ICML 2025 Papers — Page 2
International Conference on Machine Learning · 3257 papers
Accelerating LLM Inference with Lossless Speculative Decoding Algorithms for Heterogeneous Vocabularies
Nadav Timor (Weizmann Institute of Science), David Harel (Weizmann Institute of Science)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes three algorithms for lossless speculative decoding under heterogeneous vocabulary conditions, addressing the limitations of traditional methods that require vocabulary consistency.
Accelerating PDE-Constrained Optimization by the Derivative of Neural Operators
Ze Cheng (Bosch), Hang Su (Tsinghua University)
OptimizationTabularTime SeriesPhysics Related
🎯 What it does: This paper proposes a PDE-constrained optimization acceleration framework that combines neural operators with gradient optimization. The core components include a reference neural operator (RNO) that utilizes optimization trajectory data for optimization-guided training, the introduction of a Virtual-Fourier layer to enhance derivative learning, and an iterative strategy that combines mixed solvers with neural operators.
Accelerating Quantum Reinforcement Learning with a Quantum Natural Policy Gradient Based Approach
Yang Xu (Purdue University), Vaneet Aggarwal (Purdue University)
OptimizationReinforcement LearningPhysics Related
🎯 What it does: A quantum natural policy gradient (QNPG) algorithm is proposed for optimizing policies in a model-free quantum reinforcement learning (QRL) environment, utilizing a quantum oracle to access the Markov decision process (MDP).
Accelerating Spectral Clustering under Fairness Constraints
Francesco Tonin (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
OptimizationComputational EfficiencyTabular
🎯 What it does: A variant of ADMM based on the differential convex framework is proposed, placing fairness constraints on MH through variable enhancement, avoiding expensive eigenvalue decomposition, and addressing the fair spectral clustering problem.
Accelerating Unbiased LLM Evaluation via Synthetic Feedback
Zhaoyi Zhou (Carnegie Mellon University), Andrea Zanette (Carnegie Mellon University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelText
🎯 What it does: By combining human evaluation and synthetic evaluation, the control variates method is used to reduce the amount of human labeling while maintaining unbiasedness, and to assess the win rate of LLMs.
Accurate and Efficient World Modeling with Masked Latent Transformers
Maxime Burchi (University of Wurzburg), Radu Timofte (University of Wurzburg)
Robotic IntelligenceTransformerReinforcement LearningWorld ModelSequentialBenchmark
🎯 What it does: Using Transformer + MaskGIT to predict spatial latent states, constructing a high-precision, low-overhead world model, and training an agent in the latent space to achieve superhuman performance in the Crafter environment.
Accurate Identification of Communication Between Multiple Interacting Neural Populations
Belle Liu (University of Washington), Matthew D. Golub (University of Washington)
Recurrent Neural NetworkAuto EncoderTime SeriesSequential
🎯 What it does: A multi-region dynamic system latent factor analysis model (MR-LFADS) has been developed to accurately identify communication from multi-region neural recordings.
Achieving Linear Speedup and Near-Optimal Complexity for Decentralized Optimization over Row-stochastic Networks
Liyuan Liang (Peking University), Kun Yuan (Peking University)
OptimizationImageTabular
🎯 What it does: Research on non-convex stochastic decentralized optimization on row-stochastic networks, proposing theoretical lower bounds and achieving near-optimal algorithms;
Action Dubber: Timing Audible Actions via Inflectional Flow
Wenlong Wan (South China University of Technology), Shengfeng He (Singapore Management University)
RecognitionObject DetectionTransformerContrastive LearningOptical FlowVideo
🎯 What it does: This paper proposes a task for audible action temporal localization and develops the TA Net 2 model, which detects key frames in videos that produce sound by inferring the second derivative of motion (inflectional flow) and achieves sound source localization through self-supervised spatial auxiliary training.
Action-Constrained Imitation Learning
Chia-Han Yeh (National Yang Ming Chiao Tung University), Ping-Chun Hsieh (National Yang Ming Chiao Tung University)
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningAgentic AISequential
🎯 What it does: Proposes the Action-Constrained Imitation Learning (ACIL) problem and designs the DTWIL method, which aligns expert trajectories using MPC+DTW to generate alternative demonstrations that satisfy action constraints, and then learns the policy using imitation learning algorithms such as behavior cloning.
Action-Dependent Optimality-Preserving Reward Shaping
Grant Collier Forbes, David Roberts
Reinforcement LearningVideoBenchmark
🎯 What it does: A motion-related reward shaping method called ADOPS is proposed, which effectively utilizes complex intrinsic rewards while keeping the optimal policy unchanged.
Action-Minimization Meets Generative Modeling: Efficient Transition Path Sampling with the Onsager-Machlup Functional
Sanjeev Raja (University of California Berkeley), Aditi S. Krishnapriyan (University of California Berkeley)
OptimizationDrug DiscoveryDiffusion modelFlow-based ModelSequentialPhysics Related
🎯 What it does: Utilizing pre-trained molecular generation models (denoising diffusion models and flow matching models) to achieve zero-shot transition path sampling by minimizing the Onsager-Machlup action.
ActionPiece: Contextually Tokenizing Action Sequences for Generative Recommendation
Yupeng Hou (University of California), Derek Zhiyuan Cheng (Google DeepMind)
Recommendation SystemGenerative Adversarial NetworkText
🎯 What it does: A context-aware action sequence tokenization method called ActionPiece is proposed for generative recommendation.
