ICLR 2026 Papers — Page 29
International Conference on Learning Representations · 5356 papers
MATH-Beyond: A Benchmark for RL to Expand Beyond the Base Model
Prasanna Mayilvahanan, Wieland Brendel
Large Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Proposes a new benchmark, MATH-Beyond (MATH-B), specifically designed to evaluate the extension capability of reinforcement learning (RL) reasoning. By filtering out problems that existing benchmark models can solve under a large sampling budget, it constructs a zero-baseline challenging problem set to genuinely assess models' true extension capabilities on unknown problems.
Mathesis: Towards Formal Theorem Proving from Natural Languages
Yu Xuejun (Huawei), Zhenguo Li (Huawei)
Large Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Proposed a complete pipeline (Mathesis) from natural language to formal theorem proving, including automation, verification, and proof steps.
MathFimer: Enhancing Mathematical Reasoning by Expanding Reasoning Steps through Fill-in-the-Middle Task
Yuchen Yan (Zhejiang University), Yueting Zhuang (Zhejiang University)
Data-Centric LearningTransformerSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Proposed and implemented a mathematical reasoning step expansion framework called MathFimer based on the Fill-in-the-Middle (FIM) method, which generates finer-grained reasoning chains using existing CoT data to enhance LLM mathematical reasoning performance.
MATHMO: Automated Mathematical Modeling Through Adaptive Search
Tennison Liu (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
OptimizationTransformerLarge Language ModelTabularTime Series
🎯 What it does: This paper proposes MATHMO, a two-layer adaptive search framework based on large language models (LLMs), for automating the mathematical modeling process. It enables multi-objective exploration across different modeling frameworks, model forms, and algorithm implementations, combined with subjective evaluation.
MathNet: A Global Multimodal Benchmark for Mathematical Reasoning and Retrieval
Shaden Alshammari (Massachusetts Institute Of Technology), Antonio Torralba (Massachusetts Institute Of Technology)
RetrievalLarge Language ModelMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes MathNet, a multimodal, multilingual benchmark covering 47 countries, 17 languages, and over 30K Olympiad-level math problems, encompassing three tasks: solving, retrieval, and retrieval-augmented solving.
MatRIS: Toward Reliable and Efficient Pretrained Machine Learning Interatomic Potentials
Yuanchang Zhou (State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences), Weile Jia (State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences)
Computational EfficiencyGraph Neural NetworkGraphPhysics Related
🎯 What it does: Developed a novel orientation-agnostic MLIP called MatRIS, which employs separable attention to model three-body interactions, achieving high-accuracy predictions for material and molecular properties.
MATRIX: Mask Track Alignment for Interaction-aware Video Generation
Siyoon Jin (KAIST AI), Seungryong Kim (KAIST AI)
GenerationTransformerSupervised Fine-TuningDiffusion modelVideoBenchmark
🎯 What it does: Study and improve the performance of video Diffusion Transformer (DiT) in multi-instance interactive generation, proposing an alignment method based on interaction-dominant layers;
Matting Anything 2: Towards Video Matting for Anything
Chenyi Zhang (Nanjing University), Fang Zhao (Nanjing University)
SegmentationTransformerSupervised Fine-TuningPrompt EngineeringImageVideoBenchmark
🎯 What it does: Proposed Matting Anything 2 (MAM2), a full video matting model that can directly accept various user prompts (points, boxes, masks) and handle diverse objects, including transparent objects.
MAVEN: A Mesh-Aware Volumetric Encoding Network for Simulating 3D Flexible Deformation
Zhe Feng (Peking University), Yunhuai Liu (Peking University)
Graph Neural NetworkMeshPhysics Related
🎯 What it does: Proposes a Mesh-aware Volumetric Encoding Network (MAVEN) for 3D elastic and plastic deformation simulation.
Maximizing Asynchronicity in Event-based Neural Networks
Haiqing Hao, Wenhui Wang (Tsinghua University)
ClassificationObject DetectionAutonomous DrivingConvolutional Neural NetworkTransformerSequential
🎯 What it does: Proposed the EVA framework, which uses an asynchronous encoder with linear attention to update features for events individually, and obtains generalizable event features through self-supervised learning for real-time event camera data processing.
Maximizing Incremental Information Entropy for Contrastive Learning
Jiansong Zhang (Shenzhen University), Linlin Shen (Shenzhen University)
Representation LearningContrastive LearningImage
🎯 What it does: Propose the IE-CL framework, which improves contrastive learning by maximizing incremental information entropy (incremental entropy) and introduces a learnable SAIB transform in the query branch.
MC-Search: Evaluating and Enhancing Multimodal Agentic Search with Structured Long Reasoning Chains
Xuying Ning (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
RetrievalAgentic AIMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the MC-SEARCH benchmark, providing long-chain reasoning data enhanced by multi-modal retrieval.
MCbiF: Measuring Topological Autocorrelation in Multiscale Clusterings via 2-Parameter Persistent Homology
Juni Schindler (Imperial College London), Mauricio Barahona (Imperial College London)
Representation LearningGraph Neural NetworkSequentialBenchmark
🎯 What it does: Proposed the multi-scale clustering dual filter (MCbiF), which encodes non-hierarchical multi-scale clustering sequences into a dual-parameter abstract simplex complex, and obtains Hilbert function features through multi-parameter persistent homology.
