ICLR 2026 Papers — Page 30
International Conference on Learning Representations · 5356 papers
Minor First, Major Last: A Depth-Induced Implicit Bias of Sharpness-Aware Minimization
Chaewon Moon (Korea Advanced Institute of Science and Technology), Chulhee Yun (Korea Advanced Institute of Science and Technology)
OptimizationConvolutional Neural NetworkImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Investigated the implicit bias of Sharpness-Aware Minimization (SAM) in linear diagonal networks (particularly two-layer networks), revealing the impact of different norms (∞ and 2) on training dynamics and final directions.
MIRA: Memory-Integrated Reinforcement Learning Agent with Limited LLM Guidance
Narjes Nourzad (University of Southern California), Carlee Joe-Wong (Carnegie Mellon University)
Large Language ModelReinforcement LearningBenchmark
🎯 What it does: Proposed a reinforcement learning agent called MIRA, which integrates guidance from large language models (LLMs) through a structured memory graph to improve early learning efficiency in sparse reward environments.
MIRACLE: Model-free Imitation and Reinforcement Learning for Adaptive Cut-Selection
Arjun M. (Indian Institute of Technology Delhi), Manojkumar Ramteke (Indian Institute of Technology Delhi)
OptimizationReinforcement LearningTabularBenchmark
🎯 What it does: In mixed integer programming solving, memory-priority cutting plane selection is achieved through reinforcement learning, significantly reducing memory usage and improving solving efficiency.
Mirage or Method? How Model–Task Alignment Induces Divergent RL Conclusions
Haoze Wu (Zhejiang University), Junxian He (HKUST)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper systematically experimentally verifies several counterintuitive phenomena in reinforcement learning (RL) for large language model reasoning and explores whether these phenomena depend on the alignment strength between the model and the task.
Mirror Flow Matching with Heavy-Tailed Priors for Generative Modeling on Convex Domains
Yunrui Guan (Rice University), Shiqian Ma (Rice University)
GenerationData SynthesisFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: Propose a Mirror Flow Matching method tailored for convex constraint domains, introducing regularized log-barrier mirror maps and student-t priors to address issues caused by traditional log-barrier mirror maps, such as heavy-tailed distributions and poor matching with Gaussian priors, thereby achieving more stable and high-quality generative models.
Misaligned Roles, Misplaced Images: Structural Input Perturbations Expose Multimodal Alignment Blind Spots
Erfan Shayegani (University of California, Riverside), Yue Dong (University of California, Riverside)
Safty and PrivacyRepresentation LearningAdversarial AttackTransformerVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper proposes a new class of structural attacks—Role-Modal Attack (RMA)—which bypasses the rejection mechanisms of multimodal language models through role swapping and image token position transformation. The attack is evaluated on three VLMs, and the attack success rate is significantly reduced through adversarial training on the transformation.
MiSS: Revisiting the Trade-off in LoRA with an Efficient Shard-Sharing Structure
Jiale Kang (Yuanshi Inc), Qingyu Yin (Zhejiang University)
OptimizationComputational EfficiencyTextMultimodality
🎯 What it does: Proposed the MiSS structure, which uses a single shared matrix D to update weights through a sharding expansion approach, addressing the slow convergence caused by LoRA's synchronous updates.
Missingness Bias Calibration in Feature Attribution Explanations
Shailesh Sridhar (University of Pennsylvania), Eric Wong (University of Pennsylvania)
Explainability and InterpretabilityImageTextTabularBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Proposed a post-calibration method named MCal, which corrects missingness bias caused by feature masks by applying a simple linear transformation on the outputs of a frozen model, thereby enhancing the reliability of explanation methods.
Mitigating Hallucination in Vision-Language Model with Depth and Spatial-aware Key-Value Refinement
Gusang Lee (Seoul National University), Byonghyo Shim (Seoul National University)
Depth EstimationVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a lightweight, training-agnostic depth- and space-aware KV cache refinement method (DSCR), which enhances visual language models' correct attention to images and suppresses visual hallucinations by reweighting visual key vectors during inference using estimated depth and relative spatial distance.
Mitigating Mismatch within Reference-based Preference Optimization
Suqin Yuan (University of Sydney), Tongliang Liu (University of Sydney)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackSupervised Fine-TuningText
🎯 What it does: Propose Hybrid-DPO (HyPO), a lightweight improvement in DPO that conditionally prunes the reference marginal to address the training-inference mismatch caused by 'premature satisfaction';
Mitigating Noise Shift in Denoising Generative Models with Noise Awareness Guidance
Jincheng Zhong (Tsinghua University), Mingsheng Long (Tsinghua University)
GenerationSupervised Fine-TuningDiffusion modelFlow-based ModelImage
🎯 What it does: Propose and address the noise shift problem occurring during the sampling process of diffusion generative models, introducing Noise Awareness Guidance (NAG) and its classifier-free version.
