ICML 2025 Papers — Page 24
International Conference on Machine Learning · 3257 papers
Prompt-based Depth Pruning of Large Language Models
Juyun Wee (POSTECH), Jaeho Lee (POSTECH)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a prompt-based deep pruning method (PuDDing) that dynamically determines which Transformer blocks to retain for each input prompt, thereby reducing inference costs while maintaining task performance.
Prompt-to-Leaderboard: Prompt-Adaptive LLM Evaluations
Evan Frick (University of California), Ion Stoica (University of California)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes the P2L method, which trains a large language model to directly output Bradley-Terry coefficients after receiving natural language prompts, generating a model ranking for each prompt to achieve fine-grained evaluation.
ProofAug: Efficient Neural Theorem Proving via Fine-grained Proof Structure Analysis
Haoxiong Liu (Tsinghua University), Andrew C Yao (Shanghai Qi Zhi Institute)
Large Language ModelPrompt Engineering
🎯 What it does: A fine-grained proof structure analysis method named ProofAug is proposed, which utilizes complete proofs generated by LLM and extracts repairable proof structures through Maximum Compatible Semi-Proofs (MCSP), combining ATP and built-in proof methods to achieve efficient neural theorem proving.
Propagate and Inject: Revisiting Propagation-Based Feature Imputation for Graphs with Partially Observed Features
Daeho Um (Samsung Electronics), Seulki Park (University of Michigan)
Graph Neural NetworkGraph
🎯 What it does: This study investigates the problem of propagative imputation of missing features in graph data and proposes a new method, FISF, which addresses performance degradation caused by low-variance channels by injecting synthetic features.
Propagation of Chaos for Mean-Field Langevin Dynamics and its Application to Model Ensemble
Atsushi Nitanda (Agency for Science Technology and Research), Taiji Suzuki (University of Tokyo)
TextStochastic Differential Equation
🎯 What it does: An in-depth analysis of the propagation catastrophe (PoC) in mean-field Langevin dynamics (MFLD) is conducted, and based on this, improved PoC results and a PoC-based model ensemble method are proposed.
Proposer-Agent-Evaluator (PAE): Autonomous Skill Discovery For Foundation Model Internet Agents
Yifei Zhou (University of California), Li Erran Li
Autonomous DrivingTransformerLarge Language ModelReinforcement LearningVision Language ModelTextChain-of-Thought
🎯 What it does: This paper proposes the Proposer-Agent-Evaluator (PAE) framework, allowing the foundational model (VLM) network agent to autonomously generate tasks, execute them, and receive rewards from its own evaluator without human guidance, thereby discovering and learning new web browsing skills independently.
ProSec: Fortifying Code LLMs with Proactive Security Alignment
Xiangzhe Xu (Purdue University), Xiangyu Zhang (Purdue University)
Safty and PrivacyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes PROSEC, which achieves safety alignment in the post-training phase of code LLMs by utilizing vulnerability-triggering instructions and preference learning.
Protein Structure Tokenization: Benchmarking and New Recipe
Xinyu Yuan (Mila Quebec AI Institute), Huzefa Rangwala (Amazon)
Protein Structure PredictionAuto EncoderBiomedical DataBenchmark
🎯 What it does: This paper proposes a unified protein structure tokenization evaluation framework, StructTokenBench, and conducts a four-dimensional evaluation of existing tokenization methods based on this framework.
Proto Successor Measure: Representing the Behavior Space of an RL Agent
Siddhant Agarwal (University of Texas at Austin), Amy Zhang (University of Texas at Austin)
OptimizationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper proposes the Proto Successor Measure (PSM), a representation method for action spaces based on MDP that can pre-train representations capable of achieving zero-shot optimal policies under any reward function using only reward-free interaction data.
PROTOCOL: Partial Optimal Transport-enhanced Contrastive Learning for Imbalanced Multi-view Clustering
Xuqian Xue (Fudan University), Junping Zhang (Fudan University)
Representation LearningTransformerContrastive LearningImage
🎯 What it does: This study investigates the challenges of class imbalance in multi-view clustering and proposes the PROTOCOL framework to achieve imbalance awareness and representation enhancement.
Protriever: End-to-End Differentiable Protein Homology Search for Fitness Prediction
Ruben Weitzman (University of Oxford), Pascal Notin (Harvard Medical School)
RetrievalDrug DiscoveryTransformerBiomedical Data
🎯 What it does: Protriever is an end-to-end differentiable protein homology retrieval framework that combines a retriever and a reader model to predict protein fitness.
Provable and Practical Online Learning Rate Adaptation with Hypergradient Descent
Ya-Chi Chu (Stanford University), Madeleine Udell (Stanford University)
OptimizationTabular
🎯 What it does: A rigorous theoretical analysis of the convergence of Hypergradient Descent (HDM) is conducted, and an improved variant that can achieve local superlinear convergence is proposed.
Provable Benefit of Random Permutations over Uniform Sampling in Stochastic Coordinate Descent
Donghwa Kim (KAIST), Chulhee Yun (KAIST)
Optimization
🎯 What it does: This paper compares the convergence rates of Random Coordinate Descent (RCD) and Random Permutation Coordinate Descent (RPCD) on positive definite quadratic functions, and proves that RPCD converges faster in function classes that include permutation structures.