MCIF: Multimodal Crosslingual Instruction-Following Benchmark from Scientific Talks
Sara Papi (Fondazione Bruno Kessler), Jan Niehues (Karlsruhe Institute Of Technology)
Large Language ModelVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Created the MCIF benchmark to evaluate cross-lingual multimodal instruction following, covering text, speech, video, and short/long inputs;
mCLM: A Modular Chemical Language Model that Generates Functional and Makeable Molecules
Carl Edwards (University of Illinois Urbana-Champaign), Heng Ji (University of Illinois Urbana-Champaign)
Drug DiscoveryGraph Neural NetworkTransformerLarge Language ModelTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a modular chemical language model, mCLM, capable of bilingual encoding between natural language and chemical modular building blocks, generating functional molecules amenable to automated synthesis.
MCP Security Bench (MSB): Benchmarking Attacks Against Model Context Protocol in LLM Agents
Dongsen Zhang (Beijing University of Posts and Telecommunications), Wenjun Xu (Beijing University of Posts and Telecommunications)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Established the MCP security benchmark MSB, containing 12 categories of attacks, 2000 real attack instances, and evaluated the security of LLM agents in the MCP environment through real tool calls.
MCP-Bench: Benchmarking Tool-Using LLM Agents with Complex Real-World Tasks via MCP Servers
Zhenting Wang (Accenture), Eugene Siow (Accenture)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkFinance Related
🎯 What it does: Developed the MCP-BENCH benchmark, connecting 28 MCP servers (with 250 tools in total), using automatic task synthesis and fuzzy instruction evaluation to assess LLM agents' performance in multi-step, cross-domain tool collaboration and long-range planning.
MCP-SafetyBench: A Benchmark for Safety Evaluation of Large Language Models with Real-World MCP Servers
Xuanjun Zong (East China Normal University), Chao Yang (Shanghai AI Laboratory)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and implemented MCP-SafetyBench, a multi-round, multi-domain safety evaluation benchmark based on real MCP servers;
MCPMark: A Benchmark for Stress-Testing Realistic and Comprehensive MCP Use
Zijian Wu (National University of Singapore), Michael Qizhe Shieh (National University of Singapore)
Large Language ModelAgentic AITextBenchmark
🎯 What it does: Proposed the MCPMark benchmark to evaluate large language models' MCP usage capabilities in multi-tool, multi-step, real-world scenarios;
Mean Estimation from Coarse Data: Characterizations and Efficient Algorithms
Alkis Kalavasis (Yale University), Ziyu Zhu (IMC Trading)
OptimizationComputational EfficiencyTabularFinance Related
🎯 What it does: Studies Gaussian mean estimation under coarse data conditions, provides identifiability criteria for convex partitions, and offers a polynomial-time algorithm
Mean Flow Policy with Instantaneous Velocity Constraint for One-step Action Generation
Guojian Zhan (Tsinghua University), Shengbo Eben Li (Tsinghua University)
GenerationReinforcement LearningFlow-based ModelOrdinary Differential Equation
🎯 What it does: Proposed Mean Velocity Policy (MVP), a single-step action generation strategy based on flow models, and introduced Instantaneous Velocity Constraint (IVC) as a boundary condition to improve learning accuracy.
Mean-Field Neural Differential Equations: A Game-Theoretic Approach to Sequence Prediction
Sungwoo Park (Korea University), Byungseung Kong (Korea University)
OptimizationExplainability and InterpretabilityTime SeriesSequentialElectronic Health RecordsStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposed a continuous sequence predictor based on mean-field neural differential equations (MFPs), which models and generates infinite-order continuous sequences by transforming the continuous time series prediction problem into a mean-field game.
MeanCache: From Instantaneous to Average Velocity for Accelerating Flow Matching Inference
Huanlin Gao (Data Science & Artificial Intelligence Research Institute, China Unicom), Shiguo Lian (Data Science & Artificial Intelligence Research Institute, China Unicom)
GenerationComputational EfficiencyFlow-based ModelRectified FlowImageVideoBenchmark
🎯 What it does: Proposes MeanCache, a training-free caching framework based on average velocity, to accelerate Flow Matching inference and reduce error accumulation under high acceleration ratios.
Measure Twice, Cut Once: A Semantic-Oriented Approach to Video Temporal Localization with Video LLMs
Zongshang Pang (University of Osaka), Yuta Nakashima (University of Osaka)
SegmentationLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningVideoMultimodality
🎯 What it does: Propose a semantics-driven video LLM framework called MeCo, which achieves semantic segmentation and localization of video temporal events through structured token generation, query-focused description, and contrastive learning.
Measurement Score-Based Diffusion Model
Chicago Y. Park (University of Wisconsin Madison), Ulugbek S. Kamilov (University of Wisconsin Madison)
RestorationMixture of ExpertsDiffusion modelScore-based ModelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose a Measurement Score-based Diffusion Model (MSM), which generates complete measurements and solves linear inverse problems by learning partial scores dependent only on downsampling and noise measurements in the measurement domain, without requiring training on clear images.
