ICLR 2026 Papers — Page 3
International Conference on Learning Representations · 5356 papers
AFD-INSTRUCTION: A Comprehensive Antibody Instruction Dataset with Functional Annotations for LLM-Based Understanding and Design
Ling Luo (Xiamen University), Rongshan Yu (Xiamen University)
Data SynthesisDrug DiscoveryProtein Structure PredictionTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextBiomedical Data
🎯 What it does: Constructed and released the first large-scale antibody instruction dataset (AFD-Instruction), aligning antibody sequences with natural language functional annotations. This dataset was used to fine-tune large language models for antibody function understanding and instruction-based antibody design.
AFTER: Mitigating the Object Hallucination of LVLM via Adaptive Factual-Guided Activation Editing
Tianbo Wang (Beihang University), Xianglong Liu (Beihang University)
Explainability and InterpretabilityLarge Language ModelVision Language ModelMultimodality
🎯 What it does: The Adaptive Factual Guidance for Visual-Text Editing (AFTER) method edits the internal activations of large vision-language models (LVLMs) using factual text to alleviate object hallucinations caused by language bias.
Agent Data Protocol: Unifying Datasets for Diverse, Effective Fine-tuning of LLM Agents
Yueqi Song (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
Data-Centric LearningLarge Language ModelSupervised Fine-TuningAgentic AIText
🎯 What it does: Designed and implemented the Agent Data Protocol (ADP), unifying 13 different formats of agent training datasets into a standardized trajectory format, and performed large-scale supervised fine-tuning on multiple agent frameworks (OpenHands, AgentLab, SWE-Agent), significantly improving performance on various benchmarks.
Agent-X: Evaluating Deep Multimodal Reasoning in Vision-Centric Agentic Tasks
Tajamul Ashraf (Mohamed bin Zayed University of Artificial Intelligence), Salman Khan (Mohamed bin Zayed University of Artificial Intelligence)
Large Language ModelAgentic AIVision Language ModelImageVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the Agent-X benchmark to evaluate the ability of vision-centric agents to perform multi-step deep reasoning and tool calling in real multi-modal environments.
AgentFold: Long-Horizon Web Agents with Proactive Context Folding
Rui Ye (Shanghai Jiao Tong University), Yong Jiang (Alibaba Tongyi Lab)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextBenchmark
🎯 What it does: Proposed AgentFold — an architecture that achieves long-term network agents through active context folding;
AgentGym-RL: An Open-Source Framework to Train LLM Agents for Long-Horizon Decision Making via Multi-Turn RL
Zhiheng Xi (Fudan University), Yu-Gang Jiang (Fudan University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AITextChain-of-Thought
🎯 What it does: This paper develops a unified reinforcement learning framework AgentGym-RL for training large language model (LLM) agents to complete multi-round decision tasks, and proposes a phased interaction expansion method called ScalingInter-RL;
Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models
Qizheng Zhang (Stanford University), Kunle Olukotun (SambaNova Systems, Inc.)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextFinance Related
🎯 What it does: Evolutionary construction and self-improvement of context for large language models to enhance Agent performance and domain-specific reasoning;
Agentic Jigsaw Interaction Learning for Enhancing Visual Perception and Reasoning in Vision-Language Models
Yu Zeng (University of Science and TechnolZengogy of China), Feng Zhao (University of Science and TechnolZengogy of China)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision-Language-Action ModelImageTextMultimodality
🎯 What it does: Propose the AGILE framework, treating the puzzle-solving process as an interactive multi-round dialogue with the environment, leveraging VLM to generate executable Python code to perform actions such as Swap, Observe, Crop, Zoom, progressively enhancing visual perception and reasoning capabilities.
Agentic Reinforced Policy Optimization
Guanting Dong (Renmin University of China), Zhicheng Dou (Renmin University of China)
OptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Proposed a reinforcement learning algorithm ARPO for training LLM agents in multi-turn tool usage scenarios.
Agentic Reinforcement Learning with Implicit Step Rewards
Xiaoqian Liu (University of Chinese Academy of Sciences), Junge Zhang (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextMultimodality
🎯 What it does: Propose a framework named iStar, which achieves dense and low-variance credit assignment for sparse rewards in multi-step interactive reinforcement learning (RL) by alternately training an implicit process reward model (PRM) with an LLM policy model;
AgentMath: Empowering Mathematical Reasoning for Large Language Models via Tool-Augmented Agent
Haipeng Luo (Tsinghua University), Yansong Tang (Tsinghua University)
Computational EfficiencyReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextChain-of-Thought
🎯 What it does: Built an mathematical reasoning agent called AgentMath that seamlessly integrates the reasoning capabilities of large language models with the computational precision of external code interpreters.
AgentPO: Enhancing Multi-Agent Collaboration via Reinforcement Learning
Lin Sun (MatrixRobotics), Ning Wu (UAES AI Lab)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Propose the AgentPO framework, which trains lightweight collaborators (Collaborator) using reinforcement learning within a fixed multi-agent topology to optimize collaboration methods, thereby enhancing the overall reasoning performance of the primary executor (Actor).
AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems?
Guibin Zhang, Shuicheng YAN
Data SynthesisAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningAgentic AITextSequential
🎯 What it does: Proposes the AgenTracer framework, which automatically generates annotated multi-agent failure trajectories and trains a lightweight failure localization model, AgenTracer-8B, to help quickly identify root causes in LLM agent systems.