Provable Benefits of Unsupervised Pre-training and Transfer Learning via Single-Index Models
Taj Jones-McCormick (University of Waterloo), Subhabrata Sen (Harvard University)
🎯 What it does: This paper studies the impact of unsupervised pre-training and transfer learning on the sample complexity of high-dimensional supervised learning, particularly in the case of limited labeled data, and explores how to train a single-layer neural network using online stochastic gradient descent.
Provable Efficiency of Guidance in Diffusion Models for General Data Distribution
Gen Li (Chinese University of Hong Kong), Yuchen Jiao (Chinese University of Hong Kong)
GenerationData SynthesisDiffusion modelScore-based ModelImageStochastic Differential Equation
🎯 What it does: This paper analyzes the effectiveness of guidance techniques in diffusion models under general data distributions, proving that guidance can improve sample quality, particularly by reducing the proportion of low-quality samples.
Provable In-Context Vector Arithmetic via Retrieving Task Concepts
Dake Bu (City University of Hong Kong), Taiji Suzuki (University of Tokyo)
Transformer
🎯 What it does: This paper proposes a nonlinear Transformer optimization theory based on layer normalization, residual connections, and softmax attention, demonstrating that the model can achieve vector arithmetic in factual recall tasks after gradient descent training on question-answering (QA) data.
Provable Length Generalization in Sequence Prediction via Spectral Filtering
Annie Marsden (Google DeepMind), Elad Hazan (Princeton University)
Recurrent Neural NetworkTime SeriesSequential
🎯 What it does: This study investigates the length generalization problem in sequence prediction, proposes the Asymmetric-Regret metric, and proves that the spectral filtering algorithm achieves length generalization on Linear Dynamical Systems (LDS).
Provable Maximum Entropy Manifold Exploration via Diffusion Models
Riccardo De Santi (ETH Zurich), Andreas Krause (ETH Zurich)
GenerationReinforcement LearningDiffusion modelImage
🎯 What it does: By maximizing entropy on the data manifold implicitly learned by the pre-trained diffusion model, a continuous-time reinforcement learning framework S-MEME based on mirror descent is proposed, achieving maximum entropy exploration without explicit density estimation.
Provable Policy Gradient for Robust Average-Reward MDPs Beyond Rectangularity
Qiuhao Wang (Southwestern University of Finance and Economics), Marek Petrik (University of New Hampshire)
OptimizationReinforcement LearningTabular
🎯 What it does: A globally convergent policy gradient method RP2G for robust Markov decision processes (RAMDP) with steady-state rewards is proposed, along with two gradient optimization algorithms for worst-case transition evaluation.
Provable Zero-Shot Generalization in Offline Reinforcement Learning
Zhiyong Wang (Chinese University of Hong Kong), Dongruo Zhou (Indiana University Bloomington)
Reinforcement Learning
🎯 What it does: This paper addresses the Zero-Shot Generalization (ZSG) problem in offline reinforcement learning and proposes two provable algorithms: Pessimistic Experience Risk Minimization (PERM) based on models and Pessimistic Proximal Policy Optimization (PPPO) based on policies, along with corresponding theoretical convergence and generalization error upper bounds.
Provably Cost-Sensitive Adversarial Defense via Randomized Smoothing
Yuan Xin (CISPA Helmholtz Center for Information Security), Xiao Zhang (CISPA Helmholtz Center for Information Security)
OptimizationAdversarial AttackConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: A provably cost-sensitive adversarial robustness defense framework based on randomized smoothing is proposed, providing a cost-sensitive certified radius and training method.
Provably Efficient Algorithm for Best Scoring Rule Identification in Online Principal-Agent Information Acquisition
Zichen Wang (University of Illinois Urbana-Champaign), Huazheng Wang (Oregon State University)
OptimizationReinforcement Learning
🎯 What it does: The problem of Best Scoring Rule Identification in the context of online information acquisition is proposed, and two algorithms OIAFC (Fixed Confidence) and OIAFB (Fixed Budget) are designed.
Provably Efficient Exploration in Inverse Constrained Reinforcement Learning
Bo Yue (Chinese University of Hong Kong), Guiliang Liu (Chinese University of Hong Kong)
Reinforcement LearningSequential
🎯 What it does: This paper studies the sampling efficiency of Inverse Constrained Reinforcement Learning (ICRL) and proposes two exploration algorithms based on error aggregation and candidate policy constraints, BEAR and PCSE, aiming to achieve provable sample complexity.
Provably Efficient RL for Linear MDPs under Instantaneous Safety Constraints in Non-Convex Feature Spaces
Amirhossein Roknilamouki (Ohio State University), Ness Shroff
Autonomous DrivingSafty and PrivacyReinforcement LearningSequential
🎯 What it does: This paper studies safe reinforcement learning with immediate hard constraints under linear MDPs, providing sublinear regret upper bounds in both star-convex and non-star-convex feature spaces, and proposes a two-stage NCS-LSVI algorithm to achieve zero safety violations.