Activation by Interval-wise Dropout: A Simple Way to Prevent Neural Networks from Plasticity Loss
Sangyeon Park (Gwangju Institute of Science and Technology), KyungJoong Kim
ClassificationConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: A new activation method called Intervalized Dropout (AID) is proposed to prevent the loss of plasticity in neural networks.
Activation Space Interventions Can Be Transferred Between Large Language Models
Narmeen Fatimah Oozeer (Martian Learning), Amir Abdullah (Thoughtworks)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringAuto EncoderText
🎯 What it does: By learning the mapping of shared activation spaces between different large language models, this study investigates and implements activation space interventions (such as steering vectors) for transfer between models, aiming to remove backdoors, reject harmful prompts, and address the issue of 'corrupted capabilities'.
Active Evaluation Acquisition for Efficient LLM Benchmarking
Yang Li (Amazon Web Services), Graham Horwood
Computational EfficiencyLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes an Active Evaluation Acquisition (AEA) strategy and a neural process model to significantly reduce the number of prompts required during benchmark evaluations of large language models, thereby improving evaluation efficiency.
Active feature acquisition via explainability-driven ranking
Osman Berke Guney (Boston University), Vijaya B Kolachalama
OptimizationExplainability and InterpretabilityTransformerReinforcement LearningImageTabular
🎯 What it does: This paper proposes an active feature acquisition framework based on local interpretability methods, utilizing explanation techniques such as SHAP/LIME to obtain the instance feature importance ranking for each sample, and trains a decision Transformer to predict the next most important feature to be acquired, forming a complete process of dynamic feature acquisition by instance.
Active Fine-Tuning of Multi-Task Policies
Marco Bagatella (ETH Zurich), Andreas Krause (ETH Zurich)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningAgentic AISequential
🎯 What it does: This study investigates how to efficiently fine-tune pre-trained multi-task strategies through active task selection to minimize demonstration costs.
Active Learning for Efficient Discovery of Optimal Combinatorial Perturbations
Jason Qin (Neptune Bio), Yuhan Hao (Neptune Bio)
OptimizationDrug DiscoveryBiomedical Data
🎯 What it does: The NAIAD framework is proposed, which uses active learning to predict and discover the most effective gene or drug combinations based on single-gene effects, significantly reducing experimental cycles.
Active Learning of Deep Neural Networks via Gradient-Free Cutting Planes
Erica Zhang (Stanford University), Mert Pilanci (Stanford University)
ClassificationOptimizationText
🎯 What it does: This paper proposes a deep neural network training and active learning framework based on gradient-free cutting-plane methods, which can achieve geometric shrinkage in the parameter space of non-convex ReLU networks and converge to the optimal solution.
Active Learning with Selective Time-Step Acquisition for PDEs
Yegon Kim (Korea Advanced Institute of Science and Technology), Juho Lee (Korea Advanced Institute of Science and Technology)
Time SeriesBenchmarkPhysics RelatedOrdinary Differential Equation
🎯 What it does: A framework for active sampling in PDE surrogate model learning is proposed, which can selectively choose only the most important time steps for high-cost numerical solutions, while approximating the remaining time steps with a surrogate model, thereby significantly reducing data collection costs.
Active Reward Modeling: Adaptive Preference Labeling for Large Language Model Alignment
Yunyi Shen (Massachusetts Institute of Technology), Jean-Francois Ton (ByteDance)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes an active preference label sampling method based on Fisher information (Active Reward Modeling) for training reward models of large language models. It selects the most informative comparison pairs by performing D-optimal design on the features of the last layer of the reward model.
Active Treatment Effect Estimation via Limited Samples
Zhiheng Zhang (Shanghai University of Finance and Economics), Zhouchen Lin (Peking University)
Tabular
🎯 What it does: A finite sample-based active sampling framework RWAS is proposed to efficiently estimate causal effects (both individual and average) and provide non-asymptotic error bounds;
Actor-Critics Can Achieve Optimal Sample Efficiency
Kevin Tan (University of Pennsylvania), Yuting Wei (University of Pennsylvania)
OptimizationComputational EfficiencyReinforcement Learning
🎯 What it does: A new strategy-critic algorithm (DOUHUA and NORA) is designed, utilizing optimistic exploration, offline critic estimation, sparse critic updates, and policy resets, achieving a sample complexity of 1/ε² and a regret rate of √T in finite-horizon MDPs with general function approximation.
Ad Hoc Teamwork via Offline Goal-Based Decision Transformers
Xinzhi Zhang (South China University of Technology), Mengchen Zhao (South China University of Technology)
TransformerReinforcement LearningSequential
🎯 What it does: A hierarchical sequence modeling framework named TAGET is proposed, which learns to adapt real-time collaboration strategies with unknown teammates using offline multi-agent data.
Ad-Hoc Human-AI Coordination Challenge
Tin Dizdarević (University of Oxford), Jakob Nicolaus Foerster (University of Oxford)
Recurrent Neural NetworkReinforcement LearningAgentic AITabularSequentialBenchmark
🎯 What it does: The Ad-Hoc Human-AI Coordination Challenge (AH2AC2) is proposed, which evaluates the real-time collaboration ability of AI and humans by using human proxy agents in the cooperative card game Hanabi.
