explore the model's internal stereotypical attributes related to concepts and their occurrence during the generation process.
OATS: Outlier-Aware Pruning Through Sparse and Low Rank Decomposition
Stephen Zhang (University of Toronto), Vardan Papyan (University of Toronto)
CompressionAnomaly DetectionComputational EfficiencyTransformerLarge Language ModelImageText
🎯 What it does: A model compression method that does not require retraining is proposed—OATS, which decomposes the transformer weight matrix into the sum of a sparse matrix and a low-rank matrix, and utilizes the second-order moment of input embeddings for weight scaling, preserving key features of the model;
OBI-Bench: Can LMMs Aid in Study of Ancient Script on Oracle Bones?
Zijian Chen (Shanghai Jiao Tong University), Guangtao Zhai (Shanghai Jiao Tong University)
ClassificationRecognitionRetrievalTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: A comprehensive benchmark for the retrieval and interpretation of ancient script stone tablets, named OBI-Bench, has been proposed. This benchmark evaluates the performance of 23 large-scale multimodal models (LMM) on five major tasks: recognition, stitching, classification, retrieval, and interpretation.
Object-Centric Pretraining via Target Encoder Bootstrapping
Nikola Đukić (KU Leuven), Tinne Tuytelaars (KU Leuven)
Object DetectionSegmentationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: We propose OCEBO, a pre-training framework for training object-centered models from scratch based on a target encoder bootstrap;
ObscuraCoder: Powering Efficient Code LM Pre-Training Via Obfuscation Grounding
Indraneil Paul (Tu Darmstadt), Iryna Gurevych (Tu Darmstadt)
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: This paper proposes a pre-training method based on code obfuscation (ObscuraCoder), which significantly improves the performance of multilingual code models in tasks such as syntax and semantic understanding, library usage, code completion, and summarization by conducting translation learning between the source code and its obfuscated version.
OCCAM: Towards Cost-Efficient and Accuracy-Aware Classification Inference
Dujian Ding (University of British Columbia), Laks V. S. Lakshmanan (University of British Columbia)
ClassificationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A mixed reasoning framework OCCAM is proposed, which dynamically allocates different capacities of classifiers for different queries, thereby maximizing overall accuracy while satisfying user budget constraints.
Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy
Mingyang Zhao (Hong Kong Institute of Science and Innovation), Dong-ming Yan
OptimizationPoint CloudBiomedical Data
🎯 What it does: This paper proposes an unsupervised occlusion-adaptive non-rigid point cloud registration method called OAR, which achieves physically reasonable registration of occluded areas using the maximum mutual information criterion and local linear reconstruction.
OccProphet: Pushing the Efficiency Frontier of Camera-Only 4D Occupancy Forecasting with an Observer-Forecaster-Refiner Framework
Junliang Chen (Hong Kong Polytechnic University), Lap-Pui Chau (Hong Kong Polytechnic University)
Depth EstimationAutonomous DrivingComputational EfficiencyConvolutional Neural NetworkSimultaneous Localization and MappingImage
🎯 What it does: A camera-based 4D occupancy prediction framework named OccProphet is proposed, which efficiently predicts future 3D occupancy states using an Observer-Forecaster-Refiner three-step pipeline.
OCEAN: Offline Chain-of-thought Evaluation and Alignment in Large Language Models
Junda Wu (University of California San Diego), Julian McAuley (University of California San Diego)
OptimizationTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Proposes the OCEAN framework for offline evaluation and alignment of chain reasoning in large language models (LLMs);
ODE-based Smoothing Neural Network for Reinforcement Learning Tasks
Yinuo Wang (Tsinghua University), Shengbo Eben Li (Tsinghua University)
Reinforcement LearningTime SeriesOrdinary Differential Equation
🎯 What it does: A reinforcement learning (RL) policy network based on a smooth ODE neural network (SmODE) is proposed to address the issue of unsmooth control outputs in deep RL.
Offline Hierarchical Reinforcement Learning via Inverse Optimization
Carolin Schmidt (Technical University of Denmark), Filipe Rodrigues (Google DeepMind)
OptimizationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: The OHIO framework is proposed, which utilizes known low-level policy structures and approximate dynamics to convert state-only offline data into high-level action data usable for offline reinforcement learning through inverse optimization.