Provably Improving Generalization of Few-shot models with Synthetic Data
Lan-Cuong Nguyen (Hanoi University of Science and Technology), Khoat Than (Hanoi University of Science and Technology)
ClassificationData SynthesisDiffusion modelImage
🎯 What it does: Develop a theoretical framework and propose an algorithm based on prototype learning, utilizing synthetic data to enhance the generalization of few-shot classification.
Provably Near-Optimal Federated Ensemble Distillation with Negligible Overhead
Won-Jun Jang (Korea Advanced Institute of Science and Technology), Si-Hyeon Lee (Korea Advanced Institute of Science and Technology)
Federated LearningKnowledge DistillationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: Proposes the FedGO algorithm, which uses a GAN discriminator to achieve approximately optimal federated ensemble distillation, addressing the client heterogeneity issue.
PROXSPARSE: REGULARIZED LEARNING OF SEMI-STRUCTURED SPARSITY MASKS FOR PRETRAINED LLMS
Hongyi Liu (Rice University), George Karypis (Amazon Web Service)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Achieved automatic generation of semi-structured (2:4) sparse masks on pre-trained LLMs through the ProxSparse regularization learning framework, enabling efficient inference without updating weights.
Proxy-FDA: Proxy-based Feature Distribution Alignment for Fine-tuning Vision Foundation Models without Forgetting
Chen Huang (Apple), Joshua M. Susskind (Apple)
ClassificationObject DetectionDomain AdaptationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a new regularization method called Proxy-FDA, which aims to structurally align feature distributions during the fine-tuning of visual foundation models, significantly reducing the phenomenon of concept forgetting.
Prune 'n Predict: Optimizing LLM Decision-making with Conformal Prediction
Harit Vishwakarma (JPMorganChase AI Research), Sumitra Ganesh
OptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The CROQ method is proposed to improve the accuracy of LLMs by pruning multiple-choice question answers through Conformal Prediction (CP) and re-prompting the LLM on the pruned questions.
Pruning for GNNs: Lower Complexity with Comparable Expressiveness
Dun Ma (University of Chinese Academy of Sciences), Shengminjie Chen (Institute of Computing Technology, Chinese Academy of Sciences)
Graph Neural NetworkGraph
🎯 What it does: Proposes pruned versions of three GNN frameworks: MP-GNN, K-Path, and K-Hop, removing redundant structures to reduce computational complexity.
PTTA: Purifying Malicious Samples for Test-Time Model Adaptation
Jing Ma (Huazhong University of Science and Technology), Xiang Xiang (Huazhong University of Science and Technology)
SegmentationDomain AdaptationAnomaly DetectionImage
🎯 What it does: This paper proposes a plugin method called PTTA, which transforms originally filtered malicious samples into beneficial samples during the Test-Time Adaptation (TTA) process, thereby improving the model's online inference performance.
Putnam-AXIOM: A Functional & Static Benchmark for Measuring Higher Level Mathematical Reasoning in LLMs
Aryan Gulati (Stanford University), Sanmi Koyejo (Stanford University)
Large Language ModelTextBenchmark
🎯 What it does: A set of 522 problems based on the William Lowell Putnam Mathematical Competition (Putnam‑AXIOM) and its 100 programmatic variants has been proposed, along with the introduction of a lightweight Teacher Forcing Accuracy (TFA) evaluation metric.
Puzzle: Distillation-Based NAS for Inference-Optimized LLMs
Akhiad Bercovich, Ran El-Yaniv
OptimizationComputational EfficiencyKnowledge DistillationNeural Architecture SearchTransformerLarge Language ModelText
🎯 What it does: This paper studies a distillation-based NAS framework called Puzzle, which can generate sub-models that efficiently infer on a single NVIDIA H100 GPU while maintaining nearly the same performance as the original large model (e.g., Llama-70B).
PyTDC: A multimodal machine learning training, evaluation, and inference platform for biomedical foundation models
Alejandro Velez-Arce (Harvard University), Marinka Zitnik (Harvard University)
Drug DiscoveryGraph Neural NetworkTransformerMultimodalityBiomedical DataBenchmark
🎯 What it does: A PyTDC platform has been constructed, providing multimodal single-cell data retrieval, training, evaluation, and inference, and for the first time offering benchmarks and tools for single-cell drug-target naming tasks;
Q-Supervised Contrastive Representation: A State Decoupling Framework for Safe Offline Reinforcement Learning
Zhihe Yang (Chinese University of Hong Kong), Yang Zhang (Hong Polytechnic University)
Representation LearningReinforcement LearningDiffusion modelContrastive LearningBenchmark
🎯 What it does: The SDQC framework is proposed, which decomposes global observations into two representations: reward-related and cost-related. It employs Q-supervised contrastive learning to learn these two representations, thereby achieving decision-making and safety assessment in safe offline reinforcement learning.
Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers
Weilun Feng (Institute of Computing Technology, Chinese Academy of Sciences), Michele Magno (ETH Zurich)
GenerationKnowledge DistillationTransformerDiffusion modelVideo
🎯 What it does: This paper proposes a quantization framework Q-VDiT specifically for video generation diffusion Transformers (DiT), addressing the issues of quantization information loss and spatiotemporal consistency loss.