AgentSynth: Scalable Task Generation for Generalist Computer-Use Agents
Jingxu Xie (University of California, Berkeley), Dawn Song (University of California, Berkeley)
Data SynthesisTransformerLarge Language ModelAgentic AISequentialBenchmarkChain-of-Thought
🎯 What it does: Designed and implemented AgentSynth, an LLM-based automated pipeline for generating scalable, realistic, and diverse computer usage tasks and trajectory datasets.
AgilePruner: An Empirical Study of Attention and Diversity for Adaptive Visual Token Pruning in Large Vision-Language Models
Changwoo Baek (Pusan National University), Kyeongbo Kong (Pusan National University)
CompressionComputational EfficiencyVision-Language-Action ModelMultimodalityBenchmark
🎯 What it does: Systematically studied the behavior of visual token pruning in large vision-language models, quantifying the retained feature diversity and assessing its impact on hallucination generation.
Agnostics: Learning to Synthesize Code in Any Programming Language with a Universal Reinforcement Learning Environment
Aleksander Boruch-Gruszecki (Northeastern University), Arjun Guha (Northeastern University)
AI Code AssistantLarge Language ModelReinforcement LearningText
🎯 What it does: Propose a language-agnostic post-training pipeline (Agnostics) that verifies code based solely on the external behavior (I/O) of programs, thereby reducing the need for language-specific engineering in low-resource scenarios.
AI-for-Science Low-code Platform with Bayesian Adversarial Multi-Agent Framework
Zihang Zeng (Fudan University), Xi Chen (Fudan University)
OptimizationAI Code AssistantAgentic AITextBenchmark
🎯 What it does: This paper proposes a low-code platform based on a Bayesian adversarial multi-agent framework, which can automatically generate, verify, and optimize scientific code without relying on a single large language model.
AIRE-Prune: Asymptotic Impulse-Response Energy for State Pruning in State Space Models
Apurba Prasad Padhy (Georgia Institute of Technology), Saibal Mukhopadhyay (Georgia Institute of Technology)
Computational EfficiencySequentialBenchmarkAudio
🎯 What it does: Introduces a post-training structured pruning method called AIRE-Prune, which performs layer-adaptive pruning on deep state-space models based on the infinite impulse response energy of each state.
AirQA: A Comprehensive QA Dataset for AI Research with Instance-Level Evaluation
Tiancheng Huang (Shanghai Jiao Tong University), Kai Yu (Shanghai Jiao Tong University)
Data SynthesisLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Designed and released the AirQA paper QA dataset along with the automated interactive trajectory generation framework EXTRACTOR.
Algorithmic Guarantees for Distilling Supervised and Offline RL Datasets
Aaryan Gupta (Google DeepMind India), Aravindan Raghuveer (Google DeepMind India)
Knowledge DistillationReinforcement LearningTabular
🎯 What it does: Propose a model-free training data distillation algorithm based on loss matching, which can generate extremely small synthetic datasets in supervised learning and offline reinforcement learning, and provides theoretical convergence guarantees.
Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment
Yuyan Bu (Beijing Academy of Artificial Intelligence), Juntao Dai (Beijing Academy of Artificial Intelligence)
Safty and PrivacyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Propose a multilingual consistency loss to enhance the safety behavior of LLMs across different languages.
Align to Misalign: Automatic LLM Jailbreak with Meta-Optimized LLM Judges
Hamin Koo (Yonsei University), Jaehyung Kim (Yonsei University)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Through the meta-optimization framework AMIS, simultaneously evolve jailbreak prompts and scoring templates, significantly improving the jailbreak success rate of large language models.
Align Your Structures: Generating Trajectories with Structure Pretraining for Molecular Dynamics
Aniketh Iyengar (Stanford University), Stefano Ermon (Stanford University)
Drug DiscoveryGraph Neural NetworkDiffusion modelGraphTime Series
🎯 What it does: Proposed a diffusion model (EGINTERPOLATOR) based on structural pre-training and time interpolation, first training a structural generator on large-scale resonator data, then adding a time interpolation module on limited molecular dynamics (MD) trajectory data to achieve high-quality MD trajectory generation.
Align-SAM: Seeking Flatter Minima for Better Cross-Subset Alignment
Van-Anh Nguyen (Monash University), Trung Le (Monash University)
OptimizationImage
🎯 What it does: Proposes a novel method called Align-SAM, aiming to enhance the model's generalization capability by promoting optimization on both the primary subset and auxiliary subset.
Align-Then-stEer: Adapting the Vision-Language Action Models through Unified Latent Guidance
Yang Zhang (Institute of Artificial Intelligence China Telecom), Xuelong Li (Institute of Artificial Intelligence China Telecom)
Robotic IntelligenceVision-Language-Action ModelDiffusion modelScore-based ModelFlow-based ModelAuto EncoderImageTextMultimodality
🎯 What it does: Proposes the Align-Then-Steer (ATE) framework, which constructs a unified latent space to align action distributions across body types and tasks, and achieves efficient VLA adaptation by using a classifier guided by the latent space.
Aligned Agents, Biased Swarm: Measuring Bias Amplification in Multi-Agent Systems
Keyu Li (Shanghai Jiao Tong University), Dequan Wang (Shanghai Jiao Tong University)
Large Language ModelTextBenchmark
🎯 What it does: Systematically study the amplification of bias in multi-agent systems (MAS), constructing and utilizing the Discrim-Eval-Open benchmark to conduct comprehensive evaluations across different roles, topologies, depths, and model heterogeneity.