AdaDecode: Accelerating LLM Decoding with Adaptive Layer Parallelism
Zhepei Wei (University of Virginia), Yu Meng (University of Virginia)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: To address the bottleneck of autoregressive decoding speed in large language models, we propose AdaDecode, which enhances decoding throughput by making early predictions in the intermediate layers and utilizing a lightweight LM head to achieve adaptive layer-level parallelism.
Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation
Jintao Tong (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
SegmentationDomain AdaptationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper studies the cross-domain few-shot semantic segmentation task and finds that adapters can structurally separate domain-related information. It proposes the Domain Feature Navigator (DFN) structure based on this characteristic and introduces the SAM-SVN mechanism to control overfitting.
Adapting Precomputed Features for Efficient Graph Condensation
Yuan Li (National University of Singapore), Bingsheng He (National University of Singapore)
Computational EfficiencyRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes GCPA—a graph aggregation framework that first precomputes and then adapts, completely skipping trajectory matching. It generates a structure-free synthetic graph through a single message propagation and multi-scale aggregation, and then fine-tunes it using multi-class contrastive loss and diversity constraints.
Adapting to Evolving Adversaries with Regularized Continual Robust Training
Sihui Dai (CapitalOne), Arjun Bhagoji (Indian Institute of Technology Bombay)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This study investigates how to maintain the robustness of models against known and emerging attacks through Continuous Robust Training (CRT) and Adversarial ℓ2 Regularization (ALR) in the context of evolving adversarial attacks.
Adapting to Linear Separable Subsets with Large-Margin in Differentially Private Learning
Erchi Wang (Halıcıoglu Data Science Institute UC San Diego), Yu-Xiang Wang (Halıcıoglu Data Science Institute UC San Diego)
OptimizationSafty and PrivacyTransformerSupervised Fine-TuningImage
🎯 What it does: An efficient (ε,δ)-DP algorithm is proposed for unsupervised learning of half-spaces, which can adaptively utilize the large margin of the linearly separable subset obtainable from the training set, thereby achieving lower error while ensuring privacy.
Adapting While Learning: Grounding LLMs for Scientific Problems with Tool Usage Adaptation
Bohan Lyu (Tsinghua University), Rose Yu (University of California)
TransformerLarge Language ModelSupervised Fine-TuningTextPhysics Related
🎯 What it does: A two-stage post-training framework called AWL (Adapting While Learning) is proposed, enabling large language models to adaptively assess difficulty and decide whether to use tools when solving scientific problems, thereby improving answer accuracy and tool usage efficiency.
Adaptive Data Collection for Robust Learning Across Multiple Distributions
Chengbo Zang (Columbia University), Javad Ghaderi (Columbia University)
OptimizationData-Centric LearningImage
🎯 What it does: Under a limited annotation budget, an adaptive data collection framework is proposed to achieve robust learning in multi-distribution scenarios, with the goal of minimizing the maximum expected loss across all data sources.
Adaptive Elicitation of Latent Information Using Natural Language
Jimmy Wang (Columbia University), Hongseok Namkoong (Columbia University)
Meta LearningTransformerLarge Language ModelText
🎯 What it does: An adaptive information retrieval framework based on natural language is proposed, utilizing large language models to predict future answers and actively select the most informative questions to reduce uncertainty about potential entities.
Adaptive Estimation and Learning under Temporal Distribution Shift
Dheeraj Baby (Amazon), Rohit Pyati (Amazon)
ClassificationRestorationDomain AdaptationTime SeriesFinance Related
🎯 What it does: This paper studies the estimation and learning problem under time distribution drift, proposing an adaptive estimation method based on wavelet thresholding, and extending it to binary classification and total variation denoising.
Adaptive Exploration for Multi-Reward Multi-Policy Evaluation
Alessio Russo (Boston University), Aldo Pacchiano (Broad Institute of MIT and Harvard)
OptimizationNeural Architecture SearchReinforcement LearningTabular
🎯 What it does: This study investigates the sample complexity of multi-reward and multi-policy evaluation in online discount MDPs, providing a theoretical lower bound for (ε, δ)-PAC and corresponding exploration strategies.
Adaptive Flow Matching for Resolving Small-Scale Physics
Stathi Fotiadis (NVIDIA), Morteza Mardani (NVIDIA)
Diffusion modelFlow-based ModelAuto EncoderTime SeriesPhysics Related
🎯 What it does: An Adaptive Flow Matching (AFM) framework is proposed, which encodes low-resolution meteorological data into a latent distribution and generates high-resolution fine-scale physical information through flow matching.
Adaptive kernel predictors from feature-learning infinite limits of neural networks
Clarissa Lauditi (Harvard University), Cengiz Pehlevan (Harvard University)
OptimizationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: Derived the adaptive kernel predictor (aNBK and aNTK) for infinitely wide deep networks under the limit of feature learning and provided numerical solution methods, proving that even under a rich training regime, the network can still be viewed as a data-adaptive kernel machine.
