ICLR 2026 Papers — Page 31
International Conference on Learning Representations · 5356 papers
Motion-R1: Enhancing Motion Generation with Decomposed Chain-of-Thought and RL Binding
Runqi Ouyang (Institute of Automation Chinese Academy of Sciences), Xingang Wang (Institute of Automation Chinese Academy of Sciences)
GenerationTransformerLarge Language ModelReinforcement LearningVision-Language-Action ModelAuto EncoderTextSequentialBenchmarkChain-of-Thought
🎯 What it does: Propose the Motion-R1 framework, combining a decomposed Chain-of-Thought data engine and RL Binding strategy to achieve high-quality text-to-action generation.
MotionGPT3: Human Motion as a Second Modality
Bingfan Zhu (Zhejiang University), Xin Chen (ByteDance)
GenerationRetrievalTransformerLarge Language ModelDiffusion modelAuto EncoderTextMultimodality
🎯 What it does: Propose MotionGPT3, a dual-stream motion-language model that unifies the understanding and generation of human motion.
MotionSight: Boosting Fine-Grained Motion Understanding in Multimodal LLMs
Yipeng Du, Ying Tai
RecognitionObject DetectionObject TrackingTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextBenchmark
🎯 What it does: Propose MotionSight, a zero-training visual prompting method that enhances the understanding of fine-grained video motion in multimodal large language models by leveraging object spotlights and motion blur separately.
MotionStream: Real-Time Video Generation with Interactive Motion Controls
Joonghyuk Shin (Seoul National University), Xun Huang (Adobe Research)
GenerationComputational EfficiencyTransformerVision Language ModelDiffusion modelScore-based ModelAuto EncoderVideoText
🎯 What it does: Developed a real-time video generation framework called MotionStream, capable of streaming infinitely long videos at 29 FPS on a single GPU based on user-drawn motion trajectories or camera controls.
MotionWeaver: Holistic 4D-Anchored Framework for Multi-Humanoid Image Animation
Xirui Hu (Xi'an Jiaotong University), Weizhan Zhang (Xi'an Jiaotong University)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: Proposed an end-to-end framework called MotionWeaver for multi-character humanoid image animation, capable of achieving high-quality video synthesis in scenarios involving multiple characters and complex pose interactions.
Moving Beyond Diffusion: Hierarchy-to-Hierarchy Autoregression for fMRI-to-Image Reconstruction
Xu Zhang (Zhejiang University), Yi Yang (Zhejiang University)
GenerationTransformerImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose MindHier, a hierarchical recursive fMRI-to-image reconstruction framework that progressively generates images using multi-level brain signal features;
Moving Beyond Medical Exams: A Clinician-Annotated Fairness Dataset of Real-World Tasks and Ambiguity in Mental Healthcare
Max Lamparth (Stanford University), Colleen Waickman (Ohio State University)
TransformerLarge Language ModelTextBiomedical DataBenchmark
🎯 What it does: Constructed MENTAT, a multiple-choice assessment dataset manually designed and annotated by psychiatrists, covering real clinical tasks such as diagnosis, treatment, monitoring, triage, and documentation, and evaluated the impact of language models on healthcare fairness and decision accuracy by replacing variable information such as gender, age, and race.
mR3: Multilingual Rubric-Agnostic Reward Reasoning Models
David Anugraha (Stanford University), Genta Indra Winata (Capital One)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposed a multilingual, rating-scale-independent reward reasoning model called MR3, supporting evaluation and reasoning in 72 languages;
MRAD: Zero-Shot Anomaly Detection with Memory-Driven Retrieval
Chaoran Xu (Chinese Academy of Sciences), Zhengtao Zhang (Chinese Academy of Sciences)
Anomaly DetectionPrompt EngineeringVision Language ModelImageBiomedical Data
🎯 What it does: Propose a memory retrieval-based zero-shot anomaly detection framework, MRAD, which directly utilizes frozen CLIP features and a two-layer feature-label memory bank for similarity retrieval, enabling both image-level classification and pixel-level segmentation.
MRMR: A Realistic and Expert-Level Multidisciplinary Benchmark for Reasoning-Intensive Multimodal Retrieval
Siyue Zhang (Nanyang Technological University), Chen Zhao (VinUniversity)
RetrievalLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Constructed MRMR — a retrieval benchmark containing 1,435 expert-level, multidisciplinary, reasoning-intensive, and text-image interleaved query-document pairs;
MrRoPE: Mixed-radix Rotary Position Embedding
Qingyuan Tian (Shanghai Jiao Tong University), Rui Wang (Shanghai Jiao Tong University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Propose a mixed radix-based RoPE extension framework called MrRoPE, and implement two training-agnostic schemes, MrRoPE-Uni and MrRoPE-Pro.
MSCR: Exploring the Vulnerability of LLMs’ Mathematical Reasoning Abilities Using Multi-Source Candidate Replacement
Zhishen Sun (Xi'an Jiaotong University), Haishan Ye (Xi'an Jiaotong University)
Adversarial AttackTransformerPrompt EngineeringText
🎯 What it does: Propose an automated adversarial attack method called MSCR based on multi-source candidate replacement, specifically designed to evaluate the vulnerability of large language models (LLMs) in mathematical reasoning tasks;
MT-DAO: Multi-Timescale Distributed Adaptive Optimizers with Local Updates
Alex Iacob (University of Cambridge), Nicholas D. Lane (University of Cambridge)
OptimizationComputational EfficiencyLarge Language ModelText
🎯 What it does: Developed a multi-timescale distributed adaptive optimizer, MT-DAO, which integrates local updates to achieve low-communication-frequency training and matches or exceeds the performance of synchronous DDP on large language models.