Aligned Novel View Image and Geometry Synthesis via Cross-modal Attention Instillation
Min-Seop Kwak, Jin-Hwa Kim
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelImagePoint CloudMesh
🎯 What it does: Utilize diffusion models combined with geometric prediction to jointly generate images and geometry aligned with reference images
Aligner, Diagnose Thyself: A Meta-Learning Paradigm for Fusing Intrinsic Feedback in Preference Alignment
Mengyang Li (Tianjin Normal University), Zhong Zhang (Tianjin Normal University)
Meta LearningReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: Propose a preference alignment framework based on model self-diagnosis, which constructs a diagnostic vector using multiple intrinsic feedbacks (preference consistency, learning difficulty, generation confidence) and adaptively assigns weights to each preference sample through meta-learning, thereby enhancing robustness against noisy preference data.
AlignFlow: Improving Flow-based Generative Models with Semi-Discrete Optimal Transport
Lingkai Kong (Georgia Institute of Technology), Huidong Liu (Amazon.com Inc)
GenerationFlow-based ModelImage
🎯 What it does: Propose the AlignFlow method, which achieves explicit alignment between noise and data through semi-discrete optimal transport (SDOT), improving the training of flow-based generative models.
Aligning Collaborative View Recovery and Tensorial Subspace Learning via Latent Representation for Incomplete Multi-View Clustering
Youqing Wang (Beijing University of Chemical Technology), Jipeng Guo (Beijing University of Chemical Technology)
OptimizationRepresentation LearningBenchmark
🎯 What it does: Propose a missing multi-view clustering method (ARSL-IMVC) that unifies view recovery with tensor subspace learning through shared latent representations
Aligning Deep Implicit Preferences by Learning to Reason Defensively
Peiming Li (Tencent), Yang Tang (Tencent)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Proposes the Critique-Driven Reasoning Alignment (CDRA) framework, which utilizes process-level critiques to achieve deep preference understanding and defensive reasoning in LLMs.
Aligning Visual Foundation Encoders to Tokenizers for Diffusion Models
Bowei Chen (University of Washington), Kai Zhang (Adobe Research)
GenerationDiffusion modelAuto EncoderImage
🎯 What it does: Propose the AlignTok method, which constructs a continuous visual tokenizer by performing three-stage alignment on pre-trained visual encoders, thereby enhancing the performance of diffusion models in image generation.
Alignment through Meta-Weighted Online Sampling: Bridging the Gap between Data Generation and Preference Optimization
Junming Yang (Southeast University), Xin Geng (Southeast University)
OptimizationComputational EfficiencyData-Centric LearningMeta LearningReinforcement Learning from Human FeedbackTransformerReinforcement LearningTextBenchmark
🎯 What it does: Propose the MetaAPO framework, which uses a lightweight meta-learner to adaptively assess the alignment gap between offline preference data and online-generated samples during training, dynamically performing online sampling and weighted optimization to enhance LLM alignment efficiency.
Alignment-Enhanced Integration of Connectivity and Spectral Sparsity in Dynamic Sparse Training of LLM
Wenjing Wu (Tsinghua University), Carlo Vittorio Cannistraci (Tsinghua University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a unified framework that combines dynamic sparse connection training with low-rank spectral sparse training, and eliminates the destructive effects of the two branches through an alignment loss, resulting in a new parameter-efficient pretraining method called CHTsL.
Alignment-Weighted DPO: A principled reasoning approach to improve safety alignment
Mengxuan Hu (University of Virginia), Daben Liu (Capital One)
Safty and PrivacyExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Fine-tune LLMs by constructing a long-chain reasoning (CoT) safety dataset and propose a weighted direct preference optimization (AW-DPO) that separately weights reasoning and answering to enhance security against jailbreak attacks.
AlignSep: Temporally-Aligned Video-Queried Sound Separation with Flow Matching
Xize Cheng (Zhejiang University), Zhou Zhao (Zhejiang University)
RestorationTransformerFlow-based ModelAuto EncoderVideoMultimodalityAudio
🎯 What it does: Designed a generative video query audio separation framework called AlignSep based on flow matching, leveraging video temporal information to achieve high-quality target audio extraction.
All Code, No Thought: Language Models Struggle to Reason in Ciphered Language
Shiyuan Guo (Anthropic Fellows Program), Fabien Roger (Anthropic)
Explainability and InterpretabilityTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Investigate the reasoning ability of large language models on encrypted text and assess its potential threat to Chain-of-Thought monitoring.
All Patches Matter, More Patches Better: Enhance AI-Generated Image Detection via Panoptic Patch Learning
Zheng Yang, Xi Li (Zhejiang University)
Anomaly DetectionTransformerDiffusion modelContrastive LearningImage
🎯 What it does: To address the 'few-patch bias' issue in AI-generated image detection, this paper proposes a detection framework based on global patch learning called Panoptic Patch Learning (PPL). By randomly reconstructing image patches during training and introducing patch-level contrastive learning, the model is encouraged to fully utilize synthetic traces from all patches in images, thereby enhancing detection robustness and generalization performance.