Adaptive Learn-then-Test: Statistically Valid and Efficient Hyperparameter Selection
Matteo Zecchin (King's College London), Osvaldo Simeone (King's College London)
Hyperparameter SearchReinforcement LearningPrompt EngineeringSequential
🎯 What it does: An adaptive learning-testing (aLTT) framework is proposed for hyperparameter selection under limited samples through sequential multiple hypothesis testing, providing control over family-wise error rate (FWER) or false discovery rate (FDR);
Adaptive Localization of Knowledge Negation for Continual LLM Unlearning
Abudukelimu Wuerkaixi (Tsinghua University), Changshui Zhang (Tsinghua University)
Large Language ModelText
🎯 What it does: A new method for continual unlearning in large language models, called ALKN, is proposed, which maximizes the overall utility of the model while deleting sensitive or inappropriate knowledge.
Adaptive Median Smoothing: Adversarial Defense for Unlearned Text-to-Image Diffusion Models at Inference Time
Xiaoxuan Han (Institute of Automation, Chinese Academy of Sciences), Jing Dong (Institute of Automation, Chinese Academy of Sciences)
GenerationComputational EfficiencyAdversarial AttackDiffusion modelImageText
🎯 What it does: This paper proposes an Adaptive Median Smoothing defense for a text-image diffusion model with removed concepts during inference, dynamically injecting anisotropic noise based on the correlation of tokens to enhance robustness against adversarial attacks.
Adaptive Message Passing: A General Framework to Mitigate Oversmoothing, Oversquashing, and Underreaching
Federico Errica (NEC Laboratories Europe), Francesco Alesiani (NEC Laboratories Europe)
Graph Neural NetworkGraph
🎯 What it does: An Adaptive Message Passing (AMP) framework has been developed, enabling graph neural networks to autonomously determine the depth and filtering strategy of message passing during training, thereby alleviating long-range dependency issues such as over-smoothing, over-compression, and the inability to reach remote nodes.
Adaptive Multi-prompt Contrastive Network for Few-shot Out-of-distribution Detection
Xiang Fang (Nanyang Technological University), Blaise Genest (CNRS)
Anomaly DetectionTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Proposes the Adaptive Multi-prompt Contrastive Network (AMCN) to achieve OOD detection with a small number of ID samples;
Adaptive Partitioning Schemes for Optimistic Optimization
Raja Sunkara (Missouri University of Science and Technology), Ardhendu Tripathy (OpsCanvas)
Optimization
🎯 What it does: An adaptive partitioning optimistic optimization algorithm is designed, using a single-layer neural network to learn the low-dimensional subspace of a black-box function, and executing SequOOL within this subspace for efficient optimization.
Adaptive Sample Sharing for Multi Agent Linear Bandits
Hamza Cherkaoui (Huawei Noah's Ark Lab), Igor Colin (LTCI, Télécom Paris, Institut Polytechnique de Paris)
Tabular
🎯 What it does: This paper proposes an adaptive sample sharing algorithm (BASS) for collaborative learning in multi-agent linear bandit environments, which automatically determines when to share observations to balance bias and uncertainty, thereby reducing cumulative pseudo-loss.
Adaptive Self-improvement LLM Agentic System for ML Library Development
Genghan Zhang (Stanford University), Kunle Olukotun (Stanford University)
OptimizationAI Code AssistantTransformerLarge Language ModelAgentic AIText
🎯 What it does: An adaptive self-improving LLM agent system is proposed for automatically generating high-performance ML library code in an ASPL environment lacking examples.
Adaptive Sensitivity Analysis for Robust Augmentation against Natural Corruptions in Image Segmentation
Laura Yu Zheng, Ming Lin
SegmentationTransformerImage
🎯 What it does: Proposes an online data augmentation method based on adaptive sensitivity analysis to enhance the robustness of semantic segmentation models against natural noise.
AdaptiveStep: Automatically Dividing Reasoning Step through Model Confidence
Yuliang Liu (Nanjing University), Zhouhan Lin (Shanghai Jiaotong University)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes the AdaptiveStep method, which uses the model's confidence in the next word to automatically divide inference steps, and trains a Process Reward Model (ASPRM) based on this for mathematical reasoning and code generation.
AdaPTS: Adapting Univariate Foundation Models to Probabilistic Multivariate Time Series Forecasting
Abdelhakim Benechehab (Huawei Noah's Ark Lab), Balázs Kégl (Huawei Noah's Ark Lab)
Auto EncoderGenerative Adversarial NetworkTime Series
🎯 What it does: The AdaPTS framework is proposed, which utilizes feature space adapters to transfer pre-trained univariate time series foundational models (FM) to multivariate probabilistic prediction tasks.
AdaSplash: Adaptive Sparse Flash Attention
Nuno Gonçalves (Instituto Superior Tecnico), Andre Martins
RetrievalOptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper presents ADASPLASH, a hardware-friendly implementation of α-entmax that combines Hybrid Halley-bisection iteration and a custom Triton kernel to achieve efficient forward/backward computation of sparse attention.
AdaWorld: Learning Adaptable World Models with Latent Actions
Shenyuan Gao (Hong Kong University of Science and Technology), Chuang Gan (Massachusetts Institute of Technology)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerDiffusion modelAuto EncoderWorld ModelVideo
🎯 What it does: This paper presents AdaWorld, a framework that extracts latent actions through self-supervision and introduces action information during the world model pre-training phase, enabling the world model to efficiently adapt to different environments and achieve action transfer and rapid learning.