Offline Model-Based Optimization by Learning to Rank
Rong-Xi Tan (Nanjing University), Chao Qian (Nanjing University)
OptimizationReinforcement LearningTabular
🎯 What it does: A learning-to-rank based offline model optimization framework RaM is proposed, replacing the traditional MSE regression surrogate and directly optimizing the relative ranking of designs.
Offline RL in Regular Decision Processes: Sample Efficiency via Language Metrics
Ahana Deb (Universitat Pompeu Fabra), Mohammad Sadegh Talebi (University of Copenhagen)
Reinforcement Learning
🎯 What it does: Two improved offline RL algorithms are proposed, utilizing language metrics and Count-Min-Sketch to enhance the learning efficiency and sample complexity of Regular Decision Processes (RDP).
Offline RL with Smooth OOD Generalization in Convex Hull and its Neighborhood
Qingmao Yao (Beihang University), Xiao Zhang (Beihang University)
Robotic IntelligenceReinforcement LearningTabular
🎯 What it does: A smooth Q-function out-of-distribution (OOD) generalization method for offline reinforcement learning (RL) using Convex Hull Neighborhood (CHN) is proposed, which enhances the Q-value estimation in the OOD region through the Smooth Bellman Operator.
OGBench: Benchmarking Offline Goal-Conditioned RL
Seohong Park (University of California), Sergey Levine (University of California)
Robotic IntelligenceReinforcement LearningMultimodalityBenchmark
🎯 What it does: This paper presents OGBench—a unified benchmark for offline goal-conditioned reinforcement learning (GCRL), which includes 8 types of environments, 85 datasets, and 6 baseline algorithm implementations, aimed at systematically evaluating GCRL's performance on diverse tasks such as long-term reasoning, goal stitching, and randomness control.
OLMoE: Open Mixture-of-Experts Language Models
Niklas Muennighoff (Allen Institute for AI), Hannaneh Hajishirzi (Allen Institute for AI)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Proposed and open-sourced the sparse Mixture-of-Experts language model OLMOE-1B-7B and its instruction version for open-source research.
OMG: Opacity Matters in Material Modeling with Gaussian Splatting
Silong Yong (Carnegie Mellon University), Yaqi Xie (Carnegie Mellon University)
Gaussian SplattingPoint Cloud
🎯 What it does: A reverse rendering plugin module OMG based on 3D Gaussian Splatting is proposed, which couples material with opacity using the Beer-Lambert law.
OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Code
Maxence Faldor (Imperial College London), Jeff Clune (University of British Columbia)
GenerationRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Generate and learn continuously open, learnable, and interesting tasks by automatically generating task descriptions and environment code through large language models, building task profiles, and adaptively generating new tasks based on agent learning progress.
Omni-MATH: A Universal Olympiad Level Mathematic Benchmark for Large Language Models
Bofei Gao (Peking University), Baobao Chang (Peking University)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: We propose Omni-MATH, a new benchmark consisting of 4,428 Olympiad-level math problems across 33 subfields and 10+ difficulty levels, designed to assess the mathematical reasoning abilities of large language models.
OmniBind: Large-scale Omni Multimodal Representation via Binding Spaces
Zehan Wang (Zhejiang University), Zhou Zhao (Shanghai AI Lab)
RetrievalRepresentation LearningMixture of ExpertsContrastive LearningImageTextMultimodalityPoint CloudAudio
🎯 What it does: OmniBind is constructed, a unified multimodal representation model obtained by binding 5, 13, and 14 pre-trained multimodal spaces, achieving joint representation of 3D, audio, images, videos, and text.
OmniCorpus: A Unified Multimodal Corpus of 10 Billion-Level Images Interleaved with Text
Qingyun Li (Shanghai AI Laboratory), Jifeng Dai (Tsinghua University)
Data SynthesisRetrievalTransformerLarge Language ModelImageTextMultimodality
🎯 What it does: Constructed and released the OmniCorpus, an open multimodal document dataset with a scale of 10B (8.6B images, 1.696T text tokens, 2.2B documents), providing a unified streaming format and an efficient data engine.
OmniEdit: Building Image Editing Generalist Models Through Specialist Supervision
Cong Wei (University of Waterloo), Wenhu Chen (University of Waterloo)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelImageMultimodalityBenchmark
🎯 What it does: A general instruction-driven image editing model, OMNI-EDIT, has been constructed, capable of performing seven types of editing tasks at any aspect ratio and high resolution.