Mitigating Non-IID Drift in Zeroth-Order Federated LLM Fine-Tuning with Transferable Sparsity
Yide Ran (Stevens Institute of Technology), Zhaozhuo Xu (Stevens Institute of Technology)
OptimizationFederated LearningTransformerText
🎯 What it does: Propose two sparse zeroth-order optimization methods, MEERKAT and MEERKAT-VP, for federated LLM fine-tuning, leveraging extremely sparse transferable parameters to achieve efficient training with high-frequency low-communication;
Mitigating Privacy Risk via Forget Set-Free Unlearning
Aviraj Newatia (University of Toronto), Rahul G Krishnan (University of Toronto)
Safty and PrivacyComputational EfficiencyConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: Propose a partial blind machine learning model forgetting technique (PBU) and implement an algorithm called RELOAD, which can achieve high-quality forgetting without accessing the forgetting set.
Mitigating Semantic Collapse in Generative Personalization with Test-Time Embedding Adjustment
Anh Tuan Bui (Monash University), Dinh Phung (Monash University)
GenerationPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: This paper proposes a Test-time Embedding Adjustment (TEA) method to address the semantic collapse problem in personalized models, enabling personalized tokens to no longer suppress other semantics in multi-concept prompts.
Mitigating Spurious Correlation via Distributionally Robust Learning with Hierarchical Ambiguity Sets
Sung Ho Jo (Pohang University of Science and Technology), Minwoo Chae (Pohang University of Science and Technology)
ClassificationDomain AdaptationImage
🎯 What it does: This paper proposes a hierarchical uncertainty set distributed robust learning framework that can simultaneously suppress distribution drift caused by pseudo correlations across groups and within groups.
Mitigating the Curse of Detail: Scaling Arguments for Feature Learning and Sample Complexity
Noa Rubin (Hebrew University), Zohar Ringel (IBM Research)
Representation LearningConvolutional Neural NetworkPhysics Related
🎯 What it does: Proposes a heuristic method based on scale analysis and variational approximation to predict the feature learning (FL) patterns and scalable behavior of sample complexity in Bayesian neural networks.
Mitigating the Safety Alignment Tax with Null-Space Constrained Policy Optimization
Yifan Niu (Hong Kong University of Science and Technology Guangzhou), Jia Li (Hong Kong University of Science and Technology Guangzhou)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: Propose an RL framework NSPO that projects safety policy gradients into the zero space of general tasks, achieving LLM safety alignment without compromising original capabilities.
Mix-Ecom: Towards Mixed-Type E-Commerce Dialogues with Complex Domain Rules
Chenyu Zhou (Xiamen University), Rongrong Ji (Xiamen University)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AITextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed a hybrid-type e-commerce dialogue dataset called Mix-ECom, and developed a dynamic rule filtering framework (E-ReAct/E-Plan&Solve) based on ReAct and Plan&Solve, evaluating multiple LLMs on this benchmark.
Mixed-Curvature Tree-Sliced Wasserstein Distance
Duy-Tung Pham (FPT Software AI Center), Tan Minh Nguyen
ClassificationGenerationComputational EfficiencyImageGraph
🎯 What it does: Propose the hybrid curvature tree split Wasserstein distance (MCTSW), achieving efficient closed-form comparison of probability distributions in hybrid curvature spaces by constructing a hybrid curvature tree system and corresponding Radon transform.
Mixing Importance with Diversity: Joint Optimization for KV Cache Compression in Large Vision-Language Models
Xuyang Liu (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
CompressionComputational EfficiencyMultimodality
🎯 What it does: For KV cache compression in large-scale vision-language models (LVLMs), the MixKV method is proposed, which jointly optimizes the importance and diversity of KV pairs, supporting adaptive adjustment of their weights under different attention heads.