MTVCraft: Tokenizing 4D Motion for Arbitrary Character Animation
Yanbo Ding (Chinese Academy of Sciences), Xuelong Li (China Telecom)
GenerationTransformerDiffusion modelAuto EncoderImageVideoText
🎯 What it does: Propose the MTVCraft framework, which directly quantizes raw 4D motion sequences (3D joint coordinates over time) into discrete 4D motion tokens and drives arbitrary character image animation using a 4D motion attention-based Diffusion Transformer (MV-DiT).
Much Ado About Noising: Dispelling the Myths of Generative Robotic Control
Chaoyi Pan (Toyota Research Institute), Max Simchowitz (Carnegie Mellon University)
Robotic IntelligenceFlow-based ModelMultimodalityPoint Cloud
🎯 What it does: This paper evaluates the advantages of Generative Control Policies (GCP) on behavior cloning benchmarks and identifies their true driving factors.
Multi-Action Self-Improvement For Neural Combinatorial Optimization
Laurin Luttmann (Leuphana University Lüneburg), Lin Xie (Brandenburg University of Technology)
OptimizationTransformerGraphBenchmark
🎯 What it does: Propose the MACSIM framework, extending self-improvement to multi-agent joint actions by generating task assignments for all agents through a single forward pass and using set prediction loss to enhance sample efficiency and coordination.
Multi-agent Coordination via Flow Matching
Dongsu Lee (University of Texas at Austin), Amy Zhang (University of Texas at Austin)
Knowledge DistillationReinforcement LearningFlow-based ModelBenchmarkOrdinary Differential Equation
🎯 What it does: Efficient offline multi-agent reinforcement learning (MARL) is achieved by first learning a multi-agent joint policy through flow matching, then distilling it into a single-step decentralized policy via behavioral cloning and Q-guidance.
Multi-Agent Debate with Memory Masking
Hongduan Tian (TMLR Group Hong Kong Baptist University), Bo Han (TMLR Group Hong Kong Baptist University)
TransformerAgentic AITextChain-of-Thought
🎯 What it does: Propose a multi-agent debate framework called MAD-M2, which enhances the reasoning robustness of multi-agent debates by actively evaluating and masking erroneous memories generated in the previous round.
Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies
Han Zhou (Google), Sercan O Arik
OptimizationNeural Architecture SearchTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: Proposed Multi-Agent System Search (MASS), a multi-stage automated design framework for constructing efficient multi-agent systems on large language models, including block-level prompt optimization, topology search, and global prompt optimization.
Multi-Agent Guided Policy Optimization
Yueheng Li, Zongqing Lu (Peking University)
OptimizationTransformerReinforcement LearningBenchmark
🎯 What it does: Proposed the Multi-Agent Guided Policy Optimization (MAGPO) framework, which employs an autoregressive centralized guiding policy to guide decentralized learners while ensuring alignment between the two, achieving deployable collaborative multi-agent learning.
Multi-Condition Conformal Selection
Qingyang Hao (Southern University of Science and Technology), Hongxin Wei (Southern University of Science and Technology)
Convolutional Neural NetworkTransformerImageTextPoint Cloud
🎯 What it does: Propose a multi-condition adaptive composite selection method (MCCS) to achieve precise sample selection within target intervals and control the false discovery rate (FDR).
Multi-Domain Riemannian Graph Gluing for Building Graph Foundation Models
Li Sun (Beijing University of Posts and Telecommunications), Philip S. Yu (University of Illinois Chicago)
Domain AdaptationRepresentation LearningGraph Neural NetworkMixture of ExpertsContrastive LearningGraph
🎯 What it does: Proposed a Riemannian geometry-based multi-domain graph pre-training framework called GRAPHGLUE, which integrates graph data from different domains into a smooth unified Riemannian manifold and enables cross-domain knowledge transfer through this manifold.
Multi-Feature Quantized Self-Attention for Fair Large Language Models
Jaeil Park (Yonsei University), Sung-Bae Cho (Yonsei University)
Safty and PrivacyRepresentation LearningTransformerLarge Language ModelAuto EncoderGenerative Adversarial NetworkTextBenchmark
🎯 What it does: Proposes a Multi-Feature Quantized Adversarial Regularization (MQAR), directly inserting quantization bottlenecks and adversarial autoencoders into frozen large language models (BERT, T5, GPT-Neo, Mixtral, LLaMA 3.2) to eliminate attention layer biases caused by multiple sensitive attributes (e.g., gender, race).
Multi-Head Low-Rank Attention
Songtao Liu (Pennsylvania State University), Yue Guo (University of California, Los Angeles)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed a multi-head low-rank attention (MLRA) mechanism that achieves efficient 4-way tensor parallel decoding in long-context reasoning.