OmniKV: Dynamic Context Selection for Efficient Long-Context LLMs
Jitai Hao (Shandong University), Sheng Guo (MYbank, Ant Group)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes OmniKV, achieving a 1.7 times speedup in long text reasoning while maintaining the integrity of the KV cache.
OmniPhysGS: 3D Constitutive Gaussians for General Physics-Based Dynamics Generation
Yuchen Lin (Peking University), Yadong MU
GenerationData SynthesisOptimizationDiffusion modelScore-based ModelGaussian SplattingImageVideoPhysics Related
🎯 What it does: Proposes the OMNIPHYSGS framework, which uses learnable Constitutive Gaussians to achieve physics-based 3D dynamic scene synthesis, capable of automatically predicting the material properties of each particle and generating physically plausible animations consistent with text prompts.
OmniRe: Omni Urban Scene Reconstruction
Ziyu Chen (Shanghai Jiao Tong University), Yue Wang (University of Southern California)
RestorationSegmentationAutonomous DrivingGaussian SplattingPoint Cloud
🎯 What it does: This paper studies a high-fidelity digital twin reconstruction system for urban dynamic scenes called OmniRe, which can simultaneously reconstruct static backgrounds, vehicles, and non-rigid dynamic subjects (such as pedestrians and cyclists), and supports editable 3D Gaussian scene graphs.
OmniSep: Unified Omni-Modality Sound Separation with Query-Mixup
Xize Cheng (Zhejiang University), Zhou Zhao (Zhejiang University)
RetrievalPrompt EngineeringMultimodalityAudio
🎯 What it does: The OmniSep model is proposed, achieving audio separation based on arbitrary single-modal (text, image, audio) or multi-modal combination queries, and supports negative queries and open vocabulary queries.
OmnixR: Evaluating Omni-modality Language Models on Reasoning across Modalities
Lichang Chen (University of Maryland), Boqing Gong (Google DeepMind)
Data SynthesisPrompt EngineeringImageVideoTextMultimodalityBenchmarkChain-of-ThoughtAudio
🎯 What it does: Developed the Omni × R benchmark to evaluate the performance of multimodal language models in cross-modal reasoning, providing synthetic and real data subsets.
On a Connection Between Imitation Learning and RLHF
Teng Xiao (Pennsylvania State University), Vasant G Honavar (Pennsylvania State University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper re-examines RLHF from the perspective of imitation learning, proposing the DIL framework, which directly optimizes the reverse KL imitation learning objective and estimates the density ratio through Bregman divergence, thereby achieving efficient preference alignment.
On Bits and Bandits: Quantifying the Regret-Information Trade-off
Itai Shufaro (Technion), Shie Mannor (NVIDIA Research)
Large Language ModelReinforcement LearningText
🎯 What it does: This paper studies the trade-off relationship between information quantity (measured in bits) and cumulative loss (return) in Bayesian interactive decision-making problems, proposing a general information-theoretic method to obtain lower and upper bounds, and validating this theory in practical tasks.
On Calibration of LLM-based Guard Models for Reliable Content Moderation
Hongfu Liu (National University of Singapore), Ye Wang (National University of Singapore)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study investigates the confidence calibration and reliability of guardian models based on large language models (LLM) in content moderation, systematically evaluating the ECE and F1 of 9 open-source guardian models across 12 public benchmarks, and discusses post-hoc calibration methods.
On Conformal Isometry of Grid Cells: Learning Distance-Preserving Position Embedding
Dehong Xu (University of California, Los Angeles), Ying Nian Wu (University of California, Los Angeles)
Recurrent Neural NetworkSequential
🎯 What it does: This paper proposes and verifies that the reason for the formation of hexagonal lattice patterns in grid cell response maps is that their neural spatial mapping satisfies the assumption of local distance-preserving conformal isometry.
On Designing General and Expressive Quantum Graph Neural Networks with Applications to MILP Instance Representation
Xinyu Ye (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A general quantum graph neural network framework GQGLA is proposed to learn the graph representation of mixed-integer linear programming (MILP) instances, overcoming the indistinguishability of classical GNNs on certain MILP graphs;
On Discriminative Probabilistic Modeling for Self-Supervised Representation Learning
Bokun Wang (Texas A&M University), Tianbao Yang (Texas A&M University)
RetrievalRepresentation LearningContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a discriminative probability modeling framework for continuous domains, combining multiple importance sampling (MIS) to address the partition function integral challenge, and based on this, designs a new non-parametric approximation method and corresponding contrastive loss;
On Disentangled Training for Nonlinear Transform in Learned Image Compression
Han Li (Shanghai Jiao Tong University), Hongkai Xiong (Shanghai Jiao Tong University)
CompressionImage
🎯 What it does: An auxiliary linear transformation, AuxT, is proposed to help learn image compression models achieve energy compression through feature decorrelation and non-uniform energy modulation during the training phase, significantly accelerating training.