Measuring and Mitigating Rapport Bias of Large Language Models under Multi-Agent Social Interactions
Maojia Song (Singapore University of Technology and Design), Soujanya Poria (Nanyang Technological University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the KAIROS benchmark to evaluate large language models' exposure, utilization of peer information, and resistance capabilities in multi-agent social interaction scenarios.
Measuring Audio's Impact on Correctness: Audio-Contribution-Aware Post-Training of Large Audio Language Models
Haolin He (Chinese University of Hong Kong), Qiuqiang Kong (Chinese University of Hong Kong)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningMultimodalityChain-of-ThoughtAudio
🎯 What it does: Constructed the AudioMCQ audio multiple-choice question dataset with 571k samples, systematically analyzed the zero-audio contribution phenomenon in LALM, proposed Audio-Contribution Filtering and a weak-strong phased training strategy, achieving new SOTA on multiple audio QA benchmarks.
Measuring LLM Novelty As The Frontier Of Original And High-Quality Output
Vishakh Padmakumar (New York University), He He (New York University)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Assess the novelty of large language models' generated text by combining originality with task-specific quality into a unified metric.
Measuring Physical-World Privacy Awareness of Large Language Models: An Evaluation Benchmark
Xinjie Shen (Georgia Tech), Pan Li (Georgia Tech)
Safty and PrivacyLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the EAPrivacy benchmark to evaluate the privacy awareness of large language models (LLMs) in physical environments.
Measuring the Intrinsic Dimension of Earth Representations
Arjun Rao (University of Colorado Boulder), Esther Rolf (University of Colorado Boulder)
Representation LearningData-Centric LearningVision Language ModelMultimodality
🎯 What it does: This paper studies the intrinsic dimensionality of geographic implicit neural representations (INR) to quantify their information content.
Measuring Uncertainty Calibration
Kamil Ciosek (Spotify), Raphaëlle Bertrand-Lalo (Spotify)
ClassificationImageText
🎯 What it does: Propose two non-asymptotic, distribution-free upper bound estimation methods for strictly quantifying the L1 calibration error of binary classifiers under finite sample settings.
Mechanism of Task-oriented Information Removal in In-context Learning
Hakaze Cho (JAIST), Naoya Inoue (JAIST)
Explainability and InterpretabilityTransformerText
🎯 What it does: By injecting low-rank filters into zero-shot hidden states and leveraging geometric metrics to track the information removal process, it is demonstrated that in few-shot in-context learning (ICL), models achieve task focus by eliminating redundant information in queries, significantly improving prediction accuracy.
Mechanistic Detection and Mitigation of Hallucination in Large Reasoning Models
Zhongxiang Sun (Renmin University of China), Jun Xu (Renmin University of China)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper investigates the phenomenon of reasoning hallucinations in large-scale models, proposing a reasoning score based on model internal representations. This score identifies three categories of hallucination patterns, and based on this, constructs a reasoning hallucination detection framework (RHD) and incorporates a reward shaping mechanism based on the reasoning score in reinforcement learning (GRPO-R) to reduce hallucination rates.
Mechanistic Independence: A Principle for Identifiable Disentangled Representations
Stefan Matthes (Technical University of Munich), Hao Shen (Technical University of Munich)
Representation LearningAuto EncoderContrastive Learning
🎯 What it does: Propose a unified framework centered on mechanism independence for determining and achieving identifiable separable representations;
MedAgent-Pro: Towards Evidence-based Multi-modal Medical Diagnosis via Reasoning Agentic Workflow
Ziyue Wang (National University of Singapore), Yueming Jin (Second Affiliated Hospital Zhejiang University School of Medicine)
Agentic AIVision Language ModelImageTextMultimodalityBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: Propose MedAgent‑Pro, a hierarchical reasoning agent workflow that first formulates disease-level standard diagnostic plans based on retrieval-augmented generation (RAG), then performs step-by-step evidence-driven diagnosis at the patient level using specialized tools.
MedAgentGym: A Scalable Agentic Training Environment for Code-Centric Reasoning in Biomedical Data Science
Ran Xu (Emory University), Wenqi Shi (University of Washington)
Drug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIMultimodalityBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: Developed an executable and scalable medical code reasoning training platform called MedAgentGym, and trained and enhanced the medical coding capabilities of LLM agents on this platform.
MedAraBench: Large-scale Arabic Medical Question Answering Dataset and Benchmark
Mouath Abu Daoud, Farah E. Shamout (Cleveland Clinic Abu Dhabi)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataBenchmark
🎯 What it does: Constructed and released MedAraBench, a large-scale benchmark containing 24,883 Arabic medical multiple-choice questions for evaluating the medical reasoning capabilities of large language models.
MedGMAE: Gaussian Masked Autoencoders for Medical Volumetric Representation Learning
Xueming Fu (University of Science and Technology of China), S Kevin Zhou
ClassificationRestorationSegmentationRepresentation LearningTransformerAuto EncoderContrastive LearningGaussian SplattingBiomedical DataComputed Tomography
🎯 What it does: Leverages a self-supervised pre-trained 3D Gaussian mask autoencoder (MedGMAE) to predict complete 3D Gaussian parameters from sparse visible image patches, and applies them to subsequent segmentation, classification, registration, and reconstruction tasks.