Multi-LCB: Extending LiveCodeBench to Multiple Programming Languages
Maria Ivanova (Yandex School of Data Analysis Applied AI Institute), Dmitrii Babaev (GigaCode)
GenerationAI Code AssistantPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the Multi-LCB benchmark, extending LiveCodeBench to 12 programming languages, enabling continuous, pollution-aware multilingual code generation evaluation.
Multi-LLM Adaptive Conformal Inference for Reliable LLM Response
Kangjun Noh (Yonsei University), Kyungwoo Song (KAIST)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyLarge Language ModelText
🎯 What it does: Propose a synthetic reasoning method called MACI that combines multi-model fusion with group conditional calibration, aiming to ensure the authenticity of LLM responses in high-risk domains and achieve group conditional coverage guarantees.
Multi-Marginal Flow Matching with Adversarially Learnt Interpolants
Oskar Kviman (KTH), Nikolay Malkin (University of Edinburgh)
GenerationFlow-based ModelGenerative Adversarial NetworkBiomedical Data
🎯 What it does: Learned interpolation curves in multi-edge flow matching and proposed the ALI-CFM method
Multi-modal Data Spectrum: Multi-modal Datasets are Multi-dimensional
Divyam Madaan (New York University), Sumit Chopra (New York University)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningLarge Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Conduct large-scale experiments on 23 visual question answering benchmarks to systematically measure the extent to which multimodal models rely on unimodal (image, text) and interactive modalities;
Multi-Object System Identification from Videos
Chunjiang Liu (Carnegie Mellon University), Yizhou Zhao (Carnegie Mellon University)
Gaussian SplattingVideoPhysics Related
🎯 What it does: Proposed a multi-object system identification framework MOSIV based on multi-view videos, which can simultaneously reconstruct the four-dimensional geometry, motion, and continuous physical parameters of objects, achieving digital twin and long-term prediction.
Multi-objective Large Language Model Alignment with Hierarchical Experts
Zhuo Li (Harbin Institute of Technology), Jing Li (Harbin Institute of Technology)
OptimizationLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: Propose the HoE (Hierarchical Mixture-of-Experts) framework to achieve multi-objective LLM alignment, supporting arbitrary user preference weights;
Multi-ReduNet: Interpretable Class-Wise Decomposition of ReduNet
Fengrong Li (National University of Singapore), Delin Chu (National University of Singapore)
ClassificationExplainability and InterpretabilityComputational EfficiencyImageText
🎯 What it does: Propose two interpretable ReduNet extensions, Multi-ReduNet and Multi-ReduNet-LastNorm, achieving feature learning through class-level decomposition under under-sampled scenarios where m≪d
Multi-Scale Diffusion-Guided Graph Learning with Power-Smoothing Random Walk Contrast for Multi-View Clustering
Feiyang Chen (Hebei Normal University), Zhibin Gu (Hebei Normal University)
Representation LearningGraph Neural NetworkDiffusion modelContrastive LearningImageTextMultimodalityAudio
🎯 What it does: Propose a multi-view clustering framework (MANGO) based on multi-scale diffusion and power-smoothed random walk contrastive learning, which jointly learns multi-view representations and constructs a refined graph structure for clustering through four modules: self-expression, view consistency, contrastive learning, and adaptive diffusion.
Multi-Scale Hypergraph Meets LLMs: Aligning Large Language Models for Time Series Analysis
Zongjiang Shang (Zhejiang University), Ling Chen (Zhejiang University)
Graph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTime Series
🎯 What it does: Proposed the MSH-LLM framework, which utilizes multiscale hypergraphs and frozen large language models to achieve cross-modal alignment in time series analysis;
Multi-state Protein Sequence Design with DynamicMPNN
Alex Abrudan (University of Cambridge), Tuomas Knowles
Protein Structure PredictionGraph Neural NetworkBiomedical Data
🎯 What it does: Designed an inverse folding model called DynamicMPNN that can generate sequence compatible with multiple conformations for multi-state proteins.
Multi-Subspace Multi-Modal Modeling for Diffusion Models: Estimation, Convergence and Mixture of Experts
Ruofeng Yang (Shanghai Jiao Tong University), Shuai Li (Shanghai Jiao Tong University)
GenerationMixture of ExpertsDiffusion modelAuto EncoderImage
🎯 What it does: Proposed a hybrid low-rank mixture of Gaussians (MoLR-MoG) model to capture multi-subspace multimodal structures in high-dimensional data such as images, and constructed a Mixture-of-Experts (MoE) nonlinear score function based on this model, conducting analysis on estimation error and optimization convergence of diffusion models.
Multi-Synaptic Cooperation: A Bio-Inspired Framework for Robust and Scalable Continual Learning
Penghui Li (Tianjin University), Qiang Yu (Tianjin University)
Computational EfficiencyRepresentation LearningSpiking Neural NetworkImage
🎯 What it does: Proposed a multi-synapse collaborative network (MSCN), which achieves continuous learning with a fixed network structure by equipping each connection with multiple plastic synapses and modulating synapse updates based on local synaptic activity (accessibility traces).
Multi-turn Evaluation of Anthropomorphic Behaviours in Large Language Models
Lujain Ibrahim (University of Oxford), Laura Weidinger (Google DeepMind)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Design and implement AnthroBench, a multi-turn dialogue automated evaluation framework to quantify anthropomorphic behaviors in large language models (LLMs).