Mixing Mechanisms: How Language Models Retrieve Bound Entities In-Context
Yoav Gur-Arieh (Tel Aviv University), Atticus Geiger (Pr(Ai) 2 R Group)
RetrievalExplainability and InterpretabilityLarge Language ModelText
🎯 What it does: Investigated the mechanisms of large language models in contextually binding and retrieving entities, systematically evaluated three mechanisms (position, lexical, and self-referential), and proposed a causal model that combines them;
MixLinear: Extreme Low Resource Multivariate Time Series Forecasting with $0.1K$ Parameters
Aitian Ma (Florida International University), Mo Sha (Florida International University)
Computational EfficiencyTime SeriesBenchmark
🎯 What it does: Propose MixLinear, a low-parameter long-term sequential prediction model that combines piecewise linear time-domain extraction with adaptive low-rank frequency-domain filtering.
Mixture of Cognitive Reasoners: Modular Reasoning with Brain-Like Specialization
Badr AlKhamissi (EPFL), Antoine Bosselut (EPFL)
Explainability and InterpretabilityTransformerSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Proposed a modular Transformer architecture named Mixture of Cognitive Reasoners (MICRO), which utilizes four expert modules corresponding to language, logic, social reasoning, and world knowledge networks in the human brain, achieving functional specialization through a three-stage training curriculum.
Mixture of Contexts for Long Video Generation
Shengqu Cai (Stanford University), Gordon Wetzstein (Stanford University)
GenerationTransformerMixture of ExpertsDiffusion modelVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes the Mixture of Contexts (MoC) framework, which uses learnable sparse attention routing to generate long videos (minutes-level), addressing the quadratic complexity bottleneck of traditional Transformer self-attention.
Mixture of Mini Experts: Overcoming the Linear Layer Bottleneck in Multiple Instance Learning
Daniel Shao (Massachusetts Institute of Technology), Faisal Mahmood (Harvard University)
ClassificationTransformerMixture of ExpertsBiomedical Data
🎯 What it does: The paper proposes the MAMMOTH module to replace the linear layer in multi-instance learning (MIL), utilizing a multi-head soft Mixture-of-Experts to generate task-specific slice features.
Mixture-of-Experts Can Surpass Dense LLMs Under Strictly Equal Resource
Houyi Li (Fudan University), Daxin Jiang (StepFun)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Under strict constraints of equal parameter count, computational cost, and data volume, a three-step experimental framework (first optimizing MoE architecture, then identifying the optimal activation rate, and finally employing data reuse) was used to verify that Mixture-of-Experts LLM can outperform equally sized dense models.
Mixture-of-Visual-Thoughts: Exploring Context-Adaptive Reasoning Mode Selection for General Visual Reasoning
Zejun Li (Fudan University), zhongyu wei
Large Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposed the MoVT framework and the two-stage AdaVaR training method, unifying multiple visual reasoning modes and achieving context-adaptive mode selection.
Mixture-of-World Models: Scaling Multi-Task Reinforcement Learning with Modular Latent Dynamics
Boxuan Zhang (Beijing Institute of Technology), Gang Wang (Beijing Institute of Technology)
TransformerReinforcement LearningMixture of ExpertsAuto EncoderWorld ModelImage
🎯 What it does: Propose a hybrid world model (MoW) in multi-task reinforcement learning, achieving efficient visual task modeling and planning through task-specific VAEs, hybrid Transformer experts, and shared Transformers;
MLE-Smith: Scaling MLE Tasks with Automated Multi-agent Pipeline
Rushi Qiang (Georgia Institute of Technology), Bo Dai (Georgia Institute of Technology)
Data-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTextMultimodalityTabularTime SeriesBenchmark
🎯 What it does: Proposes MLE-Smith, a fully automated multi-agent generate-validate-execute pipeline for transforming raw datasets into competition-level machine learning engineering tasks.
MLP Memory: A Retriever-Pretrained Memory for Large Language Models
Rubin Wei (Shanghai Jiao Tong University), Zhouhan Lin (Shanghai Jiao Tong University)
RetrievalComputational EfficiencyLarge Language ModelText
🎯 What it does: Train a lightweight full MLP module in large language models (LLMs) to mimic the distribution of a k-NN retriever, enabling memory and invocation of external knowledge without explicit retrieval.
MM-HELIX: Boosting Multimodal Long-Chain Reflective Reasoning with Holistic Platform and Adaptive Hybrid Policy Optimization
Xiangyu Zhao (Shanghai Jiao Tong University), Xue Yang (Shanghai Jiao Tong University)
OptimizationLarge Language ModelSupervised Fine-TuningReinforcement LearningTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the MM-HELIX reflective long-chain reasoning benchmark, the MM-HELIX-100K high-quality dataset, and designed the Adaptive Hybrid Policy Optimization (AHPO) training framework to enhance the performance of multi-modal large language models in reflective reasoning tasks.