ADDQ: Adaptive distributional double Q-learning
Leif Döring (University of Mannheim), Martin Slowik (University of Mannheim)
Reinforcement LearningTabular
🎯 What it does: An Adaptive Distributed Double Q-Learning (ADDQ) algorithm is proposed, which dynamically decides whether to use standard Q-learning or double Q-learning updates based on the sample variance estimated by distributed RL, in order to alleviate the overestimation problem.
Addressing Concept Mislabeling in Concept Bottleneck Models Through Preference Optimization
Emiliano Penaloza (Mila Quebec AI Institute), Mateo Espinosa Zarlenga (University of Cambridge)
OptimizationReinforcement LearningImage
🎯 What it does: A Concept Preference Optimization (CPO) loss based on Direct Preference Optimization is proposed, which directly optimizes the concept posterior of the Concept Bottleneck Model (CBM) to alleviate the impact of concept label noise on model performance.
Addressing Imbalanced Domain-Incremental Learning through Dual-Balance Collaborative Experts
Lan Li (Nanjing University), De-Chuan Zhan (Nanjing University)
Domain AdaptationTransformerMixture of ExpertsImage
🎯 What it does: In Domain Incremental Learning (DIL), this paper proposes the Dual-Balance Collaborative Experts (DCE) framework to address the issues of internal class imbalance and cross-domain class distribution drift caused by sample imbalance.
Addressing Misspecification in Simulation-based Inference through Data-driven Calibration
Antoine Wehenkel (Apple), marco cuturi
Neural Radiance FieldAuto EncoderBenchmark
🎯 What it does: A robust posterior estimation method based on optimal transport and fine-tuning statistics (RoPE) is proposed to correct uncertainty from a small amount of real labeled data when the simulator is misconfigured.
ADHMR: Aligning Diffusion-based Human Mesh Recovery via Direct Preference Optimization
Wenhao Shen (Nanyang Technological University), Guosheng Lin (Nanyang Technological University)
Pose EstimationOptimizationTransformerDiffusion modelMesh
🎯 What it does: Proposes the ADHMR framework, which aligns the diffusion-based 3D human mesh recovery model through direct preference optimization;
ADIOS: Antibody Development via Opponent Shaping
Sebastian Rene Towers (University of Oxford), Jakob Nicolaus Foerster (University of Oxford)
OptimizationMeta LearningDrug DiscoveryReinforcement LearningBiomedical Data
🎯 What it does: The ADIOS framework is proposed, treating antibody design as a meta-learning process of adversarial shaping, using an outer loop to optimize antibodies and an inner loop to simulate the escape process of viruses evolving alongside antibodies.
Adjoint Sampling: Highly Scalable Diffusion Samplers via Adjoint Matching
Aaron J Havens, Ricky T. Q. Chen (Meta)
GenerationOptimizationDrug DiscoveryDiffusion modelSequential
🎯 What it does: An efficient algorithm called Adjoint Sampling is proposed for learning the sampling of diffusion processes from unnormalized energy functions.
Adjusting Model Size in Continual Gaussian Processes: How Big is Big Enough?
Guiomar Pescador-Barrios (Imperial College London), Mark van der Wilk (University of Oxford)
Tabular
🎯 What it does: This paper proposes an algorithm for adaptively increasing the number of inducing points in sparse Gaussian processes, enabling continuous learning models to maintain performance close to full batch processing in the face of unknown data volumes.
Adjustment for Confounding using Pre-Trained Representations
Rickmer Schulte (Ludwig Maximilian University of Munich), Thomas Nagler (Ludwig Maximilian University of Munich)
Representation LearningTransformerLarge Language ModelImageText
🎯 What it does: The study estimates the average treatment effect (ATE) using pre-trained representations for non-tabular data (images, text) and provides feasible theoretical regularization conditions.
AdvAgent: Controllable Blackbox Red-teaming on Web Agents
Chejian Xu (University of Illinois Urbana-Champaign), Bo Li (University of Illinois Urbana-Champaign)
Adversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextFinance Related
🎯 What it does: A black-box red team framework called AdvAgent is proposed to inject covert adversarial prompts into web proxies, inducing them to perform incorrect or malicious actions.
Advancing Constrained Monotonic Neural Networks: Achieving Universal Approximation Beyond Bounded Activations
Davide Sartor (University of Padova), Gian Antonio Susto (University of Padova)
ClassificationOptimizationConvolutional Neural NetworkTabular
🎯 What it does: This paper proposes a novel constrained monotonic multilayer perceptron (Monotonic MLP), which achieves global monotonicity and global approximation capability when using convex saturated activation functions like ReLU by alternately using activation functions with different saturation sides or by employing non-positive weight constraints.
Advancing Personalized Learning with Neural Collapse for Long-Tail Challenge
Hanglei Hu (East China Normal University), Bo Jiang (East China Normal University)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: The NCAL method is proposed, utilizing the idea of neural collapse to create a simple ETF structure for text representation learning in large language models, and enhancing the generalization performance of the LoRA fine-tuning model under long-tail distributions through Text-Modality Collapse (TC) regularization.
Adversarial Combinatorial Semi-bandits with Graph Feedback
Yuxiao Wen (New York University)
OptimizationReinforcement Learning from Human FeedbackGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This study investigates the combinatorial semi-bandit problem with graph structure feedback, establishing optimal lower and upper bounds under the optimal order of magnitude, and proposes an algorithm named OSMD-G to achieve these bounds.