Multi-View Encoders for Performance Prediction in LLM-Based Agentic Workflows
Patara Trirat (DeepAuto.ai), Sung Ju Hwang (KAIST)
OptimizationGraph Neural NetworkAgentic AIContrastive LearningTextMultimodalityGraphBenchmark
🎯 What it does: Propose Agentic Predictor, a lightweight predictor that rapidly estimates the performance of LLM-based agent workflows without expensive LLM calls.
Multifidelity Simulation-based Inference for Computationally Expensive Simulators
Anastasia N. Krouglova (University of Tubingen), Pedro J. Goncalves (KU Leuven)
OptimizationComputational EfficiencyFlow-based ModelBenchmark
🎯 What it does: This paper proposes a multi-level (multi-precision) simulation inference method, leveraging pre-training on low-cost low-precision simulators to accelerate parameter posterior inference on high-cost high-precision simulators.
Multilevel Control Functional
Kaiyu Li (COMAC Shanghai Aircraft Design and Research Institute), Zhuo Sun (Shanghai University of Finance and Economics)
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper proposes a novel multi-level control function (MLCF) method to efficiently estimate intractable integrals with low variance under a multi-level approximation structure.
Multilingual Routing in Mixture-of-Experts
Lucas Bandarkar (University of California, Los Angeles), Nanyun Peng (University of California, Los Angeles)
Mixture of ExpertsText
🎯 What it does: This paper reveals that Mixture of Experts (MoE) models tend to route language-specific experts in early/late layers, while exhibiting cross-lingual shared expert routing in intermediate layers by analyzing sparse routing patterns under multilingual data; further, during inference, interventions are applied to intermediate layer routing to encourage activation of experts aligned with English task experts, achieving 1-2% performance improvement on multilingual math and medical tasks.
MultiMat: Multimodal Program Synthesis for Procedural Materials using Large Multimodal Models
Jonas Belouadi (University of Mannheim), Adrien Kaiser (Adobe Research)
GenerationData SynthesisLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This paper proposes MultiMat, which introduces visual feedback using large multimodal models to achieve visual node graph synthesis during procedural material generation.
Multimodal Aligned Semantic Knowledge for Unpaired Image-text Matching
Laiguo Yin (Shandong University), Lizhen Cui (Shandong University)
RetrievalVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper proposes Multimodal Aligned Semantic Knowledge (MASK) for unpaired图文 matching, bridging words and visual prototypes through word vectors, and suppressing distribution variance via prototype consistency contrastive learning;
Multimodal Classification via Total Correlation Maximization
Feng Yu (Nanjing University of Science and Technology), Jianfeng Lu (Nanjing University of Science and Technology)
ClassificationMultimodality
🎯 What it does: Propose a hyperparameter-free TCMax loss that avoids modality competition and enhances cross-modal collaborative learning by maximizing the total correlation between multimodal features and labels
Multimodal Dataset Distillation Made Simple by Prototype-Guided Data Synthesis
Junhyeok Choi (Pohang University of Science and Technology), Minwoo Chae (Pohang University of Science and Technology)
Data SynthesisRetrievalKnowledge DistillationTransformerVision Language ModelDiffusion modelContrastive LearningMultimodality
🎯 What it does: Developed a learning-free multi-modal dataset distillation framework called PDS, which creates small-scale but information-rich multi-modal datasets by generating synthetic samples that conform to image-text prototypes using CLIP embeddings and unCLIP decoders.
Multimodal Dataset Distillation via Phased Teacher Models
Shengbin Guo (Harbin Institute of Technology), Zhuotao Tian (Harbin Institute of Technology)
Data SynthesisKnowledge DistillationConvolutional Neural NetworkTransformerMultimodality
🎯 What it does: This paper proposes a multimodal dataset distillation framework called PTM-ST based on phased teacher models and shortcut trajectories, which can extract knowledge from teacher models in different training stages and generate higher quality synthetic datasets.
Multimodal LLM-assisted Evolutionary Search for Programmatic Control Policies
Qinglong Hu (City University of Hong Kong), Qingfu Zhang (City University of Hong Kong)
Explainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision-Language-Action ModelMultimodality
🎯 What it does: Proposed and implemented a framework called MLES that combines multimodal large language models with evolutionary search to directly evolve executable and interpretable procedural control strategies through environmental interaction.
Multimodal Policy Internalization for Conversational Agents
Zhenhailong Wang (University of Illinois Urbana-Champaign), Ruhi Sarikaya (Amazon)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVision-Language-Action ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes the Multimodal Policy Internalization (MPI) task, aiming to enable large-scale multimodal models to automatically follow complex business policies during inference without requiring policies to be provided as context during inference;
Multimodal Prompt Optimization: Why Not Leverage Multiple Modalities for MLLMs
Yumin Choi (KAIST), Sung Ju Hwang (KAIST)
OptimizationTransformerLarge Language ModelPrompt EngineeringImageVideoTextMultimodality
🎯 What it does: Propose Multimodal Prompt Optimization (MPO) and design the MPO framework, which automatically jointly optimizes text and non-text prompts on MLLM to enhance the performance of multimodal tasks.