QEM-Bench: Benchmarking Learning-based Quantum Error Mitigation and QEMFormer as a Multi-ranged Context Learning Baseline
Tianyi Bao (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Graph Neural NetworkTransformerGraphBenchmarkPhysics Related
🎯 What it does: A unified quantum error mitigation benchmark QEM-Bench and a new baseline method QEMFormer are proposed for evaluating machine learning-driven quantum error mitigation techniques.
QLASS: Boosting Language Agent Inference via Q-Guided Stepwise Search
Zongyu Lin (University of California), Kai-Wei Chang (University of California)
Robotic IntelligenceReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: The QLASS method is proposed, which utilizes the Q-value at each step to guide the reasoning and actions of the language agent, achieving more efficient reasoning search.
QMamba: On First Exploration of Vision Mamba for Image Quality Assessment
Fengbin Guan (University of Science and Technology of China), Zhibo Chen (University of Science and Technology of China)
Domain AdaptationTransformerPrompt EngineeringImage
🎯 What it does: A framework for image quality assessment based on Mamba, QMamba, is proposed, modeling task-specific, general, and transferable IQA.
QoS-Efficient Serving of Multiple Mixture-of-Expert LLMs Using Partial Runtime Reconfiguration
HamidReza Imani (George Washington University), Tarek El-Ghazawi (George Washington University)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: A MoE LLM system has been built to efficiently serve multiple models on a single GPU, utilizing expert similarity to merge shared memory and dynamically replace non-expert layers at runtime to reduce memory usage.
QPRL : Learning Optimal Policies with Quasi-Potential Functions for Asymmetric Traversal
Jumman Hossain (University of Maryland), Nirmalya Roy (University of Maryland)
OptimizationRobotic IntelligenceReinforcement Learning
🎯 What it does: This paper proposes a new reinforcement learning framework—Quasi-Potential Reinforcement Learning (QPRL), specifically designed to handle asymmetric traversal costs and irreversible actions in the environment.
QT-DoG: Quantization-Aware Training for Domain Generalization
Saqib Javed (EPFL), Mathieu Salzmann (EPFL)
Domain AdaptationTransformerImage
🎯 What it does: Using Quantization-Aware Training (QAT) to induce the model to converge to flat minima by introducing weight quantization noise, thereby enhancing domain generalization performance, and proposing QT-DoG and Ensemble of Quantization (EoQ) based on quantized models;
Quadratic Upper Bound for Boosting Robustness
Euijin You (Konkuk University), Hyang-Won Lee (Konkuk University)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A quadratic upper bound (QUB) loss function is proposed to improve fast adversarial training (FAT), enhancing model robustness while keeping training time unchanged.
Quadruple Attention in Many-body Systems for Accurate Molecular Property Predictions
Jiahua Rao (Sun Yat-sen University), Yuedong Yang (Sun Yat-sen University)
Drug DiscoveryGraph Neural NetworkGraph
🎯 What it does: A MABNet based on a quaternion attention mechanism is proposed for explicitly modeling molecular four-body interactions, achieving more accurate molecular property predictions.
Quamba2: A Robust and Scalable Post-training Quantization Framework for Selective State Space Models
Hung-Yueh Chiang (University of Texas at Austin), Diana Marculescu
TransformerText
🎯 What it does: Quamba2 is proposed, a robust and scalable post-training quantization framework for selective state space models (SSM), enabling full-link 8-bit/4-bit quantization from the embedding layer to the output head;
QuanONet: Quantum Neural Operator with Application to Differential Equation
Ruocheng Wang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Physics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes QuanONet, a quantum neural operator constructed from scalar quantum circuits, for learning nonlinear operators of differential equations;
Quantifying Memory Utilization with Effective State-Size
Rom Parnichkun, Stefano Massaroli (Liquid AI)
CompressionKnowledge DistillationTransformerLarge Language ModelTextSequential
🎯 What it does: An effective state-size (ESS) metric is proposed to quantify the extent to which causal sequence models (such as Transformers, SSMs, etc.) utilize working memory, and it is applied to model initialization, distillation, regularization, and state modulation analysis of language models.
Quantifying Prediction Consistency Under Fine-tuning Multiplicity in Tabular LLMs
Faisal Hamman (University of Maryland), Sanghamitra Dutta
Large Language ModelSupervised Fine-TuningTabular
🎯 What it does: This study explores the issue of prediction consistency caused by diversity in fine-tuning tabular large language models (Tabular LLMs) and proposes a new metric to quantify the consistency of individual predictions without the need for expensive model retraining.
Quantifying Treatment Effects: Estimating Risk Ratios via Observational Studies
Ahmed BOUGHDIRI, Erwan Scornet (Sorbonne Université)
TabularBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes and theoretically studies various methods for estimating risk ratios in observational studies, and verifies their performance through theoretical derivation and simulation.
QuantSpec: Self-Speculative Decoding with Hierarchical Quantized KV Cache
Rishabh Tiwari (University of California Berkeley), Amir Gholami (University of California Berkeley)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: QuantSpec is proposed, a self-speculative decoding framework that accelerates long-context LLM inference by quantizing KV caches and weights to 4 bits.