All Roads Lead to Likelihood: The Value of Reinforcement Learning in Fine-Tuning
Gokul Swamy (Carnegie Mellon University), Drew Bagnell
GenerationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Compare the effectiveness of online reinforcement learning with reward model (RLHF) versus offline maximum likelihood estimation (MLE) in preference tuning, provide theoretical analysis using information geometry and KL projection, and validate performance differences through experiments on text summarization tasks.
All That Glitters Is Not Gold: Key-Secured 3D Secrets within 3D Gaussian Splatting
Yan Ren (Nanyang Technological University), Adams Wai-Kin Kong (Nanyang Technological University)
Safty and PrivacyVision Language ModelGaussian SplattingPoint CloudMesh
🎯 What it does: Propose the KeySS framework to achieve end-to-end key-locked 3D steganography, jointly optimizing the 3D Gaussian points in the cover scene and the key decoder, maintaining the standard 3D Gaussian Splatting format, supporting multi-secret hiding and anti-erroneous key attacks.
All-day Multi-scenes Lifelong Vision-and-Language Navigation with Tucker Adaptation
Xudong Wang (Shenyang Institute of Automation Chinese Academy of Sciences), Zhi Han (Shenyang Institute of Automation Chinese Academy of Sciences)
Autonomous DrivingMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: Proposed an efficient adaptation framework called AlldayWalker for all-day multi-scenario lifelong vision-language navigation (AML-VLN), which can continuously learn in different scenarios and lighting conditions while avoiding catastrophic forgetting.
ALM-MTA: Front-Door Causal Multi-Touch Attribution Method for Creator-Ecosystem Optimization
Yuguang Liu (Beijing Dajia Internet Information Technology Co., Ltd), Kun Gai (Beijing Dajia Internet Information Technology Co., Ltd)
Recommendation SystemOptimizationExplainability and InterpretabilityContrastive LearningTabularSequential
🎯 What it does: Developed a causal multi-touch attribution framework ALM-MTA based on front-door identification, addressing the challenge of reliable attribution in recommendation systems caused by unlabelled data, unobserved confounders, and large-scale touchpoint spaces.
Almost Bayesian: Dynamics of SGD Through Singular Learning Theory
Max Hennick (University of New Brunswick), Stijn De Baerdemacker (University of New Brunswick)
OptimizationImageTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Studied the long-term dynamics of stochastic gradient descent (SGD) in deep networks, modeling it as a fractional diffusion process on porous media, and explaining the correspondence between the steady-state distribution of SGD and Bayesian posterior through singular learning theory (SLT).
AlphaAgentEvo: Evolution-Oriented Alpha Mining via Self-Evolving Agentic Reinforcement Learning
Ziyi Tang, Liang Lin
TransformerLarge Language ModelReinforcement LearningAgentic AITabularTime SeriesFinance Related
🎯 What it does: This paper proposes AlphaAgentEvo, a self-evolving Agentic Reinforcement Learning framework that achieves continuous evolution of alpha factors from seeds to high-performance through multi-round tool calls and evaluations;
AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement Learning
Yi Zhang (University Of Science And Technology Of China), Xiang Wang
Safty and PrivacyLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Proposes AlphaAlign, a pure reinforcement learning framework that activates the intrinsic safety awareness of LLMs through verifiable safety rewards, supplemented by normalized helpfulness rewards to achieve a balance between safety and practicality;
AlphaBench: Benchmarking Large Language Models in Formulaic Alpha Factor Mining
Haochen Luo (City University of Hong Kong), Chen Liu (City University of Hong Kong)
Large Language ModelTabularTime SeriesBenchmarkFinance RelatedChain-of-Thought
🎯 What it does: Proposed and implemented AlphaBench, the first systematic benchmark for large language models (LLMs) in formulaic alpha factor mining (FAFM), covering three tasks: factor generation, evaluation, and search;
AlphaFlow: Understanding and Improving MeanFlow Models
Huijie Zhang (University of Michigan), Ivan Skorokhodov (University of Michigan)
GenerationTransformerDiffusion modelFlow-based ModelImageVideoOrdinary Differential Equation
🎯 What it does: By deeply analyzing the training objective of MeanFlow and gradient conflicts, this paper proposes the α-Flow framework, achieving progressive training from trajectory flow matching to MeanFlow, thereby improving the quality of few-step generative models trained from scratch.
AlphaSAGE: Structure-Aware Alpha Mining via GFlowNets for Robust Exploration
Binqi Chen (Peking University), Ming Zhang (Peking University)
Graph Neural NetworkReinforcement LearningFlow-based ModelTabularTime SeriesFinance Related
🎯 What it does: Proposes AlphaSAGE, a framework for automatically discovering high-quality, structurally diverse alphas, leveraging structure-aware AST encoding and generative flow networks (GFlowNet) to sample multi-modal alphas, and achieving more robust trading signals through multi-dimensional rewards and dynamic linear combination.
AlphaSteer: Learning Refusal Steering with Principled Null-Space Constraint
Leheng Sheng (National University Of Singapore), Tat-Seng Chua (National University Of Singapore)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Developed an activation-guided method called AlphaSteer, which utilizes a learnable transformation matrix and dynamically steers the activation of malicious prompts toward rejection directions within the zero space of well-activated states, while maintaining the functionality of normal prompts unaffected.