On Evaluating the Durability of Safeguards for Open-Weight LLMs
Xiangyu Qi (Princeton University), Peter Henderson (Princeton University)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical Data
🎯 What it does: Evaluate the durability of security measures for open-weight large language models (LLMs), focusing on two mainstream methods, RepNoise and TAR, systematically demonstrating various pitfalls and failure scenarios during the evaluation process.
On Generalization Across Environments In Multi-Objective Reinforcement Learning
Jayden Teoh (Singapore Management University), Peter Vamplew (Federation University Australia)
OptimizationReinforcement LearningBenchmark
🎯 What it does: This paper proposes a framework for environment generalization in Multi-Objective Reinforcement Learning (MORL) and establishes a new MORL generalization benchmark and evaluation metrics, subsequently evaluating various state-of-the-art (SOTA) algorithms on this benchmark.
On Large Language Model Continual Unlearning
Chongyang Gao (Northwestern University), Qi Zhu (Northwestern University)
Data-Centric LearningTransformerLarge Language ModelText
🎯 What it does: The O3 framework is proposed for continuous machine unlearning of large language models without using retained data.
On Linear Representations and Pretraining Data Frequency in Language Models
Jack Merullo (Brown University), Yanai Elazar (Allen Institute for AI)
TransformerLarge Language ModelText
🎯 What it does: Investigate the correlation between the linear representation of factual relationships in language models and word frequency in pre-training corpora, and construct a regression model to predict word frequency based on linear representation.
On Minimizing Adversarial Counterfactual Error in Adversarial Reinforcement Learning
Roman Belaire (Singapore Management University), Pradeep Varakantham (Singapore Management University)
Adversarial AttackReinforcement LearningSequential
🎯 What it does: This paper proposes a robust reinforcement learning method for the problem of partial observability caused by observation disturbances, defining the Adversarial Counterfactual Error (ACoE) and introducing a scalable C-ACoE objective. It then maximizes rewards while minimizing C-ACoE within the PPO/DQN framework to enhance adversarial robustness.
On Quantizing Neural Representation for Variable-Rate Video Coding
Junqi Shi (Nanjing University), Zhan Ma (Nanjing University)
CompressionVideo
🎯 What it does: Developed NeuroQuant, a post-training quantization method for variable bitrate implementation of non-general implicit neural representation video coding;
On Rollouts in Model-Based Reinforcement Learning
Bernd Frauenknecht (RWTH Aachen University), Sebastian Trimpe (RWTH Aachen University)
Reinforcement LearningTabular
🎯 What it does: This paper proposes the Infoprop mechanism, which reduces model error accumulation in model-based reinforcement learning by distinguishing between naive and aleatoric uncertainty, achieving longer and more reliable simulation rollouts.
On Scaling Up 3D Gaussian Splatting Training
Hexu Zhao (New York University), Saining Xie
GenerationOptimizationComputational EfficiencyGaussian SplattingImage
🎯 What it does: The Grendel system is proposed, achieving distributed training of 3D Gaussian Splatting on multiple GPUs, supporting large batch views, dynamic load balancing, and sparse All-to-All communication.
On Speeding Up Language Model Evaluation
Jin Peng Zhou (Cornell University), Kilian Q Weinberger
Large Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes two adaptive evaluation methods based on multi-armed bandits to quickly identify the optimal LLM or prompt under a limited budget.
On Statistical Rates of Conditional Diffusion Transformers: Approximation, Estimation and Minimax Optimality
Jerry Yao-Chieh Hu (Northwestern University), Han Liu (Northwestern University)
TransformerDiffusion model
🎯 What it does: Analyzed and derived the approximation rate and estimation rate of the classifier-free diffusion Transformer (DiT) under given conditions, exploring the statistical limits under generic and strong Hölder smoothness assumptions;
On Stochastic Contextual Bandits with Knapsacks in Small Budget Regime
Hengquan Guo (ShanghaiTech University), Xin Liu (ShanghaiTech University)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the random context bandit with budget constraints (CBwK) problem under small budget scenarios and proposes a single-stage adaptive primal-dual algorithm AUPD to achieve optimal or near-optimal regret without requiring prior safety margins or strict feasibility assumptions.