Quantum Algorithms for Finite-horizon Markov Decision Processes
Bin Luo (Chinese University of Hong Kong), John C.S. Lui (Chinese University of Hong Kong)
OptimizationReinforcement Learning from Human FeedbackPhysics Related
🎯 What it does: This paper proposes four quantum algorithm-based methods for solving finite time-domain, time-varying Markov Decision Processes (MDP), namely QVI-1, QVI-2, QVI-3, and QVI-4;
Quantum Optimization via Gradient-Based Hamiltonian Descent
Jiaqi Leng (University of California), Bin Shi (Fudan University)
OptimizationPhysics Related
🎯 What it does: A gradient-based Quantum Hamiltonian Descent algorithm (QHD) is proposed for unconstrained continuous optimization.
Quantum Speedup for Hypergraph Sparsification
Chenghua Liu (Chinese Academy of Sciences), Mingsheng Ying (University of Technology Sydney)
Physics Related
🎯 What it does: This study presents the first quantum algorithm for hypergraph sparsification, capable of constructing ε-spectral sparsification within near-linear size and achieving quantum speedup;
Quantum Speedups in Regret Analysis of Infinite Horizon Average-Reward Markov Decision Processes
Bhargav Ganguly (Purdue University), Vaneet Aggarwal (Purdue University)
Reinforcement LearningPhysics Related
🎯 What it does: A quantum reinforcement learning algorithm Q-UCRL for infinite average reward Markov decision processes is proposed, along with theoretical scheduling and reward analysis.
QuEst: Enhancing Estimates of Quantile-Based Distributional Measures Using Model Predictions
Zhun Deng (UNCat Chapel Hill), Richard Zemel (Columbia University)
ImageTextFinance Related
🎯 What it does: The QuEst framework is proposed, which combines a small amount of real observation data with a large amount of model prediction data to estimate quantile-related distribution quantities (such as CVaR, interval VaR, etc.) and provides strict confidence intervals.
QuEST: Stable Training of LLMs with 1-Bit Weights and Activations
Andrei Panferov (International Scientific and Technological Academy), Dan Alistarh (International Scientific and Technological Academy)
TransformerLarge Language ModelText
🎯 What it does: A quantization-aware training method named QuEST is proposed, which can stably train large language models with weights and activations ranging from 1-bit to 4-bit.
QuRe: Query-Relevant Retrieval through Hard Negative Sampling in Composed Image Retrieval
Jaehyun Kwak (KAIST), Sung-Ju Lee (KAIST)
RetrievalTransformerReinforcement LearningVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Proposes the QURE method, which improves composite image retrieval (CIR) by optimizing objectives through a reward model and implementing hard negative sample sampling to enhance retrieval quality;
QUTE: Quantifying Uncertainty in TinyML models with Early-exit-assisted ensembles for model-monitoring
Nikhil Pratap Ghanathe (University of British Columbia), Steven J E Wilton (University of British Columbia)
Anomaly DetectionComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkImageAudio
🎯 What it does: A resource-efficient TinyML uncertainty quantification method called QUTE is proposed, which forms a diverse single forward inference ensemble by adding a lightweight output head at the end of the network and utilizing early exit knowledge distillation.
R.I.P.: Better Models by Survival of the Fittest Prompts
Ping Yu (Meta), Jing Xu (Meta)
Data SynthesisOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: The Rejecting Instruction Preferences (RIP) method is proposed, which utilizes the quality of rejection responses (reward value and length) for each prompt, as well as the reward gap, to filter training data; and derives SelfRIP to generate high-quality synthetic instructions.
R*: Efficient Reward Design via Reward Structure Evolution and Parameter Alignment Optimization with Large Language Models
Pengyi Li (Tianjin University), YAN ZHENG
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningMultimodality
🎯 What it does: An automated reward function design framework R* is proposed, which splits reward design into two parts: reward structure evolution and parameter alignment optimization.
R2-T2: Re-Routing in Test-Time for Multimodal Mixture-of-Experts
Zhongyang Li (Johns Hopkins University), Tianyi Zhou (University of Maryland)
OptimizationKnowledge DistillationMixture of ExpertsMultimodalityBenchmark
🎯 What it does: This paper proposes a dynamic re-routing (R2-T2) during the testing phase of a multi-modal Mixture-of-Experts (MoE) model, enhancing model performance through local optimization of routing weights.
R3DM: Enabling Role Discovery and Diversity Through Dynamics Models in Multi-agent Reinforcement Learning
Harsh Goel (University of Texas at Austin), Sandeep P. Chinchali (Honda Research Institute)
Reinforcement LearningContrastive LearningTabularBenchmark
🎯 What it does: Achieve multi-agent collaborative training through role learning and dynamic models to enhance coordination.
Radio: Rate–Distortion Optimization for Large Language Model Compression
Sean I. Young (Harvard Medical School)
CompressionOptimizationTransformerLarge Language ModelText
🎯 What it does: Proposes an information-theoretic rate-distortion optimization framework for post-training quantization compression of large language models, generating weight quantization models with controllable bit rates.