Alternating Diffusion for Proximal Sampling with Zeroth Order Queries
Hirohane Takagi (University of Tokyo), Atsushi Nitanda (Agency for Science, Technology and Research)
OptimizationDiffusion modelScore-based ModelStochastic Differential Equation
🎯 What it does: Propose an approximate proximal sampler that utilizes only the zeroth-order information of the objective function, directly simulating the forward/backward iterations of proximal sampling through Gaussian convolution and reverse diffusion.
Ambig-SWE: Interactive Agents to Overcome Underspecificity in Software Engineering
Sanidhya Vijayvargiya (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
AI Code AssistantLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Built the Ambig-SWE evaluation framework to systematically assess LLMs' capabilities in detecting missing information, clarifying questions, and completing tasks interactively in software engineering scenarios.
AMemGym: Interactive Memory Benchmarking for Assistants in Long-Horizon Conversations
Cheng Jiayang (Hong Kong University of Science and Technology), Xunliang Cai (Meituan)
OptimizationLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose AMEMGYM, an interactive environment for on-policy evaluation and optimization of long-term conversational memory systems.
AMiD: Knowledge Distillation for LLMs with $\alpha$-mixture Assistant Distribution
Donghyeok Shin (Korea Advanced Institute of Science and Technology), Il-chul Moon
Knowledge DistillationLarge Language ModelText
🎯 What it does: Proposed the α-mixture assistant distribution and the AMiD framework for knowledge distillation in large language models, unifying and extending existing assistant distribution and divergence methods.
AMLRIS: Alignment-aware Masked Learning for Referring Image Segmentation
Tongfei Chen (Beihang University), Baochang Zhang (Beihang University)
SegmentationContrastive LearningMultimodality
🎯 What it does: Proposed an Alignment-Aware Masked Learning (AML) training strategy for Referring Image Segmentation (RIS), which enhances model alignment and segmentation performance by computing pixel-level visual-language alignment and masking unreliable pixels.
Amortising Inference and Meta-Learning Priors in Neural Networks
Tommy Rochussen (Helmholtz AI), Vincent Fortuin (Helmholtz AI)
Meta LearningTransformerImageTabularTime SeriesBiomedical Data
🎯 What it does: This paper proposes a novel Bayesian Neural Network Process (BNNP), which treats the weights of Bayesian neural networks as latent variables, uses amortized linear layers to achieve fast variational inference of weights, and views the model as a neural process to realize meta-learning prior and inference on multi-task datasets.
AMPED: Adaptive Multi-objective Projection for balancing Exploration and skill Diversification
Geonwoo Cho (Gwangju Institute of Science and Technology), Sundong Kim (Gwangju Institute of Science and Technology)
Reinforcement LearningContrastive Learning
🎯 What it does: Propose the AMPED method to simultaneously balance exploration and skill diversity during the skill learning phase, and achieve adaptive skill deployment through a skill selector during the fine-tuning phase.
An Agentic Framework with LLMs for Solving Complex Vehicle Routing Problems
Ni Zhang (Singapore Management University), Yew-Soon Ong (Nanyang Technological University)
OptimizationTransformerLarge Language ModelAgentic AIGraphBenchmark
🎯 What it does: Propose a fully automated, external-module-free LLM agent framework called AFL for end-to-end solving complex vehicle routing problems.
An Efficient SE(p)-Invariant Transport Metric Driven by Polar Transport Discrepancy-based Representation
Junyi Lin (Renmin University of China), Cheng Meng (Renmin University of China)
ClassificationGenerationDrug DiscoveryGenerative Adversarial NetworkPoint CloudBiomedical Data
🎯 What it does: Proposed a new SE(p) invariant optimal transport metric called SEINT, and introduced unsupervised, training-free representations named Polar Transport Discrepancy (PTD) and Distance-convoluted PTD (DcPTD), while applying them as regularization terms in molecular generation models.
An efficient, provably optimal algorithm for the 0-1 loss linear classification problem
Xi He (Peking University), Max A Little
ClassificationTabular
🎯 What it does: Proposed a rigorously proven efficient algorithm called ICE that can precisely solve the 0-1 loss linear classification problem
An Ensemble Framework for Unbiased Language Model Watermarking
Yihan Wu (University of Maryland), Heng Huang (University of Maryland)
GenerationSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Propose an ENS (Ensemble) framework that stacks multiple independent unbiased watermark layers to enhance the detectability and robustness of watermarks in text generation models while maintaining unbiasedness
An evolutionary perspective on modes of learning in Transformers
Alexander Ku (Google DeepMind), Stephanie C.Y. Chan (Princeton University)
TransformerContrastive LearningImage
🎯 What it does: Investigate the learning dynamics of Transformer models using in-context learning (ICL) and in-weight learning (IWL) under different levels of environmental predictability, and explain their transition from an evolutionary perspective
An Improved Model-free Decision-estimation Coefficient with Applications in Adversarial MDPs
Haolin Liu (University of Virginia), Julian Zimmert (Google Research)
Reinforcement Learning
🎯 What it does: Proposed an improved model-free decision estimation coefficient (Dig-DEC) for decision-making in adversarial Markov Decision Processes (MDPs), addressing some unresolved issues from prior research.
An Information Theoretic Perspective on Agentic System Design
Shizhe He (Stanford University), Dan Biderman (Stanford University)
CompressionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAgentic AITextBiomedical DataFinance Related
🎯 What it does: Studies the design of compressor-predictor systems from an information theory perspective, proposing the use of mutual information to estimate compression quality and perform rate-distortion analysis, systematically comparing the impact of different compressor and predictor models and scales on downstream performance.