Adversarial Cooperative Rationalization: The Risk of Spurious Correlations in Even Clean Datasets
Wei Liu (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
ClassificationExplainability and InterpretabilityAdversarial AttackRecurrent Neural NetworkGraph Neural NetworkTransformerGenerative Adversarial NetworkTextGraph
🎯 What it does: This study investigates the model self-induced spurious correlations that may arise when the generator and predictor collaborate within the self-explanatory Rationalization framework, and proposes an attack-based inspection and instruction mechanism (A2I) to identify and suppress these spurious correlations, thereby enhancing the quality of rationalization and the robustness of the model.
Adversarial Inception Backdoor Attacks against Reinforcement Learning
Ethan Rathbun (Northeastern University), Christopher Amato (Northeastern University)
Adversarial AttackReinforcement LearningSequential
🎯 What it does: For deep reinforcement learning models, a trigger and reward deception are implanted during the training phase to achieve backdoor attacks on specific actions during deployment.
Adversarial Inputs for Linear Algebra Backends
Jonas Möller (Berlin Institute for the Foundations of Learning and Data), Konrad Rieck (Berlin Institute for the Foundations of Learning and Data)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper studies the subtle differences in floating-point operations among different linear algebra backends (such as Intel MKL, Nvidia cuBLAS, Apple Accelerate, etc.) and utilizes these differences to construct 'Chimera' adversarial inputs, causing the same model to produce contradictory predictions across different backends.
Adversarial Perturbations Are Formed by Iteratively Learning Linear Combinations of the Right Singular Vectors of the Adversarial Jacobian
Thomas Paniagua (North Carolina State University), Tianfu Wu (North Carolina State University)
OptimizationAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a white-box ordered Top-K adversarial attack method called RisingAttacK based on Sequential Quadratic Programming (SQP), which directly learns perturbations in the image space that satisfy a specified ranking order.
Adversarial Reasoning at Jailbreaking Time
Mahdi Sabbaghi (University of Pennsylvania), Hamed Hassani (University of Pennsylvania)
Adversarial AttackTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper proposes an automatic jailbreak method based on adversarial reasoning, which efficiently generates prompts that can bypass security defenses by utilizing the continuous loss signals of the target model to guide reasoning, verification, and search.
Adversarial Robust Generalization of Graph Neural Networks
Chang Cao (Huazhong Agricultural University), Hong Chen (Huazhong Agricultural University)
Adversarial AttackGraph Neural NetworkGraph
🎯 What it does: This paper presents a high-probability generalization error upper bound for Graph Neural Networks (GNNs) under adversarial attacks in a transductive learning scenario, and provides specific covering number estimates and generalization analyses for mainstream models such as GCN, APPNP, and GCNII. It further verifies the impact of model architecture, graph filtering, perturbation budget, feature dimensions, and weight norms on adversarial generalization.
Adversarial Robustness in Two-Stage Learning-to-Defer: Algorithms and Guarantees
Yannis Montreuil (National University of Singapore), Wei Tsang Ooi (National University of Singapore)
Object DetectionOptimizationAdversarial AttackImageTabular
🎯 What it does: This paper studies the adversarial robustness of the Learning-to-Defer (L2D) framework, proposing both untargeted and targeted attacks, and designing a provably Bayes-consistent adversarially smoothed surrogate loss along with the corresponding SARD algorithm.
Adversarial Robustness via Deformable Convolution with Stochasticity
Yanxiang Ma (University of Sydney), Chang Xu (University of Sydney)
Adversarial AttackConvolutional Neural NetworkImageStochastic Differential Equation
🎯 What it does: A deformable convolution with random offsets (DCS) is proposed to achieve structural-level random defense in convolutional layers;
Adversaries Can Misuse Combinations of Safe Models
Erik Jones (University of California Berkeley), Jacob Steinhardt (University of California Berkeley)
GenerationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study investigates the risks of utilizing ensemble models to complete malicious tasks in a multi-model ecosystem, demonstrating that even if a single model is secure, attackers can still combine cutting-edge models with weaker models through task decomposition to execute attacks such as vulnerability code, malicious scripts, explicit images, and manipulated tweets.
AdvI2I: Adversarial Image Attack on Image-to-Image Diffusion Models
Yaopei Zeng (Pennsylvania State University), Lu Lin (Pennsylvania State University)
Image TranslationGenerationAdversarial AttackDiffusion modelAuto EncoderGenerative Adversarial NetworkImageText
🎯 What it does: A framework for adversarial image attacks on image-to-image diffusion models (AdvI2I) has been designed and implemented, which generates adversarial images to induce the model to produce NSFW content and can bypass existing safety checks.
AdvPrompter: Fast Adaptive Adversarial Prompting for LLMs
Anselm Paulus (University of Tuebingen), Yuandong Tian (Meta)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A language model named AdvPrompter is proposed and trained to quickly generate readable and security-bypassing adversarial suffixes, enabling automated red teaming attacks and adversarial training on large language models.
AEQA-NAT : Adaptive End-to-end Quantization Alignment Training Framework for Non-autoregressive Machine Translation
Xiangyu Qu (Shandong University), Liang Li (Shandong University)
GenerationOptimizationKnowledge DistillationTransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes the AEQA-NAT framework, which utilizes the Semantic Quantization Space (SQS) to achieve consistency in training and inference for Non-Autoregressive Translation (NAT), eliminating the training-inference gap, improving translation quality, and maintaining extremely high decoding speed.