RAGGED: Towards Informed Design of Scalable and Stable RAG Systems
Jennifer Hsia (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
GenerationRetrievalTransformerPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes the RAGGED framework for systematic evaluation of retrieval-augmented generation systems under different retrieval depths, reader models, retrievers, and noise conditions.
Raising the Bar: Investigating the Values of Large Language Models via Generative Evolving Testing
Han Jiang (Tongji University), Xing Xie (Microsoft Research Asia)
GenerationTransformerLarge Language ModelText
🎯 What it does: This study proposes an evaluation framework based on Generative Evolutionary Testing (GETA) for dynamically measuring the value and ethical alignment levels of large language models (LLMs).
Random Feature Representation Boosting
Nikita Zozoulenko (Imperial College London), Lukas Gonon (University of St. Gallen)
Point CloudTabular
🎯 What it does: A method for constructing deep residual random feature neural networks using Random Feature Boosting (RFRBoost) is proposed.
Random Policy Evaluation Uncovers Policies of Generative Flow Networks
Haoran He (Hong Kong University of Science and Technology), Ling Pan (Hong Kong University of Science and Technology)
GenerationReinforcement LearningFlow-based ModelSequential
🎯 What it does: The researchers view the training of Generative Flow Networks (GFlowNets) as flow iteration in dynamic programming, discovering that it is equivalent to the value evaluation of random policies in tree structures or DAGs that satisfy path invariance. They propose the Rectified Random Policy Evaluation (RPE) algorithm based on random policy evaluation, achieving the same reward matching effect as GFlowNets.
Random Registers for Cross-Domain Few-Shot Learning
Shuai Yi (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
Domain AdaptationTransformerPrompt EngineeringImage
🎯 What it does: In cross-domain few-shot learning, the authors found that using random registers instead of learnable prompts can significantly improve the cross-domain generalization performance of Vision Transformers. They proposed the REAP method, which enhances attention perturbation by randomly replacing clustering blocks in the image semantic region, resulting in a flatter loss landscape.
Randomized Dimensionality Reduction for Euclidean Maximization and Diversity Measures
Jie Gao (Rutgers University), Erik Waingarten (University of Pennsylvania)
OptimizationConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: This study investigates the effects of random dimensionality reduction on Euclidean maximization and diversity measurement problems, proving that a target dimension of O(λ) under hyperbolic dimension λ is sufficient to maintain approximate solutions for problems such as maximum matching, maximum TSP, and maximum spanning tree with a 1+ε error.
Rank-One Modified Value Iteration
Arman Sharifi Kolarijani, Mohamad Amin Sharifi Kolarijani (Delft University of Technology)
OptimizationReinforcement LearningGraph
🎯 What it does: Proposed rank-one approximation-based value iteration and Q-learning algorithms (R1-VI, R1-QL), achieving more efficient updates by using rank-one approximations of the transition probability matrix in the policy evaluation step.
Ranked Entropy Minimization for Continual Test-Time Adaptation
Jisu Han (Ajou University), Wonjun Hwang (Korea University)
Domain AdaptationOptimizationTransformerSupervised Fine-TuningImage
🎯 What it does: A Ranked Entropy Minimization (REM) method is proposed, which achieves efficient and stable updates for Continual Test-Time Adaptation (CTTA) using an advanced masking chain and dual losses (mask consistency loss + entropy ranking loss).
Ranked from Within: Ranking Large Multimodal Models Without Labels
Weijie Tu (Australian National University), Tongliang Liu (Sydney AI Centre, University of Sydney)
Recommendation SystemTransformerVision Language ModelMultimodality
🎯 What it does: This paper studies the method of unsupervised ranking for large multimodal models (LMM) in the absence of labeled data, exploring how to utilize the model's own uncertainty signals to assess its relative performance.
Ranking with Multiple Oracles: From Weak to Strong Stochastic Transitivity
Tao Jin (University of Virginia), Farzad Farnoud (University of Virginia)
Recommendation SystemOptimization
🎯 What it does: This paper studies the ranking problem under multi-source preference information and proposes two active learning algorithms, RMO-WST and RMO-SST, which achieve efficient ranking under weak and strong stochastic transitivity assumptions, respectively.
Rapid Overfitting of Multi-Pass SGD in Stochastic Convex Optimization
Shira Vansover-Hager (Tel Aviv University), Roi Livni (Tel Aviv University)
Optimization
🎯 What it does: This paper studies the overfitting behavior of multi-round (multi-epoch) stochastic gradient descent (SGD) in stochastic convex optimization (SCO), providing upper and lower bounds on the population loss of multi-round SGD, revealing the phenomenon of step overfitting from the first round to the second round, and generalizing the results to both sampling with and without replacement.
RAPID: Long-Context Inference with Retrieval-Augmented Speculative Decoding
Guanzheng Chen (National University of Singapore), Michael Qizhe Shieh (National University of Singapore)
GenerationRetrievalComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the RAPID method, which combines retrieval-augmented speculative decoding to accelerate long-context LLM inference and improve generation quality.