An Information-Theoretic Framework For Optimizing Experimental Design To Distinguish Probabilistic Neural Codes
Po-Chen Kuo (University of Washington), Edgar Y. Walker (University of Washington)
OptimizationBiomedical Data
🎯 What it does: The research objective is to optimize experimental design through an information theory framework to distinguish whether early sensory neural populations encode likelihood functions or posterior distributions.
An Information-Theoretic Parameter-Free Bayesian Framework for Probing Labeled Dependency Trees from Attention Score
Hongxu Liu (Nanyang Technological University), Fangxiang Feng (Beijing University of Posts and Telecommunications)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Propose an untrained network, information-theoretic Bayesian framework (IPBP), which directly reconstructs labeled syntactic dependency trees by estimating the mutual information between attention scores and dependencies.
An Open-Ended Benchmark and Formal Framework for Adjuvant Research with MLLM
yi chen, Cheng-Lin Liu (Chinese Academy of Sciences)
Drug DiscoveryTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the first open-ended question-answering benchmark and formal description framework for adjuvants, and conducted systematic evaluations of various multimodal large language models (MLLMs);
An Optimal Diffusion Approach to Quadratic Rate-Distortion Problems: New Solution and Approximation Methods
Dror Freirich (Technion Israel Institute Of Technology), Nir Weinberger (Technion Israel Institute Of Technology)
CompressionDiffusion modelImageStochastic Differential EquationOrdinary Differential EquationAudio
🎯 What it does: This paper proposes a new framework based on terminal entropy regularized stochastic control (TEC), which solves the rate-distortion (RD) function of continuous sources under mean squared error (MSE) distortion using diffusion processes, and provides a numerical estimation method called R2D2;
An Orthogonal Learner for Individualized Outcomes in Markov Decision Processes
Emil Javurek (LMU Munich), Stefan Feuerriegel (LMU Munich)
Meta LearningReinforcement LearningBenchmark
🎯 What it does: Propose a new DR Q-learner for estimating the Q-function in Markov Decision Processes (MDPs) from observational data, combining double robustness, Neyman-orthogonality, and quasi-likelihood efficiency.
Analysis of approximate linear programming solution to Markov decision problem with log barrier function
Donghwan Lee (Korea Advanced Institute of Science and Technology), Bumgeun Park (Korea Advanced Institute of Science and Technology)
OptimizationReinforcement LearningBenchmark
🎯 What it does: This paper establishes a theoretical framework by analyzing the linear programming (LP) form of Markov decision processes (MDP), and proposes a method using a logarithmic barrier function to solve MDP.
Analytica: Soft Propositional Reasoning for Robust and Scalable LLM-Driven Analysis
Junyan Cheng (Dartmouth College), Peter Chin (Dartmouth College)
Large Language ModelAgentic AITabularFinance RelatedChain-of-Thought
🎯 What it does: Propose a Soft Proposition Reasoning (SPR) framework and implement an LLM-driven analysis agent named Analytica, which recursively decomposes and evaluates uncertain propositions through three stages: analyzer, grounder, and synthesizer.
Analyzing and Evaluating Unbiased Language Model Watermark
Yihan Wu (University of Maryland), Heng Huang (University of Maryland)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed an open-source benchmark called UWBENCH specifically for evaluating unbiased watermarks, including theoretical analysis and experimental evaluation
Analyzing the Training Dynamics of Image Restoration Transformers: A Revisit to Layer Normalization
MinKyu Lee (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)
RestorationSuper ResolutionTransformerImage
🎯 What it does: Analyzes the training dynamics of image restoration Transformers, revealing that traditional LayerNorm causes feature amplitude explosion and sudden drop in channel entropy, leading to the proposal of an i-LN normalization scheme tailored for image restoration tasks.
Anatomy-aware Representation Learning for Medical Ultrasound
Seok-Hwan Oh (Barreleye Inc), Hyeonmin Bae
ClassificationSegmentationKnowledge DistillationRepresentation LearningTransformerAuto EncoderGenerative Adversarial NetworkContrastive LearningImageBiomedical DataUltrasound
🎯 What it does: Developed an anatomy-aware representation learning framework (ARL) for medical ultrasound, achieving adaptive feature extraction for ultrasound images through anatomy-conditioned deformable Transformer (ACDT) and multi-objective self-supervised training (MIM, adversarial loss, knowledge distillation).
Anchor Frame Bridging for Coherent First-Last Frame Video Generation
Xuehan Hou (Peking University), Jie Chen (Harbin Institute of Technology)
GenerationVision Language ModelDiffusion modelVideoText
🎯 What it does: Propose the Anchor Frame Bridging (AFB) method, which significantly enhances semantic continuity between the first and last frames by inserting adaptively selected anchor frames during the video generation process, addressing the issue of semantic degradation in intermediate frames.
Anchored Supervised Fine-Tuning
He Zhu (Southern University of Science and Technology), Guanhua Chen (Shanghai Artificial Intelligence Laboratory)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningTextBiomedical Data
🎯 What it does: Proposed and evaluated an Anchored Supervised Fine-Tuning (ASFT) method based on reward-weighted regression to enhance the generalization of large language models (LLMs) while maintaining the efficiency of supervised fine-tuning (SFT).