MULTIMODALITY AS SUPERVISION: SELF-SUPERVISED SPECIALIZATION TO THE TEST ENVIRONMENT VIA MULTIMODALITY
Kunal Pratap Singh (Swiss Federal Institute of Technology Lausanne), Amir Zamir (Swiss Federal Institute of Technology Lausanne)
RecognitionObject DetectionSegmentationTransformerImageMultimodality
🎯 What it does: Collect multimodal (RGB, depth, normals, Canny edges, etc.) data in the test environment and build a dedicated visual model using cross-modal self-supervised pre-training.
Multiplayer Nash Preference Optimization
Fang Wu (Stanford University), Yejin Choi (Stanford University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: Propose a framework called MNPO that extends Nash learning in RLHF from two-player to multi-player settings, and theoretically analyze multi-player equilibrium, dual gaps, etc.; simultaneously design algorithms for time-related opponent mixing and heterogeneous preference source adaptation;
Multiple Token Divergence: Measuring and Steering In-Context Computation Density
Vincent Herrmann (Swiss AI Lab IDSIA USI SUPSI), Jürgen Schmidhuber (Swiss AI Lab IDSIA USI SUPSI)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes Multiple Token Divergence (MTD) as a metric to measure the computational investment of language models in context, and introduces a Divergence Steering decoding method that can dynamically adjust the computational density of generated text based on MTD.
Multiple-Prediction-Powered Inference
Charlie Cowen-Breen (Massachusetts Institute of Technology), Adam Fisch (Google DeepMind)
OptimizationComputational EfficiencyTextBenchmark
🎯 What it does: Propose the MultiPPI framework, which efficiently estimates expectations under budget constraints by leveraging multiple predictors with varying costs, automatically allocates sampling resources, and applies linear weighting;
Multiplicative Diffusion Models: Beyond Gaussian Latents
Robert Gruhlke (Freie Universitaet Berlin), Merveille Talla (INRAE)
GenerationData SynthesisDiffusion modelScore-based ModelImageTabularPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a conservative diffusion model based on multiplicative noise (MSGM), which achieves a non-Gaussian latent space by maintaining the data norm distribution under random rotations, enabling better reproduction of extreme events.
Multiverse Mechanica: A Testbed for Learning Game Mechanics via Counterfactual Worlds
Robert Ness, Lars Kunze
Explainability and InterpretabilityData-Centric LearningDiffusion modelContrastive LearningWorld ModelImageVideoBenchmark
🎯 What it does: Developed the Multiverse Mechanica game testing platform, formalizing game mechanics as causal counterfactual inference, followed by causal consistency fine-tuning of the world model on this platform.
Muon Outperforms Adam in Tail-End Associative Memory Learning
Shuche Wang (National University of Singapore), Vincent Y. F. Tan (National University of Singapore)
OptimizationTransformerLarge Language ModelText
🎯 What it does: The paper investigates the advantages of the Muon optimizer over Adam in training large language models (LLMs), particularly for the associative memory parameters of Transformers (VO, FFN);
MuonBP: Faster Muon via Block-Periodic Orthogonalization
Ahmed Khaled (Princeton University), Youngsuk Park (Amazon Web Services)
OptimizationText
🎯 What it does: Proposes the MuonBP algorithm: Introduces block-periodic orthogonalization on top of the Muon optimizer, performing local orthogonalization of the gradient matrix independently on each device and conducting global orthogonalization periodically to reduce communication overhead while maintaining data efficiency.
Musculoskeletal simulation of limb movement biomechanics in Drosophila melanogaster
Pembe Gizem Ozdil (Harvard University), Pavan Ramdya (EPFL)
OptimizationReinforcement LearningImageVideoBiomedical DataComputed Tomography
🎯 What it does: Constructed the first 3D kinematic simulation of the fruit fly's forelimb based on the Hill-type muscle model, and implemented muscle-driven behavior replay in OpenSim and MuJoCo.
MUSE: Model-Agnostic Tabular Watermarking via Multi-Sample Selection
Liancheng Fang (University of Illinois Chicago), Philip S. Yu (University of Illinois Chicago)
GenerationSafty and PrivacyScore-based ModelTabular
🎯 What it does: Proposes a table generation model watermarking method called MUSE based on multi-sample selection, which utilizes multiple sampling and selects the highest-scoring samples to embed watermarks, achieving watermark embedding and detection without inversion.
Music Flamingo: Scaling Music Understanding in Audio Language Models
Sreyan Ghosh (NVIDIA), Bryan Catanzaro (NVIDIA)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningMultimodalityChain-of-ThoughtAudio
🎯 What it does: Developed the Music Flamingo model and constructed two large datasets, MF-Skills and MF-Think, to enhance music understanding and reasoning capabilities.
MVAR: Visual Autoregressive Modeling with Scale and Spatial Markovian Conditioning
Jinhua Zhang (University of Electronic Science and Technology of China), Shuhang Gu (University of Electronic Science and Technology of China)
GenerationTransformerImage
🎯 What it does: Propose the MVAR model, which generates images using a next-scale autoregressive framework with scale and spatial Markov assumptions
MVCustom: Multi-View Customized Diffusion via Geometric Latent Rendering and Completion
Minjung Shin (Yonsei University), Youngjung Uh (Yonsei University)
GenerationTransformerDiffusion modelNeural Radiance FieldImageVideoMesh
🎯 What it does: Proposes the MVCustom framework, achieving multi-view personalized generation under a given camera pose, preserving object identity while generating complete scenes based on text prompts.