Raptor: Scalable Train-Free Embeddings for 3D Medical Volumes Leveraging Pretrained 2D Foundation Models
Ulzee An (University of California), Sriram Sankararaman (University of California)
ClassificationCompressionRepresentation LearningTransformerContrastive LearningBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A Raptor method is designed, utilizing a frozen 2D visual foundation model (such as DINOv2-L) to encode medical 3D volumes into three-axis slices, and compressing them into low-dimensional embeddings through random projection, without any 3D training;
RATE: Causal Explainability of Reward Models with Imperfect Counterfactuals
David Reber (University of Chicago), Victor Veitch (University of Chicago)
Recommendation SystemExplainability and InterpretabilityLarge Language ModelText
🎯 What it does: The RATE (Rewrite-based Attribute Treatment Estimator) method is proposed, which uses dual LLM rewriting to estimate the causal effects of reward models on high-level attributes such as sentiment, helpfulness, length, etc.
RBench: Graduate-level Multi-disciplinary Benchmarks for LLM & MLLM Complex Reasoning Evaluation
Meng-Hao Guo (Tsinghua University), Shi-min Hu
Large Language ModelVision Language ModelTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: A graduate-level multidisciplinary bilingual reasoning benchmark R Bench covering 108 disciplines and 1,094 questions has been constructed, along with 665 mixed image-text questions designed for multimodal models.
RE-Bench: Evaluating Frontier AI R&D Capabilities of Language Model Agents against Human Experts
Hjalmar Wijk (METR), Elizabeth Barnes (METR)
TransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: The RE-Bench benchmark is proposed to evaluate the automation capabilities of AI agents in cutting-edge AI research and development tasks, providing a direct comparison with human experts.
RE-IMAGINE: Symbolic Benchmark Synthesis for Reasoning Evaluation
Xinnuo Xu (Microsoft Research), Javier Gonzalez (Microsoft Research)
Large Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: By using symbolic representation and automated variant generation, the RE-IMAGINE framework is proposed for hierarchical evaluation of LLM's reasoning capabilities.
Re-ranking Reasoning Context with Tree Search Makes Large Vision-Language Models Stronger
Qi Yang (Alibaba Cloud Computing), Shiming Xiang (University of Chinese Academy of Sciences)
GenerationRetrievalTransformerVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: A multi-modal retrieval-enhanced generation framework RCTS is proposed, which constructs a knowledge base using reasoning context and improves the VQA performance of large visual language models through tree search reordering.
Reaction Graph: Towards Reaction-Level Modeling for Chemical Reactions with 3D Structures
Yingzhao Jian (Zhejiang University), Yi Yang (Zhejiang University)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes the Reaction Graph (RG) - a unified representation of chemical reaction graphs that can simultaneously capture the molecular structures of reactants and products, the reaction edges generated by atom mapping, and three-dimensional geometric information (edge lengths and angle edges), thereby directly learning the transformation features of reactions during the message passing process of graph neural networks.
RealRAG: Retrieval-augmented Realistic Image Generation via Self-reflective Contrastive Learning
Yuanhuiyi Lyu (Hong Kong University of Science and Technology), Xuming Hu (Hong Kong University of Science and Technology)
GenerationRetrievalDiffusion modelContrastive LearningImageMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes RealRAG, a text-to-image generation framework enhanced by real image retrieval, which uses retrieved missing knowledge to supplement the visual memory of the generative model, thereby improving the generation quality of fine-grained and novel objects.
Reasoning Limitations of Multimodal Large Language Models. A case study of Bongard Problems
Mikołaj Małkiński (Warsaw University of Technology), Jacek Mańdziuk (AGH University of Krakow)
TransformerLarge Language ModelPrompt EngineeringContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Evaluate and compare the reasoning and answer generation capabilities of multimodal large language models on Bongard problems (BPs).
Reasoning Through Execution: Unifying Process and Outcome Rewards for Code Generation
Zhuohao Yu (Peking University), Shikun Zhang (Peking University)
GenerationAI Code AssistantTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A unified framework for unsupervised process and outcome rewards, named ORPS, is proposed for code generation, utilizing executable feedback and LLM self-criticism to guide reasoning and implementation.
Reasoning-as-Logic-Units: Scaling Test-Time Reasoning in Large Language Models Through Logic Unit Alignment
Cheryl Li (Independent Researcher), Steven Y. Guo
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper proposes a framework RaLU that constructs reliable reasoning paths by extracting and aligning program logic units during inference, aiming to reduce the reasoning hallucinations produced by LLMs.
Recommendations with Sparse Comparison Data: Provably Fast Convergence for Nonconvex Matrix Factorization
Suryanarayana Sankagiri (Ecole Polytechnique Federale de Lausanne), Matthias Grossglauser (Ecole Polytechnique Federale de Lausanne)
Recommendation SystemOptimization
🎯 What it does: A non-convex matrix factorization recommendation system based on sparse comparative data is proposed and theoretically analyzed.