AnesSuite: A Comprehensive Benchmark and Dataset Suite for Anesthesiology Reasoning in LLMs
Xiang Feng (Wuhan University), Jing Zhang (Wuhan University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: Proposed AnesSuite, which includes a cross-lingual evaluation benchmark called AnesBench, along with three training datasets (AnesCorpus, AnesQA, AnesR1), and trained the Morpheus series of LLMs as a baseline for anesthesia reasoning.
Angle K-Means
Shenfei Pei (Hangzhou City University), Zengwei Zheng (Hangzhou City University)
OptimizationComputational EfficiencyImage
🎯 What it does: Propose Angle k-means, a parameter-free, accelerated exact k-means algorithm based on angles and distances between samples and centers;
Animal behavioral analysis and neural encoding with transformer-based self-supervised pretraining
Yanchen Wang (Columbia University), Matthew R Whiteway
Pose EstimationTransformerAuto EncoderContrastive LearningVideo
🎯 What it does: Propose the BEAST framework, combining masked autoencoding with temporal contrastive learning, pretraining Vision Transformer on laboratory-specific unlabeled behavioral videos, supporting multiple behavioral neuroscience tasks including neural encoding, pose estimation, and action segmentation.
Animating the Uncaptured: Humanoid Mesh Animation with Video Diffusion Models
Marc Benedí San Millán (Technical University of Munich), Matthias Nießner (Technical University of Munich)
GenerationPose EstimationVision-Language-Action ModelDiffusion modelTextMesh
🎯 What it does: Propose a text-driven human mesh animation method based on video diffusion models, capable of generating 4D animations from static 3D meshes and text prompts;
Anime-Ready: Controllable 3D Anime Character Generation with Body-Aligned Component-Wise Garment Modeling
Jiachen Qian (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
GenerationTransformerMixture of ExpertsVision Language ModelDiffusion modelImageTextMesh
🎯 What it does: Proposes a complete pipeline for generating animatable anime-style 3D characters from text or single images, including the Anime-SMPL skeleton, body-aligned modular clothing generation, and high-resolution texture synthesis.
Annotation-Efficient Honesty Alignment via Confidence Elicitation and Calibration
Shiyu Ni (State Key Laboratory of AI Safety), Xueqi Cheng (State Key Laboratory of AI Safety)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed a two-stage annotation-efficient honest alignment framework called EliCal, which first uses self-consistency unsupervised signals to enable LLMs to express internal confidence, and then calibrates using a small amount of true annotations; simultaneously released a large-scale benchmark called HonestyBench;
Antibody: Strengthening Defense Against Harmful Fine-Tuning for Large Language Models via Attenuating Harmful Gradient Influence
Quoc Minh Nguyen (Monash University), Mehrtash Harandi (Monash University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a defense framework named Antibody, aiming to counteract harmful fine-tuning attacks in the Fine-tuning-as-a-Service (FTaaS) environment, ensuring large language models maintain both safety and task performance when receiving user data.
AntigenLM: Structure-Aware DNA Language Modeling for Influenza
Yue Pei (Computer Network Information Center Chinese Academy Of Sciences), Yu Kang (Beijing Institute Of Genomics Chinese Academy Of Sciences)
ClassificationGenerationTransformerLarge Language ModelSupervised Fine-TuningSequentialBiomedical Data
🎯 What it does: This paper proposes AntigenLM, a generative DNA language model for the whole genome of influenza viruses, used to predict future antigenic variations and perform subtyping.
Antislop: A Comprehensive Framework for Identifying and Eliminating Repetitive Patterns in Language Models
Samuel J Paech (Liquid AI), Ravid Shwartz-Ziv (New York University)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the Antislop framework to address the problem of repeated words and phrases (called slop) in the outputs of large language models (LLMs). The framework includes: 1) Antislop Sampler — which detects and suppresses already-occurred slop during inference through backtracking; 2) an automated pipeline — which generates slop fingerprints and automatically constructs training samples by comparing the model with human baselines; 3) Final Token Preference Optimization (FTPO) — which precisely adjusts the logits of individual tokens during training to enable the model to naturally avoid slop.
Antithetic Noise in Diffusion Models
Jing Jia (Rutgers University), Guanyang Wang (Rutgers University)
GenerationData SynthesisDiffusion modelScore-based ModelImageTextMultimodality
🎯 What it does: Systematically studied the impact of using antithetic noise in diffusion models on generative results, finding that it can produce strong negative correlation and improve uncertainty quantification and diversity
Any-Depth Alignment: Unlocking Innate Safety Alignment of LLMs to Any-Depth
Jiawei Zhang (ByteDance Seed), Xiaojun Xu (ByteDance Seed)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose Any-Depth Alignment (ADA), achieving automatic rejection at any generation depth by re-injecting assistant head safety tokens during inference to counteract depth Prefill attacks.
Any-Order Flexible Length Masked Diffusion
Jaeyeon Kim (Harvard University), Michael Samuel Albergo (Harvard University)
GenerationTransformerDiffusion modelText
🎯 What it does: Proposed FlexMDM, a discrete diffusion model capable of handling variable-length sequences, demonstrating its effectiveness in pre-training and large-scale fine-tuning
Any-step Generation via N-th Order Recursive Consistent Velocity Field Estimation
Peng Sun (Zhejiang University), Tao Lin (Westlake University)
GenerationTransformerDiffusion modelFlow-based ModelAuto EncoderImageTextMultimodalityStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposed the RCGM (Recursive Consistent velocity field Estimation for Generative Modeling) framework, which unifies and generalizes first-order consistency models, MeanFlow, and other few-step generative models, supporting arbitrary-order recursive learning to achieve high-quality image and multi-modal generation in as few as 1-2 steps.