MMaDA-Parallel: Multimodal Large Diffusion Language Models for Thinking-Aware Editing and Generation
Ye Tian (Peking University), Xiangtai Li (ByteDance)
GenerationLarge Language ModelReinforcement LearningVision Language ModelDiffusion modelMultimodalityBenchmark
🎯 What it does: Designed and implemented a parallel multimodal diffusion framework, MMaDA-Parallel, for thought-aware image editing and generation, and proposed an evaluation benchmark called ParaBench.
MMDuet2: Enhancing Proactive Interaction of Video MLLMs with Multi-Turn Reinforcement Learning
Yueqian Wang (Peking University), Dongyan Zhao (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodality
🎯 What it does: Developed a video multimodal large language model called MMDuet2 that can actively decide the timing of responses, and constructed a proactive question-answering dataset with 52k videos.
MME-Emotion: A Holistic Evaluation Benchmark for Emotional Intelligence in Multimodal Large Language Models
Fan Zhang (Chinese University of Hong Kong), Pheng-Ann Heng (Chinese University of Hong Kong)
RecognitionLarge Language ModelAgentic AIVideoTextMultimodalityBenchmarkChain-of-ThoughtAudio
🎯 What it does: Constructed the MME-Emotion benchmark, collecting 6,500 video clips along with corresponding emotion recognition and reasoning question-answer pairs, covering 8 categories of emotion tasks and 27 different scenarios.
MME-Unify: A Comprehensive Benchmark for Unified Multimodal Understanding and Generation Models
Wulin Xie (CASIA), Liang Wang (CASIA)
GenerationLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Created the MME-UNIFY benchmark for unified evaluation of multimodal understanding, generation, and hybrid modal generation capabilities.
MMedAgent-RL: Optimizing Multi-Agent Collaboration for Multimodal Medical Reasoning
Peng Xia (University Of North Carolina Chapel Hill), Huaxiu Yao (University Of North Carolina Chapel Hill)
OptimizationLarge Language ModelReinforcement LearningAgentic AIVision Language ModelMultimodalityBiomedical Data
🎯 What it does: Proposed a multi-agent framework based on reinforcement learning, MMedAgent-RL, for multimodal medical reasoning, simulating the clinical workflow from triage to specialists and then to attending physicians;
MMPD: Diverse Time Series Forecasting via Multi-Mode Patch Diffusion Loss
Yunhao Zhang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
TransformerDiffusion modelTime Series
🎯 What it does: Proposed a multi-modal differentiable time series prediction loss based on diffusion models (MMPD), which can achieve multi-modal (multiple possible futures) prediction on any patch-based prediction backbone while also considering deterministic prediction.
MMR-Life: Piecing Together Real-life Scenes for Multimodal Multi-image Reasoning
Jiachun Li (University of Chinese Academy of Sciences), Jun Zhao (University of Chinese Academy of Sciences)
ImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes MMR-Life, a benchmark for evaluating multimodal large language models on seven categories of real-world multi-image reasoning tasks, containing 2,646 multiple-choice questions and 19,108 real images.
MMR-V: What's Left Unsaid? A Benchmark for Multimodal Deep Reasoning in Videos
Kejian Zhu (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences), Jun Zhao (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences)
TransformerLarge Language ModelPrompt EngineeringVideoMultimodalityBenchmarkChain-of-ThoughtAudio
🎯 What it does: Proposes MMR-V, a manually annotated benchmark for video multimodal deep reasoning, covering 317 videos and 1257 multimodal reasoning questions;
MMReD: a Cross-Modal Benchmark for Dense Context Reasoning
Maxim Kurkin (FusionBrain Lab), Andrey Kuznetsov (FusionBrain Lab)
TransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextBenchmark
🎯 What it does: Propose the MMReD benchmark to evaluate the reasoning capabilities of LLMs and LVLMs in dense contexts.
MMSearch-Plus: Benchmarking Provenance-Aware Search for Multimodal Browsing Agents
Xijia Tao (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
RetrievalAgentic AIVision Language ModelImageVideoTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the MMSearch-Plus benchmark, specifically designed to evaluate the capabilities of multi-modal browsing agents in visual fine-grained reasoning and cross-source retrieval.