On Targeted Manipulation and Deception when Optimizing LLMs for User Feedback
Marcus Williams (University of California Berkeley), Anca Dragan (University of California Berkeley)
OptimizationSafty and PrivacyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The study investigates how reinforcement learning optimizes LLMs to obtain user feedback, revealing that the model learns targeted manipulation and deceptive behaviors aimed at easily deceived users, while assessing the risks and effectiveness of existing mitigation methods.
On the Adversarial Risk of Test Time Adaptation: An Investigation into Realistic Test-Time Data Poisoning
Yongyi Su (South China University of Technology), Xun Xu (Institute for Infocomm Research)
Domain AdaptationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a testing-time data poisoning (RTTDP) framework that is more aligned with practical deployment scenarios. It designs a complete protocol for gray-box attacks, without the need to access other normal samples, online attack sequences, and attack budget constraints. Based on this, it develops adaptive 'in-distribution' poisoning methods, feature consistency regularization, and two attack targets for testing-time adaptation (TTA) methods (high-entropy attack and low-entropy attack).
On the Adversarial Vulnerability of Label-Free Test-Time Adaptation
Shahriar Rifat (Northeastern University), Francesco Restuccia (Northeastern University)
ClassificationDomain AdaptationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This study quantifies the vulnerability of test-time adaptive (TTA) methods to adversarial attacks in the absence of labels and proposes a novel attack algorithm that does not rely on true labels (Feature Collapse Attack, FCA).
On the Almost Sure Convergence of the Stochastic Three Points Algorithm
Taha EL BAKKALI EL KADI (Mohammed VI Polytechnic University), Omar Saadi
OptimizationTabular
🎯 What it does: Proposed and theoretically analyzed the almost sure convergence and convergence rate of the Stochastic Three-Point (STP) algorithm on three types of functions: smooth, convex, and smooth strongly convex.
On the Benefits of Attribute-Driven Graph Domain Adaptation
Ruiyi Fang (Western University), Charles Ling (Western University)
ClassificationDomain AdaptationGraph Neural NetworkGraph
🎯 What it does: This paper proposes an attribute-driven dual-view alignment framework GAA in graph domain adaptation (GDA) to address cross-network node classification tasks.
On the Benefits of Memory for Modeling Time-Dependent PDEs
Ricardo Buitrago, Andrej Risteski (Carnegie Mellon University)
Time SeriesPhysics Related
🎯 What it does: This paper proposes and validates a memory network structure (MemNO) for solving time-varying partial differential equations, demonstrating through theory and experiments that memory can significantly enhance prediction accuracy under low-resolution or noisy observations.
On the Byzantine-Resilience of Distillation-Based Federated Learning
Christophe Roux (Zuse Institute Berlin), Sebastian Pokutta (Technische Universität Berlin)
Federated LearningKnowledge DistillationAdversarial AttackConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This study investigates the robustness of knowledge distillation-based federated learning (FedDistill) under Byzantine attacks, proposing two novel attacks (LMA, CPA) and designing an experience-weighted defense method called ExpGuard and an attack concealment framework named HIPS.
On the Completeness of Invariant Geometric Deep Learning Models
Zian Li (Peking University), Muhan Zhang (Peking University)
Graph Neural NetworkPoint Cloud
🎯 What it does: A systematic analysis of the theoretical expressiveness of several classes of point cloud-based invariant geometric deep learning models (such as DisGNN, GeoNGNN, DimeNet, GemNet, SphereNet, etc.) is conducted, proving the near completeness of DisGNN and proposing a stronger GeoNGNN to achieve E(3)-completeness, while also providing completeness proofs for almost all invariant models under the condition of complete connectivity.
On the Convergence of No-Regret Dynamics in Information Retrieval Games with Proportional Ranking Functions
Omer Madmon (Technion Israel Institute of Technology), Moshe Tennenholtz (Technion Israel Institute of Technology)
RetrievalOptimizationTabular
🎯 What it does: This study investigates the no-regret learning dynamics of publishers using the proportional ranking function (PRF) in information retrieval games, providing theoretical results that equate the concavity of the activation function with the convexity/social convexity of the game, proving that a concave activation function ensures the convergence of no-regret dynamics and the existence of a unique Nash equilibrium.