Medical Interpretability and Knowledge Maps of Large Language Models
Razvan Marinescu (Lumos AI), Diego Fajardo V.
Explainability and InterpretabilityTransformerLarge Language ModelBiomedical Data
🎯 What it does: Systematically evaluate and map medical knowledge across five open-source LLMs (Llama3.3-70B, Gemma3-27B, MedGemma-27B, Qwen-32B, GPT-OSS-120B), investigating how age, symptoms, diseases, drugs, and dosage information are stored in model layers; employ four techniques—UMAP projection, gradient-weighted significance, layer ablation, and activation patching—to multidimensionally interpret internal model representations; reveal nonlinear and fragmented age representations, cyclic disease progression, drug clustering by specialty, and activation collapse in Gemma series mid-layers through a unified cross-model, multi-layer analysis framework.
Medical thinking with multiple images
Zonghai Yao (University of Massachusetts Amherst), hong yu
Large Language ModelVision Language ModelMultimodalityBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: Proposed MedThinkVQA—a multi-modal medical question-answering benchmark based on real multi-image cases, incorporating expert-annotated step-by-step diagnostic traces;
MedLesionVQA: A Multimodal Benchmark Emulating Clinical Visual Diagnosis for Body Surface Health
Deli Yu (Xiaohe Team, Bytedance), Haihua Yang (Xiaohe Team, Bytedance)
Large Language ModelPrompt EngineeringVision Language ModelMultimodalityBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes MedLesionVQA—a multimodal question-answering benchmark for skin health, aiming to evaluate the performance of large models in real clinical visual diagnosis processes.
MedVR: Annotation-Free Medical Visual Reasoning via Agentic Reinforcement Learning
Zheng Jiang (Tsinghua University), Minfeng Xu (DAMO Academy, Alibaba Group)
Reinforcement LearningAgentic AIVision Language ModelMultimodalityBiomedical DataBenchmark
🎯 What it does: Achieving unannotated visual reasoning in medical vision-language models by enabling the model to actively use image tools (e.g., zooming) during text reasoning through reinforcement learning to verify reasoning outcomes.
MEGS^{2}: Memory-Efficient Gaussian Splatting via Spherical Gaussians and Unified Pruning
Jiarui Chen (HKUST), Wenping Wang (TAMU)
OptimizationComputational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: Developed the MEGS 2 framework to achieve efficient rendering of 3D Gaussian Splatting by simultaneously compressing the number of 3D Gaussian points and the parameters of each point.
MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents
Zijian Zhou (Singapore-MIT Alliance for Research and Technology Centre), Paul Pu Liang (Massachusetts Institute of Technology)
Computational EfficiencyTransformerReinforcement LearningAgentic AITextSequentialRetrieval-Augmented Generation
🎯 What it does: Propose the MEM1 framework, training language model agents to achieve memory compression and reasoning through a unified internal state in multi-round interactions, maintaining nearly constant context size.
MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent
Hongli Yu (Institute for AI Industry Research, Tsinghua University), Hao Zhou (Institute for AI Industry Research, Tsinghua University)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Train a fixed-length memory module using reinforcement learning, enabling LLMs to read long texts in segments and update memory, thus achieving reasoning on infinitely long texts.
Memba: Membrane-driven Parameter-Efficient Fine-Tuning for Mamba
Donghyun Lee (University of Southern California), Priyadarshini Panda (University of Southern California)
ClassificationRecognitionImageTextBenchmark
🎯 What it does: A parameter-efficient fine-tuning (PEFT) method called Memba, based on membrane potential, is proposed for the Mamba model, specifically designed for the SSM architecture.
Membership Inference Attacks Against Fine-tuned Diffusion Language Models
Yuetian Chen (Purdue University), Ninghui Li (Cisco Research)
Safty and PrivacyAdversarial AttackSupervised Fine-TuningDiffusion modelText
🎯 What it does: This paper investigates the privacy risks of diffusion language models (Diffusion Language Models, DLMs) and proposes a membership inference attack (Membership Inference Attack, MIA) method targeting fine-tuned DLMs.
Membership Privacy Risks of Sharpness Aware Minimization
Young In Kim (Purdue University), Rajiv Khanna (Purdue University)
OptimizationSafty and PrivacyAdversarial AttackImageTabular
🎯 What it does: Studied that Sharpness-Aware Minimization (SAM) enhances generalization performance while amplifying the privacy leakage risk to membership inference attacks (MIA).
Memento: Toward an All-Day Proactive Assistant for Ultra-Long Streaming Video
Hongxiang Jiang (Deepeleph Intelligent Technology), Yan Wang (Deepeleph Intelligent Technology)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the Memento framework to enable active audio-visual language interaction with ultra-long video streams.
MemGen: Weaving Generative Latent Memory for Self-Evolving Agents
Guibin Zhang (National University of Singapore), Shuicheng YAN (National University of Singapore)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose MemGen, a framework that dynamically generates latent memories within LLM agents, intertwining memory and reasoning at the token level.