MVR: Multi-view Video Reward Shaping for Reinforcement Learning
Lirui Luo (Peking University), Qing Li (State Key Laboratory of General Artificial Intelligence)
Robotic IntelligenceReinforcement LearningVision Language ModelVideoTextMultimodality
🎯 What it does: Propose a multi-view video reward shaping framework named MVR, which utilizes pre-trained vision-language models (VLM) to evaluate text similarity in multi-view videos, thereby providing visual guidance for reinforcement learning;
NAB: Neural Adaptive Binning for Sparse-View CT reconstruction
Wangduo Xie (KU Leuven), Matthew B. Blaschko (KU Leuven)
Biomedical DataComputed Tomography
🎯 What it does: Proposes a self-supervised CT reconstruction method based on Neural Adaptive Binning (NAB), leveraging differentiable binning encoding to capture rectangular shape priors in industrial CT.
NAIPv2: Debiased Pairwise Learning for Efficient Paper Quality Estimation
Penghai Zhao (Nankai University), Xiang Li (NKIARI)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextTabularReview/Survey Paper
🎯 What it does: Propose the NAIPv2 framework, which achieves efficient and debiased paper quality estimation by leveraging debiased paired learning and review confidence fusion signals.
Naming to Learn: Class Incremental Learning for Vision-Language Model with Unlabeled Data
Qiwei Li (Peking University), Jiahuan Zhou (Peking University)
ClassificationRepresentation LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose a visual-language model incremental learning method called N2L, which addresses the problem of noisy pseudo-labels in unlabeled incremental learning by providing only class names and unannotated images for each incremental task.
Nano3D: A Training-Free Approach for Efficient 3D Editing Without Masks
Junliang Ye, Jun Zhu
GenerationData SynthesisFlow-based ModelAuto EncoderPoint CloudMesh
🎯 What it does: Proposed a training-free, mask-free 3D object local editing framework called Nano3D, and constructed the first large-scale 3D editing dataset named Nano3DEdit-100k
NarrLV: Towards a Comprehensive Narrative-Centric Evaluation for Long Video Generation
Xiaokun Feng (University of Chinese Academy of Sciences), Kaiqi Huang (University of Chinese Academy of Sciences)
GenerationLarge Language ModelPrompt EngineeringVideoTextMultimodalityBenchmark
🎯 What it does: Proposed the NarrLV benchmark, specifically designed to evaluate the narrative expression capabilities of long video generation models;
Narrow Finetuning Leaves Clearly Readable Traces in Activation Differences
Julian Minder (EPFL Ecole Normale Superieure Paris Saclay), Neel Nanda (EPFL Ecole Normale Superieure Paris Saclay)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningAgentic AIText
🎯 What it does: This paper studies how fine-tuning large language models on finite domains leaves readable biases in model activations during the fine-tuning process.
Nasty Adversarial Training: A Probability Sparsity Perspective for Robustness Enhancement
Yuhang Zhou (Harbin Institute of Technology), Leo Yu Zhang (Griffith University)
Explainability and InterpretabilityComputational EfficiencyAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposed Malicious Adversarial Training (NAT), leveraging probabilistic sparse regularization to enhance model adversarial robustness
NatADiff: Adversarial Boundary Guidance for Natural Adversarial Diffusion
Max Collins (University of Western Australia), Ajmal Saeed Mian (University of Western Australia)
Adversarial AttackDiffusion modelAuto EncoderImage
🎯 What it does: Propose a natural adversarial example generation method based on diffusion models called NatADiff, which achieves efficient and transferable adversarial attacks by leveraging adversarial boundary guidance, enhanced classifier guidance, and time-travel sampling
Native Adaptive Solution Expansion for Diffusion-based Combinatorial Optimization
Yu Wang (School of Artificial Intelligence Jilin University), Yi Chang (School of Artificial Intelligence Jilin University)
OptimizationGraph Neural NetworkDiffusion modelGraphBenchmark
🎯 What it does: Propose NEXCO, a CO-direction mask diffusion framework that integrates adaptive expansion directly into the generative model itself;
Native Reasoning Models: Training Language Models to Reason on Unverifiable Data
Yuanfu Wang (Shanghai Artificial Intelligence Laboratory), Chao Yang (Shanghai Artificial Intelligence Laboratory)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose Native Reasoning Training (NRT), a method that cultivates complex reasoning capabilities by utilizing only question-answer pairs, without requiring expert demonstrations or external verifiers, through intrinsic rewards based on the model's self-generated reasoning trajectories.
Natural Identifiers for Privacy and Data Audits in Large Language Models
Lorenzo Rossi (CISPA Helmholtz Center for Information Security), Adam Dziedzic (CISPA Helmholtz Center for Information Security)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Proposes a method for posterior differential privacy auditing and dataset inference (DI) in large language models (LLMs) using natural identifiers (NIDs).