On the Crucial Role of Initialization for Matrix Factorization
Bingcong Li (ETH Zurich), Niao He (ETH Zurich)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelImageText
🎯 What it does: This paper proposes an initialization method based on Nystrom sampling and applies it to the fine-tuning of ScaledGD and LoRA, significantly improving the convergence speed and final performance of low-rank matrix decomposition and LLM/diffusion models.
On the Expressive Power of Sparse Geometric MPNNs
Yonatan Sverdlov (Technion Israel Institute of Technology), Nadav Dym (Technion Israel Institute of Technology)
Graph Neural NetworkGraph
🎯 What it does: This paper studies the expressive power of message passing networks on sparse geometric graphs, proving that E-GGNN possesses generic separability on connected graphs, and proposes the EGENNET architecture that can achieve this property.
On the expressiveness and spectral bias of KANs
Yixuan Wang (California Institute of Technology), Thomas Y. Hou (California Institute of Technology)
Time Series
🎯 What it does: This paper compares the expression and approximation capabilities of Kolmogorov-Arnold Networks (KAN) and Multi-Layer Perceptrons (MLP) through theoretical analysis and experimental validation, and studies their spectral bias differences, finding that KAN performs better in high-frequency learning.
On the Expressiveness of Rational ReLU Neural Networks With Bounded Depth
Gennadiy Averkov (Brandenburg University of Technology Cottbus-Senftenberg), Maximilian Merkert (Technische Universität Braunschweig)
🎯 What it does: This paper studies the expressive power of rational ReLU neural networks with bounded depth, particularly exploring the representation capability of the function F_n = max(0, x_1, ..., x_n).
On the Feature Learning in Diffusion Models
Andi Han (RIKEN AIP), Difan Zou (University of Hong Kong)
ClassificationData SynthesisRepresentation LearningConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: A theoretical framework is proposed and validated to compare the differences in feature learning processes between diffusion models and traditional classification models, focusing on the dynamics of learning signal and noise features.
On the Fourier analysis in the SO(3) space : the EquiLoPO Network
Dmitrii Zhemchuzhnikov (AIRI), Sergei Grudinin (Univ. Grenoble Alpes)
RecognitionSegmentationConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A trainable filter and local activation function convolutional network based on the continuous rotation group SO(3) is proposed, capable of achieving rotation equivariance for 3D volumetric data.
On the Hölder Stability of Multiset and Graph Neural Networks
Yair Davidson (Technion), Nadav Dym (Technion)
ClassificationOptimizationGraph Neural NetworkGraph
🎯 What it does: A new framework based on Hölder stability expectations is proposed to evaluate the separation quality of multi-set and graph neural networks, and SortMPNN, a sorting-based aggregation MPNN, is designed.
On the Identification of Temporal Causal Representation with Instantaneous Dependence
Zijian Li (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
Data SynthesisAnomaly DetectionRepresentation LearningTime SeriesSequential
🎯 What it does: A framework IDOL is proposed to identify time series latent variables with instantaneous causal relationships;
On the Importance of Language-driven Representation Learning for Heterogeneous Federated Learning
Yunlu Yan (Hong Kong University of Science and Technology), Lei Zhu (Hong Kong University of Science and Technology)
Federated LearningRepresentation LearningContrastive LearningImageText
🎯 What it does: Proposes the FedGLCL framework, which employs language-driven contrastive learning in federated learning, aligning global text embeddings with local image features, replacing traditional label-driven training to address the performance degradation caused by non-IID data.
On the Learn-to-Optimize Capabilities of Transformers in In-Context Sparse Recovery
Renpu Liu (University of Virginia), Jing Yang (University of Virginia)
OptimizationTransformerTabular
🎯 What it does: The paper proves that the Transformer can implement a Learning to Optimize (L2O) algorithm (LISTA-VM) and achieve linear convergence in context-sparse recovery tasks.
On the Linear Speedup of Personalized Federated Reinforcement Learning with Shared Representations
GUOJUN XIONG, Jian Li (Stony Brook University)
Federated LearningReinforcement Learning
🎯 What it does: The PFEDRL-REP framework is proposed, which can learn shared feature representations and personalized weights in multi-agent heterogeneous environments, and two instances are provided: PFEDTD-REP (linear TD) and PFEDDQN-REP (deep Q-network).