Memorization Through the Lens of Sample Gradients
Deepak Ravikumar (Purdue University), Kaushik Roy (Purdue University)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposes cumulative sample gradient (CSG) as an efficient proxy for memorization, based on the theoretical relationship between input gradients and learning time;
Memorizing Long-tail Data Can Help Generalization Through Composition
Mo Zhou (University of Washington), Rong Ge (Duke University)
OptimizationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: Studied the synergistic effects of memorization and simple composition in deep learning models under long-tailed data distributions, provided theoretical proofs for linear models, and conducted experimental validation on custom tasks using MNIST and Omniglot.
Memory-Free Continual Learning with Null Space Adaptation for Zero-Shot Vision-Language Models
Yujin Jo (Seoul National University), Taesup Kim (Seoul National University)
Computational EfficiencyMeta LearningVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Proposed NuSA-CL, a memory-free, low-resource continual learning framework for incrementally learning new knowledge in multi-task sequences while maintaining the zero-shot capabilities of pre-trained Vision-Language Models (VLMs).
Memory-Statistics Tradeoff in Continual Learning with Structural Regularization
Haoran Li (Shenzhen University), Vladimir Braverman (Johns Hopkins University)
🎯 What it does: This paper investigates the trade-off between statistical performance and memory complexity in structured regularization (GRCL) within continual learning (CL), focusing on the joint overfitting risk in two-task linear regression under random designs (one-hot, Gaussian).
Memory-T1: Reinforcement Learning for Temporal Reasoning in Multi-session Agents
Yiming Du (Chinese University of Hong Kong), Kam-Fai Wong (Huawei Technologies Co Ltd)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextSequentialRetrieval-Augmented Generation
🎯 What it does: Propose the MEMORY-T1 framework, which uses reinforcement learning to precisely filter time-related memories in multi-session dialogues, thereby enhancing temporal reasoning capabilities in long contexts.
Memory, Benchmark & Robots: A Benchmark for Solving Complex Tasks with Reinforcement Learning
Egor Cherepanov (AXXX), Aleksandr Panov (AXXX)
Robotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningVision-Language-Action ModelDiffusion modelMultimodalitySequentialBenchmark
🎯 What it does: Built and released a unified memory reinforcement learning benchmark called MIkaSA, which includes two environments: MIkaSA-Base and MIkaSA-Robo. The latter contains 32 desktop robot manipulation tasks covering four categories: object memory, spatial memory, sequence memory, and memory capacity.
MemoryVLA: Perceptual-Cognitive Memory in Vision-Language-Action Models for Robotic Manipulation
Hao Shi (Tsinghua University), Gao Huang (Tsinghua University)
Robotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose MemoryVLA, a long-term sequential robot manipulation framework that integrates vision, language, and actions, utilizing working memory and hippocampal-like perceptual-cognitive memory banks to capture non-Markovian temporal dependencies.
MENLO: From Preferences to Proficiency – Evaluating and Modeling Native-like Quality Across 47 Languages
Chenxi Whitehouse (Meta Superintelligence Labs), Denise Diaz (Meta Superintelligence Labs)
GenerationTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper constructs a large-scale multilingual evaluation framework called MENLO, covering localized dialogue quality assessment across 47 languages.
Merge before Forget: A Single LoRA Continual Learning via Continual Merging
Fuli Qiao (Pennsylvania State University), Mehrdad Mahdavi (Pennsylvania State University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a single LoRA continual learning framework SLAO, achieving a shared LoRA through a continuous merging method to avoid catastrophic forgetting.
MergeMix: A Unified Augmentation Paradigm for Visual and Multi-Modal Understanding
Xin Jin (Westlake University), Huan Wang (Westlake University)
ClassificationRecognitionTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: Propose MergeMix, a unified mixed augmentation framework that generates hybrid images through token merge for image classification and alignment training of multimodal large models.
MergePRAG: Orthogonal Merging of Passage-experts for Multi-hop Parametric RAG
Xuebing Liu (Jeonbuk National University), Seung-Hoon Na (Ulsan National Institute of Science and Technology)
RetrievalComputational EfficiencyTransformerMixture of ExpertsTextRetrieval-Augmented Generation
🎯 What it does: Proposed the MERGEPRAG framework, extending Parametric RAG to multi-hop reasoning scenarios and achieving continuous knowledge injection through orthogonal merging and key layer parameterization.
MergeTune: Continued Fine-Tuning of Vision-Language Models
Wenqing Wang (University of Surrey), Josef Kittler (University of Surrey)
ClassificationRetrievalKnowledge DistillationPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Propose a post-hoc continuous fine-tuning (MERGETUNE) method that identifies weights combining knowledge from pre-trained CLIP and fine-tuned models by leveraging linear mode connectivity, thereby recovering pre-training knowledge lost during fine-tuning.
MergOPT: A Merge-Aware Optimizer for Robust Model Merging
Enneng Yang (Sun Yat-sen University), Li Shen (Sun Yat-sen University)
OptimizationSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Propose a fine-tuning optimizer called MergOPT for model fusion, enabling individual expert models to proactively adapt to subsequent merging operations during fine-tuning, thereby enhancing the performance of the merged model.