Natural Language PDDL (NL-PDDL) for Open-world Goal-oriented Commonsense Regression Planning in Embodied AI
Xiaotian Liu (University of Toronto), Scott Sanner (University of Toronto)
OptimizationRobotic IntelligenceLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: Propose the NL-PDDL framework, integrating natural language descriptions with classical PDDL, supporting goal-directed planning in open-world scenarios;
Navigating the Accuracy-Size Trade-Off with Flexible Model Merging
Akash Dhasade (EPFL), Anne-Marie Kermarrec (EPFL)
OptimizationComputational EfficiencyTransformerImageTextBenchmark
🎯 What it does: Proposed a data-free model fusion framework called FLEXMERGE, which can merge multi-task models at any size (including non-integer sizes).
Navigating the Latent Space Dynamics of Neural Models
Marco Fumero (IST Austria), Francesco Locatello (IST Austria)
Anomaly DetectionRepresentation LearningDiffusion modelAuto EncoderImageOrdinary Differential Equation
🎯 What it does: Investigate the dynamics of autoencoders in latent space, treating the encoding-decoding iterations as a hidden vector field, and explore how attractors in this vector field reflect the model's memory and generalization properties. Further utilize noise-generated attractors to probe model weight information in data-free scenarios and detect distribution drift using latent trajectories.
NC-Bench and NCfold: A Benchmark and Closed-Loop Framework for RNA Non-Canonical Base-Pair Prediction
Heqin Zhu (University of Science and Technology of China), S Kevin Zhou
TransformerBiomedical DataBenchmark
🎯 What it does: Constructed the NC-Bench benchmark dataset targeting RNA atypical base pairing, and proposed a closed-loop dual-branch framework NCfold for prediction
NDAD: Negative-Direction Aware Decoding for Large Language Models via Controllable Hallucination Signal Injection
Panjia Qiu (East China Normal University), Daixin Wang (AntGroup)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose NDAD, which utilizes masked key attention heads to extract hallucination signals and decodes them as negative interventions.
Near Optimal Robust Federated Learning Against Data Poisoning Attack
Jingfan Yu (Tsinghua University), Zhixuan Fang (Tsinghua University)
Federated LearningAdversarial AttackImage
🎯 What it does: Propose a robust training mechanism in federated learning against data poisoning attacks, which leverages the variance characteristics of worker node data to identify poisoned nodes and assign weights, then employs weighted averaging during target model training to achieve attack defense.
Near-Optimal Online Deployment and Routing for Streaming LLMs
Shaoang Li (Stony Brook University), Jian Li (Stony Brook University)
OptimizationLarge Language ModelText
🎯 What it does: Propose the StageRoute algorithm to address the two-layer decision problem of LLM online deployment and routing, balancing model flow, concurrency limits, cost budget, and throughput constraints.
Near-Optimal Sample Complexity Bounds for Constrained Average-Reward MDPs
Yukuan Wei (Fudan University), Lin F. Yang (University of California, Los Angeles)
Data SynthesisOptimizationReinforcement Learning
🎯 What it does: This study investigates the sample complexity of learning constrained average reward Markov decision processes (CAMDP) under a generative model, proposing a model-based algorithm that returns an ε-optimal policy under both relaxed and strict feasibility settings.
Near-Optimal Second-Order Guarantees for Model-Based Adversarial Imitation Learning
Shangzhe Li (UNC Chapel Hill), Weitong Zhang (UNC Chapel Hill)
OptimizationReinforcement LearningBenchmark
🎯 What it does: Propose a model-driven adversarial imitation learning algorithm, MB-AIL, which integrates reward learning, model learning, and planning, providing a novel first-order unbounded second-order sample complexity upper bound, and achieving near-optimal sampling efficiency with a small number of expert demonstrations and online interactions.
Nearly Space-Optimal Graph and Hypergraph Sparsification in Insertion-Only Data Streams
Vincent Cohen-Addad (Google Research), Samson Zhou (Texas A&M University)
OptimizationComputational EfficiencyGraph
🎯 What it does: Under the insertion-only model, we propose nearly optimal space graph and hypergraph sparsification algorithms that maintain (1+ε) approximation in sliding window and adversarial environments.
Nearly-Optimal Bandit Learning in Stackelberg Games with Side Information
Maria Florina Balcan, Steven Wu
OptimizationReinforcement LearningTabular
🎯 What it does: The study achieves near-optimal online learning in a Stackelberg game with side information under bandit feedback, obtaining a regret upper bound of ~O(√T).
Nef-Net v2: Adapting Electrocardio Panorama in the wild
Zehui Zhan (Hong Kong University of Science and Technology), Jintai Chen (Hong Kong University of Science and Technology)
RestorationConvolutional Neural NetworkBiomedical DataElectrocardiogram
🎯 What it does: Propose NEF-NET V2, constructing an ECG panoramic reconstruction system with arbitrary length and arbitrary perspective, supporting multi-device and on-site calibration;
Negative Pre-activations Differentiate Syntax
Linghao Kong (MIT), Nir N Shavit
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Investigate the negative pre-activation of smooth activation functions in large language models and explore their role in syntactic processing, with a focus on sparse Wasserstein neurons;
Negotiated Reasoning: On Provably Addressing Relative Over-Generalization
Junjie Sheng (Tongji University), Xiangfeng Wang (East China Normal University)
Reinforcement Learning
🎯 What it does: This paper proposes a theoretical avoidance and practical solution for relative overgeneralization (RO) in multi-agent reinforcement learning through a 'negotiation reasoning' mechanism.