Aequa: Fair Model Rewards in Collaborative Learning via Slimmable Networks
Nurbek Tastan (Mohamed bin Zayed University of Artificial Intelligence), Karthik Nandakumar (Michigan State University)
Federated LearningSafty and PrivacyImage
🎯 What it does: The AEQUA framework is proposed, which achieves model rewards based on participant contributions through slimmable neural networks in collaborative learning, addressing the issue of fair reward distribution.
AffectGPT: A New Dataset, Model, and Benchmark for Emotion Understanding with Multimodal Large Language Models
Zheng Lian (Institute of Automation Chinese Academy of Sciences), Jianhua Tao (Tsinghua University)
ClassificationGenerationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodalityBenchmark
🎯 What it does: A large-scale emotional description dataset MER-Caption has been established, a pre-fusion multimodal large language model AffectGPT has been proposed outside of LLMs, and a unified evaluation benchmark MER-UniBench has been created.
AffinityFlow: Guided Flows for Antibody Affinity Maturation
Can Chen, Ron Benson (Amazon)
OptimizationDrug DiscoveryFlow-based ModelBiomedical Data
🎯 What it does: Aiming at antibody affinity maturation, an iterative optimization framework is proposed that only uses antibody and antigen sequences. It first uses structure generation to guide affinity enhancement, and then achieves sequence mutation through inverse folding, iterating in cycles.
AGAV-Rater: Adapting Large Multimodal Model for AI-Generated Audio-Visual Quality Assessment
Yuqin Cao (Shanghai Jiao Tong University), Guangtao Zhai (Shanghai Jiao Tong University)
Recommendation SystemTransformerLarge Language ModelVision Language ModelVideoMultimodalityAudio
🎯 What it does: A dataset called AGAVQA-3k was constructed, and the AGAV-Rater model was proposed to evaluate the audio quality, audio-video consistency, and overall quality of AI-generated audio and video, as well as to select the best output.
Agent Reviewers: Domain-specific Multimodal Agents with Shared Memory for Paper Review
Kai Lu (Chinese Academy of Sciences), Gaofeng Meng (Chinese Academy of Sciences)
Large Language ModelAgentic AITextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: A multi-modal, multi-agent paper review system called Agent Reviewers is proposed, simulating the real review process;
Agent Workflow Memory
Zora Zhiruo Wang (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
Large Language ModelAgentic AITextBenchmark
🎯 What it does: Proposes Agent Workflow Memory (AWM), which guides language model-driven web navigation agents by extracting reusable workflows from completed task trajectories.
Agent-as-a-Judge: Evaluate Agents with Agents
Mingchen Zhuge (King Abdullah University of Science and Technology), Jürgen Schmidhuber
AI Code AssistantLarge Language ModelAgentic AIText
🎯 What it does: This paper proposes the Agent-as-a-Judge framework, which utilizes agentic systems to evaluate other agentic systems, and assesses code generation tasks based on this framework.
Agent-Centric Actor-Critic for Asynchronous Multi-Agent Reinforcement Learning
Whiyoung Jung (LGAI Research), Woohyung Lim (LGAI Research)
Recurrent Neural NetworkReinforcement LearningAgentic AISequential
🎯 What it does: The Agent-Centric Actor-Critic (ACAC) algorithm is proposed to address coordination and learning efficiency issues in asynchronous multi-agent reinforcement learning with macro actions.
Aggregation Buffer: Revisiting DropEdge with a New Parameter Block
Dooho Lee (Korea Advanced Institute of Science and Technology), Jaemin Yoo (Korea Advanced Institute of Science and Technology)
ClassificationOptimizationKnowledge DistillationGraph Neural NetworkGraph
🎯 What it does: A new parameter block called Aggregation Buffer (AGG B) is proposed to enhance the edge robustness of Graph Neural Networks (GNNs). By adding this block to a pre-trained GNN and using DropEdge for two-stage training, significant improvements in node classification performance are achieved.
Aggregation of Dependent Expert Distributions in Multimodal Variational Autoencoders
Rogelio A. Mancisidor (BI Norwegian Business School), Michael Kampffmeyer (Norwegian Computing Center)
GenerationData SynthesisAuto EncoderImageMultimodality
🎯 What it does: A method considering the dependency between expert distributions, CoDE, is proposed, and based on this method, a CoDE-VAE multimodal variational autoencoder is constructed.
Aguvis: Unified Pure Vision Agents for Autonomous GUI Interaction
Yiheng Xu (University of Hong Kong), Caiming Xiong (Salesforce Research)
OptimizationRobotic IntelligenceTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: A purely visual unified GUI interaction framework AGUVIS has been designed and implemented, supporting cross-platform automation task completion.
AI for Global Climate Cooperation: Modeling Global Climate Negotiations, Agreements, and Long-Term Cooperation in RICE-N
Tianyu Zhang (Mila - Quebec AI Institute), Stephan Zheng (Asari AI)
Reinforcement LearningTime Series
🎯 What it does: RICE-N is proposed—a framework that combines a multi-regional integrated assessment model with multi-agent reinforcement learning to simulate global climate negotiations, agreements, and long-term cooperation; it implements two types of negotiation agreements (bilateral negotiations and basic clubs), compares different benchmark scenarios, and assesses their impact on climate and economic outcomes.