MesaNet: Sequence Modeling by Locally Optimal Test-Time Training
Johannes von Oswald (Google), Joao Sacramento (Google)
Computational EfficiencyRepresentation LearningTransformerText
🎯 What it does: Proposed a parallelizable Mesa layer that achieves optimal fast weight learning by solving quadratic regularized least squares problems during inference, enabling constant memory and computational cost in autoregressive sequence modeling tasks.
Mesh Splatting for End-to-end Multiview Surface Reconstruction
Ruiqi Zhang (Hong Kong Baptist University), Jie Chen (Hong Kong Baptist University)
OptimizationComputational EfficiencyGaussian SplattingImageMeshBenchmark
🎯 What it does: Propose a method that softens the baseline grid into multi-layer semi-transparent grids and performs voxelization rendering using a differentiable mesh scattering renderer, achieving end-to-end mesh reconstruction and optimization.
MeSH: Memory-as-State-Highways for Recursive Transformers
Chengting Yu (Alibaba Group), Bo Zheng (Alibaba Group)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed and implemented a recursive Transformer architecture named MeSH (Memory-as-State-Highways), which externalizes state management through explicit memory buffers and learnable step-level routers, addressing issues of indiscriminate computation and information overload in traditional recursive Transformers.
Meta-Adaptive Prompt Distillation for Few-Shot Visual Question Answering
Akash Gupta (University of Edinburgh), Mirella Lapata (University of Edinburgh)
Knowledge DistillationMeta LearningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose Meta-Adaptive Prompt Distillation (MAPD), achieving rapid adaptation on LMM for few-shot visual question answering through meta-learning based soft prompt distillation and attention mapper.
Meta-Learning Theory-Informed Inductive Biases using Deep Kernel Gaussian Processes
Bahti Zakirov (Institute of Science and Technology Austria), Gasper Tkacik
Meta LearningImageBiomedical Data
🎯 What it does: Built a Bayesian meta-learning framework that automatically converts normative theories (e.g., efficient coding) into trainable kernel functions usable for data fitting, termed Theory-Informed Kernel (TIK)
Meta-RL Induces Exploration in Language Agents
Yulun Jiang (EPFL), Maria Brbic
Meta LearningLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Propose LAMER—a Meta-RL-based training framework for large language model agents, enabling agents to actively explore and adaptively learn through environmental feedback in multi-round tasks.
Meta-Router: Bridging Gold-standard and Preference-based Evaluations in LLM Routing
Yichi Zhang (Indiana University Bloomington), Chong Wu (University Of Texas Md Anderson Cancer Center)
Data-Centric LearningMeta LearningTextBenchmark
🎯 What it does: Constructed a causal meta-learning based LLM router that jointly trains routing decisions using gold standard and preference data.
Meta-UCF: Unified Task-Conditioned LoRA Generation for Continual Learning in Large Language Models
ShiLin Xiao, Chuangxin Zhao (Ningbo University)
Representation LearningMeta LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose a task-conditioned LoRA generation framework called Meta-UCF based on a hypernetwork, achieving memory-efficient continual learning on large language models (LLMs);
MetaCaptioner: Towards Generalist Visual Captioning with Open-source Suites
Zhenxin Lei (Shanghai AI Laboratory), Gen Luo (Shanghai AI Laboratory)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringVision Language ModelImageVideoTextMultimodality
🎯 What it does: This paper proposes the CapFlow multi-agent collaboration framework for generating high-quality visual descriptions across domains, and builds MetaCaptioner based on this, becoming a general-purpose visual caption model in the open-source field.
MetaEmbed: Scaling Multimodal Retrieval at Test-Time with Flexible Late Interaction
Zilin Xiao (Meta Superintelligence Labs), Vijai Mohan (Meta Superintelligence Labs)
RetrievalRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningMultimodality
🎯 What it does: METAEMBED proposes an expandable multimodal retrieval framework that generates multi-vector Late Interaction embeddings in Vision-Language Models by leveraging learnable Meta Tokens;
MetaMuse: Algorithm Generation via Creative Ideation
Ruiying Ma (University of California Berkeley), Francis Y. Yan (University of Illinois Urbana Champaign)
OptimizationTransformerLarge Language ModelTabularChain-of-Thought
🎯 What it does: Propose the MetaMuse framework, which leverages large language models for algorithm generation and addresses availability bias
MetaSpatial: Reinforcing 3D Spatial Reasoning in VLMs for the Metaverse
Zhenyu Pan (Northwestern University), Han Liu (Northwestern University)
GenerationOptimizationTransformerLarge Language ModelReinforcement LearningVision Language ModelImageText
🎯 What it does: MetaSpatial proposes a reinforcement learning-based framework that utilizes VLM to perform multi-round iterative generation of 3D scene layouts, directly optimizing spatial layouts through a three-tier reward mechanism without requiring post-processing.