NeMo-map: Neural Implicit Flow Fields for Spatio-Temporal Motion Mapping
Yufei Zhu (Örebro University), Martin Magnusson (Örebro University)
Object TrackingFlow-based ModelVideoTime SeriesSequential
🎯 What it does: Proposed a continuous spatiotemporal motion graph (NeMo-map) that uses an implicit neural network to predict motion distributions at any position and time.
Nemotron-CC-Math: A 133 Billion-Token-Scale High Quality Math Pretraining Dataset
Rabeeh Karimi mahabadi, Bryan Catanzaro (NVIDIA)
Data SynthesisData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Constructed and publicly released a high-quality mathematical pre-training dataset named Nemotron-CC-Math with a scale of 133B tokens, and trained models to verify its effectiveness.
Nemotron-Research-Tool-N1: Exploring Tool-Using Language Models with Reinforced Reasoning
Shaokun Zhang (NVIDIA), Guilin Liu (NVIDIA)
TransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Proposed a rule-based reinforcement learning method to train the tool call language model Tool-N1, eliminating the need for manually annotated reasoning trajectories.
NEO — No-Optimization Test-Time Adaptation through Latent Re-Centering
Alexander Murphy (University of Birmingham), Abhirup Ghosh (University of Birmingham)
Domain AdaptationTransformerImage
🎯 What it does: Proposes NEO, a no-hyperparameter, no-optimization test-time adaptation method that aligns source and target domain features by centroid-centering the global embedding means.
Neodragon: Mobile Video Generation Using Diffusion Transformer
Animesh Karnewar (Qualcomm AI Research), Amir Habibian (Qualcomm AI Research)
GenerationData SynthesisKnowledge DistillationTransformerDiffusion modelAuto EncoderVideoTextMultimodality
🎯 What it does: This paper proposes Neodragon, a low-power text-to-video Diffusion Transformer (DiT) model for mobile phones and laptops, achieving end-to-end performance of generating 49 frames of 640×1024 resolution video in 6.7 seconds through four technical optimizations.
Neologism Learning for Controllability and Self-Verbalization
John Hewitt (Google DeepMind), Been Kim (Google DeepMind)
Representation LearningTransformerPrompt EngineeringText
🎯 What it does: By freezing the original model and introducing new word embeddings in natural language, training only modifies the new word embeddings to achieve controllability over model concepts.
Neon: Negative Extrapolation From Self-Training Improves Image Generation
Sina Alemohammad (University of Texas at Austin), Richard Baraniuk (Rice University)
GenerationDiffusion modelFlow-based ModelAuto EncoderImageBenchmark
🎯 What it does: Propose the Neon method, which improves the quality of image generation models through negative extrapolation (reverse self-training degradation direction).
NePTune: A Neuro-Pythonic Framework for Tunable Compositional Reasoning on Vision-Language
Danial Kamali (Michigan State University), Parisa Kordjamshidi (Michigan State University)
Large Language ModelVision Language ModelMultimodality
🎯 What it does: This paper proposes NePTune, a neural symbolic framework that combines LLM-generated Python programs with VLM for visual-language compositional reasoning.
NeRV-Diffusion: Diffuse Implicit Neural Representation for Video Synthesis
Yixuan Ren (University of Maryland), Abhinav Shrivastava (University of Maryland)
GenerationData SynthesisTransformerDiffusion modelNeural Radiance FieldGenerative Adversarial NetworkVideo
🎯 What it does: Propose a two-stage video generation framework called NeRV-Diffusion, which compresses videos into neural network weights (INR) and performs diffusion generation in this weight space, achieving efficient video synthesis.
NerVE: Nonlinear Eigenspectrum Dynamics in LLM Feed-Forward Networks
Nandan Kumar Jha (New York University), Brandon Reagen (New York University)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Proposes the NerVE framework, which utilizes four spectral metrics (Spectral Entropy, Participation Ratio, Eigenvalue Early Enrichment, Jensen-Shannon Divergence) to online and memory-efficiently track the pre- and post-activation feature spectrum changes in the FFN of large language models, thereby revealing how FFN nonlinearity actively redistributes variance and regulates information flow.
Nesterov Finds GRAAL: Optimal and Adaptive Gradient Method for Convex Optimization
Ekaterina Borodich (Higher School of Economics), Dmitry Kovalev (Yandex Research)
Optimization
🎯 What it does: Proposed an adaptive accelerated gradient method that combines Nesterov acceleration with GRAAL's local curvature estimation, addressing the challenge of achieving accelerated convergence without requiring line search or hyperparameter tuning.
NetArena: Dynamic Benchmarks for AI Agents in Network Automation
Yajie Zhou (University of Maryland), Zaoxing Liu (University of Maryland)
OptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIBenchmarkChain-of-Thought
🎯 What it does: Proposed the NETARENA framework for dynamically generating benchmark tasks for network automation and real-time verification of the correctness, safety, and latency performance of LLM agents in simulation environments.