On the Modeling Capabilities of Large Language Models for Sequential Decision Making
Martin Klissarov (Mila McGill University), Bogdan Mazoure (Apple)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelMultimodalitySequentialChain-of-Thought
🎯 What it does: This study investigates the decision-making capabilities of large language models (LLM) in reinforcement learning (RL), exploring two strategies: direct action generation and indirect reward model construction.
On the Optimal Memorization Capacity of Transformers
Tokio Kajitsuka (University of Tokyo), Issei Sato (University of Tokyo)
TransformerSequential
🎯 What it does: This paper studies the parameter efficiency of Transformers in memorization tasks, providing upper and lower bounds in both next-step prediction and sequence-to-sequence prediction settings, and proving that these upper bounds are optimal in terms of the number of parameters (differing only by a logarithmic factor).
On the Optimization and Generalization of Two-layer Transformers with Sign Gradient Descent
Bingrui Li (Tsinghua University), Jianfei Chen (Tsinghua University)
OptimizationTransformerTabular
🎯 What it does: Analyzed the training dynamics of Sign Gradient Descent (SignGD) on a two-layer Transformer, proving that it converges quickly on noisy linearly separable binary classification datasets but has poor generalization performance.
On the Optimization Landscape of Low Rank Adaptation Methods for Large Language Models
Xu-Hui Liu (Nanjing University), Yang Yu (Nanjing University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper analyzes the optimization landscape of low-rank adaptation methods through theoretical analysis and experimental evaluation, and proposes a new algorithm called GaRare.
On the Performance Analysis of Momentum Method: A Frequency Domain Perspective
Xianliang Li (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences), Sheng Xu (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences)
OptimizationReinforcement LearningImageText
🎯 What it does: By viewing the momentum method as a time-varying filter, a frequency domain analysis framework is proposed to explore the impact of different momentum coefficients on high and low frequency gradients, and based on this, an optimizer FSGDM is designed to dynamically adjust the momentum filtering characteristics, verifying its performance improvement in multiple tasks.
On the Price of Differential Privacy for Hierarchical Clustering
Chengyuan Deng (Rutgers University), Samson Zhou (Texas A&M University)
OptimizationSafty and PrivacyGraphTabular
🎯 What it does: A polynomial-time hierarchical clustering algorithm is designed under the weight-level differential privacy model, achieving a multiplicative approximation of O(log^(5/2) n / ε) by adding noise to edge weights and recursively solving for a balanced sparse cut.
On the Relation between Trainability and Dequantization of Variational Quantum Learning Models
Elies Gil-Fuster (Freie Universitat Berlin), Vedran Dunjko (Universiteit Leiden)
Physics Related
🎯 What it does: This paper explores the relationship between the trainability of variational quantum learning models and dequantization, proposing relevant definitions and a theoretical framework, and demonstrating how to construct variational quantum models that are both trainable and non-dequantizable.
On the Role of Attention Heads in Large Language Model Safety
Zhenhong Zhou (Tongyi Lab), Yongbin Li (Tongyi Lab)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: This paper quantifies and attributes the attention heads responsible for safety in large language models by defining the Attention Head Importance Score (Ships) and the Safety Attention Head Attribution Algorithm (Sahara);
On the self-verification limitations of large language models on reasoning and planning tasks
Kaya Stechly (Arizona State University), Subbarao Kambhampati (Arizona State University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Systematically evaluate the effect of self-criticism in LLMs on reasoning and planning tasks, and conduct ablation experiments by splitting into three layers: validation, critical generation, and critical consideration.
On the Transfer of Object-Centric Representation Learning
Aniket Rajiv Didolkar (MILA & University of Montreal), Maximilian Seitzer (Max Planck Institute for Intelligent Systems & University of Tübingen)
Object DetectionRepresentation LearningTransformerContrastive LearningImageBenchmark
🎯 What it does: This paper studies the transferability of center representation learning, proposes a zero-shot evaluation benchmark, and evaluates the performance of different models on various real and synthetic datasets. It also introduces the FT-DINOSAUR method for task-specific fine-tuning of pre-trained features to enhance performance.
On-the-fly Preference Alignment via Principle-Guided Decoding
Mingye Zhu (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
Recommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A framework named OPAD is proposed for preference alignment during inference without fine-tuning. It dynamically adjusts token probabilities through a principle-guided reward mechanism, enabling the model to adhere to user preferences and principles during the inference phase.
Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaptation
Anqi Li (Beijing Jiaotong University), Huihui Bai (Beijing Jiaotong University)
GenerationCompressionGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes Control-GIC, a unified controllable generative image compression framework that enables fine-grained bitrate adjustment and high-fidelity reconstruction within a single model.
One for all and all for one: Efficient computation of partial Wasserstein distances on the line
Laetitia Chapel (Institut Agro Rennes Angers), Romain Tavenard (Universite de Rennes)
Domain AdaptationOptimizationComputational EfficiencyImagePoint Cloud
🎯 What it does: The PAWL algorithm has been designed and implemented for efficiently computing the partial Wasserstein distance in one-dimensional space, and a slicing strategy for Partial OT has been proposed, which can obtain the solution for all transport quantities at once.
One Hundred Neural Networks and Brains Watching Videos: Lessons from Alignment
Christina Sartzetaki (University of Amsterdam), Iris Groen (University of Amsterdam)
TransformerVideoMagnetic Resonance ImagingBenchmark
🎯 What it does: A large-scale benchmark test was conducted on the similarity of representations between 99 video models and the human brain while watching videos, using Representational Similarity Analysis (RSA) for hierarchical comparisons across multiple brain regions.
One Model Transfer to All: On Robust Jailbreak Prompts Generation against LLMs
Linbao Li (Harbin Institute of Technology), YU LI
GenerationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: The ArrAttack framework is proposed, which utilizes rewriting attacks and a general robustness evaluation model to automatically generate jailbreak prompts that can bypass various defense strategies, and quickly produces high-quality, semantically consistent robust prompts through fine-tuning of the generative model.
One Step Diffusion via Shortcut Models
Kevin Frans (University of California Berkeley), Pieter Abbeel (University of California Berkeley)
GenerationData SynthesisRobotic IntelligenceTransformerDiffusion modelImage
🎯 What it does: A single network, single-stage training Shortcut Models is proposed, capable of generating high-quality images at any number of steps (including single-step).
One-for-All Few-Shot Anomaly Detection via Instance-Induced Prompt Learning
Wenxi Lv (Sun Yat-sen University), Wenchao Xu (Hong Kong Polytechnic University)
Anomaly DetectionTransformerPrompt EngineeringVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes a few-shot anomaly detection framework based on an instance-induced prompt generator and a category-aware memory bank, capable of uniformly handling multi-class anomaly detection tasks under the 'one-to-all' paradigm.
One-Prompt-One-Story: Free-Lunch Consistent Text-to-Image Generation Using a Single Prompt
Tao Liu (Nankai University), Ming-Ming Cheng (Nankai University)
GenerationData SynthesisPrompt EngineeringDiffusion modelImageText
🎯 What it does: This paper proposes a training-free consistency text-to-image generation method called One-Prompt-One-Story. By concatenating identity prompts with multiple frame descriptions into a single long prompt, and applying singular value reweighting and identity-preserving cross-attention on the prompt embeddings during generation, it can maintain the identity consistency of the same subject and text alignment across different scenes.
Online Clustering with Nearly Optimal Consistency
T-H. Hubert Chan (University of Hong Kong), Mengshi Zhao (University of Hong Kong)
OptimizationTabular
🎯 What it does: An online k-MEANS (and more generally, (k,z)-CLUSTERING) algorithm is proposed, achieving almost optimal consistency (recourse) while maintaining a competitive ratio of (1+ϵα)^2, and converting any α-approximate offline algorithm into an online version.
ONLINE EPSILON NET & PIERCING SET FOR GEOMETRIC CONCEPTS
Sujoy Bhore (Indian Institute of Technology Bombay), Satyam Singh (Indian Institute of Technology Bombay)
🎯 What it does: This paper studies the online ε-net and online piercing set problems for geometric concepts, providing online algorithms for intervals, axis-aligned rectangles/boxes, ellipses, and similarly sized α-fat objects, along with their competitive ratios.
Online Preference Alignment for Language Models via Count-based Exploration
Chenjia Bai (China Telecom), Xuelong Li (Tencent AI Lab)
Recommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The COPO algorithm is proposed, which incorporates count-based exploration based on UCB into online RLHF, dynamically generating more diverse prompt-response pairs to enhance the model's preference alignment effect.