MetaVLA: Unified Meta Co-Training for Efficient Embodied Adaptation
Chen Li (Carnegie Mellon University), Marios Savvides
Robotic IntelligenceMeta LearningTransformerSupervised Fine-TuningVision-Language-Action ModelImageTextMultimodality
🎯 What it does: Propose a unified post-training framework called MetaVLA, which integrates multiple tasks and auxiliary tasks into a single model through Context-Aware Meta Co-Training, achieving efficient and scalable Vision-Language-Action (VLA) adaptation.
Metis: Training LLMs with FP4 Quantization
Hengjie Cao (Fudan University), Li Shang (Fudan University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Propose a spectral domain quantization framework, Metis, in LLM training, which splits wide distributions into narrow sub-distributions by leveraging singular value spectral heterogeneity, achieving efficient training in FP4 format.
Metric $k$-clustering using only Weak Comparison Oracles
Rahul Raychaudhury (Duke University), Stavros Sintos (University Of Illinois Chicago)
OptimizationTabular
🎯 What it does: This paper proposes an algorithm to construct an approximate optimal k-clustering core set under the constraint of accessing only a noisy quadruple comparator (R-model);
MHLA: Restoring Expressivity of Linear Attention via Token-Level Multi-Head
Kewei Zhang (Peking University), Daquan Zhou (Peking University)
ClassificationGenerationTransformerImageVideoText
🎯 What it does: Proposed a new linear attention mechanism called Multi-Head Linear Attention (MHLA), which divides tokens into multiple small blocks (heads) along the token dimension, calculates local KV summaries for each block, and then recovers the context collapse problem caused by global aggregation in linear attention by weighting each block with learnable mixing coefficients.
MIAM: Modality Imbalance-Aware Masking for Multimodal Ecological Applications
Robin Zbinden (École Polytechnique Fédérale de Lausanne), Devis Tuia (École Polytechnique Fédérale de Lausanne)
ClassificationTransformerMultimodality
🎯 What it does: Proposed and implemented a dynamic masking strategy called MIAM for ecological multimodal learning to handle missing data and enhance model robustness.
MICLIP: Learning to Interpret Representation in Vision Models
Yingdong Shi (ShanghaiTech University), Kan Ren (ShanghaiTech University)
Explainability and InterpretabilityTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Propose the MICLIP framework, which maps internal units of visual models (such as neurons or SAE features) into the CLIP semantic space through contrastive learning, achieving interpretability and controllability of model internal mechanisms;
Micro-Macro Coupled Koopman Modeling on Graph for Traffic Flow Prediction
Bairan Xiang (Hong Kong University of Science and Technology), Huan Yu (Hong Kong University of Science and Technology)
Autonomous DrivingGraph Neural NetworkMixture of ExpertsGraphTime SeriesSequential
🎯 What it does: Developed a micro-macro coupled Koopman model to achieve unified prediction of vehicle trajectories and traffic flow density.
Micro-Macro Retrieval: Reducing Long-Form Hallucination in Large Language Models
Yujie Feng (Solar System Of Ovb Tencent), Xiao-Ming Wu (Hong Kong Polytechnic University)
RetrievalTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposes Micro-Macro Retrieval (M2R), which first performs macro retrieval in multi-turn reasoning and then conducts micro retrieval during the answer generation phase. Key evidence is stored in an internal warehouse and positioned near the output to reduce hallucinations in long-text generation.
MicroMix: Efficient Mixed-Precision Quantization with Microscaling Formats for Large Language Models
Wenyuan Liu (Tianjin University), Xindian Ma (Tianjin University)
Computational EfficiencyTransformerText
🎯 What it does: Proposed a mixed-precision quantization framework called MicroMix, achieving efficient inference on NVIDIA Blackwell architecture using MXFP4/6/8 formats.
MicroVerse: A Preliminary Exploration Toward a Micro-World Simulation
Rongsheng Wang (Chinese University of Hong Kong Shenzhen), Benyou Wang (Chinese University of Hong Kong Shenzhen)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelAuto EncoderVideoTextBenchmark
🎯 What it does: This paper proposes the concept of Micro-World Simulation and constructs the MicroWorldBench evaluation benchmark, the MicroSim-10K experimental dataset, and a specialized video generation model for micro-scale simulation called MicroVerse;
MIDAS: Multi-Image Dispersion and Semantic Reconstruction for Jailbreaking MLLMs
Yilian Liu (Beijing University of Posts and Telecommunications), Yang Liu (Nanyang Technological University)
Safty and PrivacyPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose the MIDAS multi-graph dispersion and semantic reconstruction framework, leveraging multi-graph game-like visual reasoning and role-driven text guidance to achieve jailbreaking attacks on multimodal large language models.
Midway Network: Learning Representations for Recognition and Motion from Latent Dynamics
Christopher Hoang (New York University), Mengye Ren (New York University)
RecognitionSegmentationTransformerContrastive LearningOptical FlowVideo
🎯 What it does: Designed the Midway Network self-supervised learning framework, which learns visual representations that balance object recognition and motion understanding using natural videos.