Reconstructing Cell Lineage Trees from Phenotypic Features with Metric Learning
Da Kuang (University of Pennsylvania), Junhyong Kim (University of Pennsylvania)
Representation LearningTransformerBiomedical Data
🎯 What it does: The CellTreeQM method is proposed, which uses metric learning to map single-cell transcriptomes into an embedding space to reconstruct cell lineage trees.
Rectifying Conformity Scores for Better Conditional Coverage
Vincent Plassier (Lagrange Mathematics and Computing Research Center), Eric Moulines (Mohamed bin Zayed University of Artificial Intelligence)
TabularBenchmark
🎯 What it does: A new method is proposed to generate confidence sets through trainable transformations to improve conditional coverage while ensuring the accuracy of marginal coverage.
Reducing Confounding Bias without Data Splitting for Causal Inference via Optimal Transport
Yuguang Yan (Guangdong University of Technology), Zhifeng Hao (Shantou University)
Optimization
🎯 What it does: This study proposes a method to reduce confounding bias in causal inference without the need for data partitioning, utilizing an optimal transport framework to learn balanced representations, applicable to both binary and continuous treatments;
Reducing Tool Hallucination via Reliability Alignment
Hongshen Xu (Shanghai Jiao Tong University), Kai Yu (Shanghai Jiao Tong University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: A systematic evaluation and mitigation method for the tool hallucination problem that arises when large language models call external tools is proposed, constructing the RelyToolBench benchmark and designing the Relign reliability alignment framework.
Reducing Variance of Stochastic Optimization for Approximating Nash Equilibria in Normal-Form Games
Linjian Meng (Nanjing University), Yang Gao (Nanjing University)
OptimizationReinforcement Learning from Human FeedbackTabular
🎯 What it does: This paper proposes a new unbiased and significantly variance-reduced Nash Advantage Loss (NAL) loss function for solving approximate Nash equilibria in large-scale normal-form games using non-convex stochastic optimization.
Redundancy Undermines the Trustworthiness of Self-Interpretable GNNs
Wenxin Tai (University of Electronic Science and Technology of China), Fan Zhou (University of Electronic Science and Technology of China)
Explainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: This study investigates the explanation credibility of self-explanatory graph neural networks, systematically discovering that redundancy leads to inconsistent and erroneous explanations, and proposes a simple explanation ensemble (EE) to mitigate the issue.
ReferSplat: Referring Segmentation in 3D Gaussian Splatting
Shuting He (Shanghai University of Finance and Economics), Henghui Ding (Fudan University)
Object DetectionSegmentationContrastive LearningGaussian SplattingTextPoint Cloud
🎯 What it does: A novel task called R3DGS is proposed for object segmentation in 3D Gaussian Splatting based on natural language descriptions, along with the construction of the corresponding dataset Ref-LERF.
Refined generalization analysis of the Deep Ritz Method and Physics-Informed Neural Networks
Xianliang Xu (Tsinghua University), Zhongyi Huang (Tsinghua University)
TabularSequentialPhysics Related
🎯 What it does: This paper derives refined generalization bounds for the Deep Ritz Method (DRM) and Physics-Informed Neural Networks (PINNs), focusing on prototype problems of the Poisson equation and the static Schrödinger equation.
Refining Adaptive Zeroth-Order Optimization at Ease
Yao Shu (Hong Kong University of Science and Technology), Zhongxiang Dai (Chinese University of Hong Kong)
OptimizationAdversarial AttackImage
🎯 What it does: An improved zero-order adaptive optimization method R-AdaZO is proposed, significantly enhancing the convergence speed of zero-order optimization in scenarios without gradient information.
Reflect-then-Plan: Offline Model-Based Planning through a Doubly Bayesian Lens
Jihwan Jeong (Google Research), Pascal Poupart (University of Waterloo)
Reinforcement LearningBenchmark
🎯 What it does: This paper proposes RefPlan, a dual Bayesian method that integrates offline model prediction with planning to achieve adaptive decision-making in offline reinforcement learning.
Reflection-Bench: Evaluating Epistemic Agency in Large Language Models
Lingyu Li (Shanghai Artificial Intelligence Laboratory), Yingchun Wang (Shanghai Artificial Intelligence Laboratory)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: The Reflection-Bench benchmark is proposed to evaluate the intrinsic knowledge subjectivity of large language models (LLMs) in seven cognitive processes (prediction, decision-making, perception, memory, counterfactual thinking, belief updating, and meta-reflection).
Reflection-Window Decoding: Text Generation with Selective Refinement
Zeyu Tang (Carnegie Mellon University), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A reflective window decoding framework is proposed, which alternates between refinement and generation during the generation process using a sliding window and a pause criterion, thereby alleviating the suboptimal problem of autoregressive decoding.
ReFocus: Visual Editing as a Chain of Thought for Structured Image Understanding
Xingyu Fu (University of Pennsylvania), Cha Zhang (Microsoft)
RecognitionObject DetectionSegmentationTransformerLarge Language ModelVision Language ModelImageMultimodalityTabularChain-of-Thought
🎯 What it does: By allowing a multimodal LLM to generate Python code for visual editing of input images (such as drawing boxes, masking, and highlighting), a visual thinking chain is realized, enhancing selective attention and multi-hop visual reasoning.