Any-Subgroup Equivariant Networks via Symmetry Breaking
Abhinav Goel, Ningyuan Huang (Flatiron Institute)
Pose EstimationRepresentation LearningGraph Neural NetworkImageGraphSequential
🎯 What it does: Propose a generic framework called Any-Subgroup Equivariant Networks (ASEN), achieving equivariance for any subgroup by introducing spin symmetry breaking in the base model, with the same network capable of handling multiple data types and tasks;
Any-to-Bokeh: Arbitrary-Subject Video Refocusing with Video Diffusion Model
Yang Yang (Zhejiang University), Peng-Tao Jiang (vivo BlueImage Lab)
Image TranslationGenerationDiffusion modelNeural Radiance FieldAuto EncoderVideo
🎯 What it does: Designed a one-time diffusion framework for controllable video bokeh rendering, allowing users to freely set focal planes and bokeh intensity.
AnyBCQ: Hardware Efficient Flexible Binary-Coded Quantization for Multi-Precision LLMs
Gunho Park (NAVER Cloud), Dongsoo Lee (NAVER Cloud)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes AnyBCQ, a multi-precision LLM framework based on binary-coded quantization, which enables dynamic selection of any bit-width for inference within a single model.
AnyTouch 2: General Optical Tactile Representation Learning For Dynamic Tactile Perception
Ruoxuan Feng (Renmin University of China), Di Hu (Renmin University of China)
Representation LearningRobotic IntelligenceVision Language ModelAuto EncoderMultimodality
🎯 What it does: Developed the AnyTouch 2 universal tactile representation learning framework and constructed the large-scale tactile dataset ToucHD, which covers three-tier dynamic levels to achieve hierarchical dynamic tactile perception.
AnyUp: Universal Feature Upsampling
Thomas Wimmer (Max Planck Institute for Informatics), Jan Eric Lenssen (Max Planck Institute for Informatics)
SegmentationDepth EstimationSuper ResolutionConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Proposed a generic feature upsampling method called AnyUp, which can upsample features extracted from any visual encoder during inference;
AP-OOD: Attention Pooling for Out-of- Distribution Detection
Claus Hofmann (Johannes Kepler University), Werner Zellinger (Johannes Kepler University)
Anomaly DetectionTransformerTextAudio
🎯 What it does: Proposed a semi-supervised OOD detection method called AP-OOD based on attention pooling, which can utilize token-level information to detect OOD in natural language text under unsupervised or few abnormal sample conditions.
APC-RL: Exceeding data-driven behavior priors with adaptive policy composition
Finn Rietz (Örebro University), Johannes A. Stork
Robotic IntelligenceReinforcement LearningFlow-based ModelBenchmark
🎯 What it does: Propose Adaptive Policy Composition (APC), which combines multiple data-driven Normalizing Flow behavior priors with prior-free actors through a hierarchical architecture, and achieves robustness and efficient exploration via a learning-free arbitrator and reward-sharing mechanism.
APPLE: Toward General Active Perception via Reinforcement Learning
Tim Schneider (TU Darmstadt), Jan Peters (TU Darmstadt)
OptimizationRobotic IntelligenceTransformerReinforcement LearningImageVideo
🎯 What it does: Developed an active perception framework called APPLE, which leverages reinforcement learning combined with Transformer to jointly train information acquisition strategies and prediction models, achieving adaptive learning across various active perception tasks.
APT: Towards Universal Scene Graph Generation via Plug-in Adaptive Prompt Tuning
Ruikun Luo (National Engineering Research Center for Big Data Technology and System), Xiaoyu Xia (Royal Melbourne Institute of Technology)
GenerationTransformerPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Proposes a pluggable Adaptive Prompt Tuning (APT) module that converts frozen language model semantic features into context-aware dynamic representations to enhance scene graph generation.
AQER: A Scalable and Efficient Data Loader for Digital Quantum Computers
Kaining Zhang (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
OptimizationComputational EfficiencyImageTextPhysics Related
🎯 What it does: Propose a unified approximate quantum loading framework, derive information-theoretic entanglement bounds, and design scalable AQER loading methods based on maximum entanglement reduction, enabling efficient loading of classical and quantum data.
AQuA: Toward Strategic Response Generation for Ambiguous Visual Questions
Jihyoung Jang (POSTECH), Hyounghun Kim (POSTECH)
RecognitionGenerationLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: This study addresses the ambiguity issue in Visual Question Answering (VQA) by proposing the AQUA dataset, and trains Vision-Language Models (VLMs) through supervised fine-tuning (SFT) and reinforcement learning (GRPO) to achieve strategic answering, enabling the model to select direct answers, infer intentions, list alternative answers, or request clarification based on the ambiguity level.
Arbitrary Generative Video Interpolation
Guozhen Zhang (Nanjing University), Limin Wang (Nanjing University)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: Achieved video frame interpolation at arbitrary timestamps, enabling the generation of intermediate frames at any moment.