MILCO: Learned Sparse Retrieval Across Languages via a Multilingual Connector
Thong Nguyen (University of Amsterdam), Andrew Yates (Johns Hopkins University)
RetrievalTransformerContrastive LearningText
🎯 What it does: Propose MILCO, a sparse retrieval model that maps multilingual queries and documents into a shared English vocabulary space;
MILPnet: A Multi-Scale Architecture with Geometric Feature Sequence Representations for Advancing MILP Problems
Ruobing Wang (Beijing Institute of Technology), Mingzhong Wang (University of the Sunshine Coast)
OptimizationTransformerSequentialBenchmark
🎯 What it does: This paper proposes MILPnet, a multi-scale hybrid attention framework based on sequence modeling, which represents mixed integer linear programming (MILP) problems using geometric feature sequences;
MILR: Improving Multimodal Image Generation via Test-Time Latent Reasoning
Yapeng Mi (University of Science and Technology of China), Qing Li (State Key Laboratory of General Artificial Intelligence, BIGAI)
GenerationTransformerReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Propose MILR, an inference-enhanced method for multimodal image generation that optimizes image and text latent vectors through reinforcement learning in a unified latent space during testing.
MIMIC-Bench: Exploring the User-Like Thinking and Mimicking Capabilities of Multimodal Large Language Models
Jiajie Teng (Shanghai Jiao Tong University), Guangtao Zhai (Shanghai Jiao Tong University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: Constructed MIMIC-Data (over 150k user videos along with their titles, tags, comments, and other metadata), and selected 4,000 high-quality videos from it to create the MIMIC-Bench evaluation system, which includes seven multiple-choice tasks and a comment imitation task for user thinking understanding and user simulation.
MIMIC: Mask-Injected Manipulation Video Generation with Interaction Control
Tianxiao Chen (Zhejiang University), Qi Ye (Zhejiang University)
GenerationConvolutional Neural NetworkPrompt EngineeringVision Language ModelDiffusion modelVideoTextMultimodality
🎯 What it does: Proposed the MIMIC framework, a two-stage image-to-video diffusion model that generates realistic and controllable manipulated videos using reference videos and text descriptions.
MindMix: A Multimodal Foundation Model for Auditory Perception Decoding via Deep Neural-Acoustic Alignment
RUI LIU, KC Tan
Representation LearningTransformerContrastive LearningMultimodalityBiomedical DataAudio
🎯 What it does: Propose MindMix, a multimodal foundation model that utilizes deep neural-acoustic alignment to achieve cross-modal representations between EEG and audio, addressing the limitation of single-modal pre-training in capturing acoustic information.
MindPilot: Closed-loop Visual Stimulation Optimization for Brain Modulation with EEG-guided Diffusion
Dongyang Li (Southern University of Science and Technology), Quanying Liu (Southern University of Science and Technology)
GenerationOptimizationVision Language ModelDiffusion modelImageTime SeriesBiomedical Data
🎯 What it does: Proposed MindPilot, a closed-loop EEG-guided visual stimulus generation framework that continuously optimizes natural images through non-invasive EEG feedback to modulate brain states.
Mini-cluster Guided Long-tailed Deep Clustering
Zhixin Li (Southeast University), Junhui Hou (City University of Hong Kong)
Representation LearningContrastive LearningImage
🎯 What it does: Propose a mini-cluster guided long-tailed deep clustering method called MiniClustering to address the class bias problem in unsupervised long-tailed clustering.
Mini-o3: Scaling Up Reasoning Patterns and Interaction Turns for Visual Search
Xin Lai (ByteDance), Hengshuang Zhao (University of Hong Kong)
RetrievalTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Build the Mini-o3 system to enable vision-language models to perform deep, hierarchical reasoning through multi-round tool calls (e.g., image cropping) and achieve high accuracy in visual search tasks.
Minimax Optimal Adversarial Reinforcement Learning
Yudan Wang (Arizona State University), Shaofeng Zou (Arizona State University)
Reinforcement Learning
🎯 What it does: This paper studies the reinforcement learning problem under adversarial transition kernels, proposing a new algorithm AD-FTRL that achieves sublinear regret bounds in fully adversarial environments.
Minimax Rates for Learning Pairwise Interactions in Attention-Style Models
Shai Zucker (Tel Aviv University), Inbar Seroussi (Tel Aviv University)
OptimizationRepresentation LearningTransformer
🎯 What it does: Analyze the learning convergence rate of interactions between token pairs in a single-layer self-attention model;
Minimax Sample Complexity of Graph Neural Networks: Lower Bounds and Structural Effects
Ahmad Ghasemi (University of Massachusetts Amherst), Hossein Pishro-Nik (University of Massachusetts Amherst)
Representation LearningGraph Neural NetworkGraph
🎯 What it does: Studied the optimal sample complexity of graph neural networks (GNNs) under two learning settings—inductive (graph-level) and transductive (node-level)—and provided corresponding lower bounds.
Minimax-Optimal Aggregation for Density Ratio Estimation
Lukas Gruber (Johannes Kepler University Linz), Werner Zellinger (Johannes Kepler University Linz)
Domain AdaptationOptimizationImageTextTime Series
🎯 What it does: Proposed and implemented a density ratio estimation model that aggregates different hyperparameter settings, ensuring minimax optimal error convergence under unknown smoothness.