AKORN: Adaptive Knots generated Online for RegressioN splines
Sunil Madhow (University of California San Diego), Yu-Xiang Wang (University of California San Diego)
Tabular
🎯 What it does: A non-parametric algorithm named AKORN is proposed for offline non-parametric regression, which can adaptively select nodes to fit the geometric features of real data.
AKRMap: Adaptive Kernel Regression for Trustworthy Visualization of Cross-Modal Embeddings
Yilin Ye (Hong Kong University of Science and Technology), Wei Zeng (Hong Kong University of Science and Technology)
Representation LearningDiffusion modelImageTextMultimodality
🎯 What it does: Proposes AKRMap, an adaptive kernel regression supervised dimensionality reduction method for trustworthy visualization of cross-modal embedding metrics.
Alberta Wells Dataset: Pinpointing Oil and Gas Wells from Satellite Imagery
Pratinav Seth (Mila Quebec AI Institute), David Rolnick (McGill University)
Object DetectionSegmentationConvolutional Neural NetworkTransformerImageBenchmark
🎯 What it does: This paper constructs the Alberta Wells dataset and conducts benchmark experiments on the localization and segmentation tasks of oil and gas wells (including abandoned, suspended, and active wells) using high-resolution multispectral satellite images from Planet Labs.
Algorithm Development in Neural Networks: Insights from the Streaming Parity Task
Loek van Rossem (University College London), Andrew M Saxe
Recurrent Neural NetworkSequential
🎯 What it does: Experiments and theoretical analysis on streaming parity tasks using RNNs, proposing and validating an implicit representation merging mechanism, demonstrating that the network ultimately forms a finite automaton through two-stage learning (tree fitting → state merging), achieving perfect generalization for sequences of arbitrary length.
Algorithmic Recourse for Long-Term Improvement
Kentaro Kanamori (Fujitsu Limited), Takuya Takagi
Recommendation SystemOptimizationExplainability and InterpretabilityReinforcement LearningTabular
🎯 What it does: This paper proposes an online learning-based algorithm interpretability framework (ARLIM) that can dynamically suggest executable actions for each incoming instance, which can improve real-world outcomes (h*‑valid) in the long term.
Algorithms and Hardness for Active Learning on Graphs
Vincent Cohen-Addad (Google), Simon Meierhans (ETH Zurich)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A resource augmentation algorithm for the active learning label selection (GLS) problem on general weighted graphs is proposed, which can achieve optimal Psi(L) with O(log n) resources; it is also proven that the problem is NP-hard for threshold determination of 2 and 3 on unweighted graphs.
Algorithms with Calibrated Machine Learning Predictions
Judy Hanwen Shen (Stanford University), Anders Wikum (Stanford University)
TabularTime SeriesBiomedical Data
🎯 What it does: Introduce the calibration of machine learning predictions into online algorithms, design learning-enhanced algorithms suitable for ski rental and job scheduling, and provide the relationship between theoretical competitive ratio and expected error.
Aligned Multi Objective Optimization
Yonathan Efroni (Meta AI), Karen Ullrich (Meta AI)
Optimization
🎯 What it does: This paper proposes the Aligned Multi-Objective Optimization (AMOO) framework, studying how to enhance convergence speed through Adaptive Weighted Gradient Descent (CAMOO, PAMOO) under the condition that all objectives share the same optimal solution, and provides corresponding theoretical guarantees.
Aligning LLMs by Predicting Preferences from User Writing Samples
Stéphane Aroca-Ouellette (University of Colorado), Katherine Metcalf (Apple)
GenerationRecommendation SystemTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: The PROSE method is proposed to infer preference descriptions from user writing samples through iterative refinement and multi-sample validation, which is then used for LLM generation to achieve more accurate personalized writing.
Aligning Multimodal Representations through an Information Bottleneck
Antonio Almudévar (University of Zaragoza), Alfonso Ortega (University of Zaragoza)
RetrievalRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImageTextMultimodality
🎯 What it does: This paper explores the issue of alignment deficiency in multimodal representation learning and proposes integrating information bottleneck (IB) regularization into contrastive learning loss to suppress modality-specific information and enhance the alignment of the representation space.
Aligning Protein Conformation Ensemble Generation with Physical Feedback
Jiarui Lu (Mila - Quebec AI Institute), Jian Tang (HEC Montreal)
Protein Structure PredictionDiffusion modelBiomedical Data
🎯 What it does: Proposes the Energy-based Alignment (EBA) method, which integrates physical energy feedback into the diffusion model for protein conformation generation, achieving alignment with the Boltzmann distribution and enhancing the physical consistency of the generated conformation set.
Aligning Spoken Dialogue Models from User Interactions
Anne Wu (Cornell University), Alexandre Défossez (Kyutai)
Data SynthesisOptimizationTransformerLarge Language ModelSupervised Fine-TuningMultimodalityAudio
🎯 What it does: In real-time full-duplex voice dialogue systems, a preference contrast dataset is constructed using a large number of real user interaction voice records, directly aligning the multi-stream (text + audio) model offline, thereby improving the model's performance in both content and temporal dimensions.