MMSI-Bench: A Benchmark for Multi-Image Spatial Intelligence
Sihan Yang (Shanghai AI Laboratory), Jiangmiao Pang (Shanghai AI Laboratory)
TransformerPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed MMSI-Bench - a VQA benchmark for multi-image spatial reasoning, containing 1000 multi-step multi-image multiple-choice questions with human-annotated reasoning processes;
MMSU: A Massive Multi-task Spoken Language Understanding and Reasoning Benchmark
Dingdong WANG (Chinese University of Hong Kong), Helen M. Meng (Chinese University of Hong Kong)
Large Language ModelPrompt EngineeringMultimodalityBenchmarkAudio
🎯 What it does: Constructed a large-scale multi-task speech understanding and reasoning benchmark, MMSU, containing 5,000 audio-question pairs across 47 tasks, with a focus on speech perception and reasoning capabilities.
MMTok: Multimodal Coverage Maximization for Efficient Inference of VLMs
Sixun Dong (Arizona State University), Qi Qian (Zoom Communications)
Computational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: Propose a training-free multi-modal maximum coverage method called MMTok, which selects the most informative subset from a large number of visual tokens to enhance the inference efficiency of Vision-Language Models (VLMs).
MnemoDyn: Learning Resting State Dynamics from $40$K FMRI sequences
Sourav Pal (University of Wisconsin Madison), Vikas Singh (University of Wisconsin Madison)
ClassificationRepresentation LearningConvolutional Neural NetworkSupervised Fine-TuningAuto EncoderTime SeriesSequentialBiomedical DataMagnetic Resonance ImagingAlzheimer's DiseaseStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose a dynamic system model called MnemoDyn, which combines multiscale wavelet subbasis functions with pseudodifferential operators, for large-scale self-supervised pretraining and fine-tuning on multiple prediction tasks under resting-state fMRI (rs-fMRI);
MOAI: Module-Optimizing Architecture for Non-Interactive Secure Transformer Inference
Linru Zhang (Nanyang Technological University), Kwok-Yan Lam (Nanyang Technological University)
Computational EfficiencyTransformerText
🎯 What it does: MOAI proposes an efficient non-interactive secure Transformer inference framework
MoAlign: Motion-Centric Representation Alignment for Video Diffusion Models
Aritra Bhowmik (University of Amsterdam), Mohsen Ghafoorian (Qualcomm AI Research)
GenerationRepresentation LearningTransformerDiffusion modelAuto EncoderOptical FlowVideoMultimodality
🎯 What it does: Improve the temporal consistency and physical plausibility of video generation by learning a motion subspace extracted from a frozen video encoder and aligning the latent features of text-to-video diffusion models.
MoBE: Mixture-of-Basis-Experts for Compressing MoE-based LLMs
Xiaodong Chen (Renmin University of China), Jianguo Li (Inclusion AI)
CompressionComputational EfficiencyKnowledge DistillationLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: Propose Mixture-of-Basis-Experts (MoBE) by performing rank decomposition on MoE expert weights and sharing a basis matrix, achieving parameter compression while maintaining performance.
MobiEdit: Resource-efficient Knowledge Editing for Personalized On-device LLMs
Zhenyan Lu (Beijing University of Posts and Telecommunications), Mengwei Xu (Beijing University of Posts and Telecommunications)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Developed the MobiEdit framework, enabling resource-efficient knowledge editing of 3B-scale LLMs on mobile devices.
Mobile-GS: Real-time Gaussian Splatting for Mobile Devices
Xiaobiao Du (University of Technology Sydney), Xin Yu (Adelaide University)
GenerationComputational EfficiencyGaussian SplattingImage
🎯 What it does: Proposes Mobile-GS, a novel method for achieving real-time high-quality Gaussian splatting specifically optimized for mobile devices;
MobileIPL: Enhancing Mobile Agents Thinking Process via Iterative Preference Learning
Kun Huang (Xiaomi Inc), Bo An (Xiaomi Inc)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringVision Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Propose the MobileIPL framework, achieving self-training for mobile GUI agents through instruction evolution, CoaT tree iterative sampling, and thinking layer DPO based on rule-based rewards.
MobileKGQA: On-Device KGQA System on Dynamic Mobile Environments
Junyong Ahn (Seoul National University), Sungroh Yoon (Seoul National University)
Computational EfficiencyGraph Neural NetworkTransformerLarge Language ModelGraph
🎯 What it does: Developed a KGQA system called MobileKGQA that can perform local training and inference on mobile devices, supporting dynamic updates of knowledge graphs and answer generation.
MobileLLM-R1: Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes
Changsheng Zhao (Meta AI), Vikas Chandra (Meta AI)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Built a series of small-scale (≤1B parameters) inference language models called MobileLLM‑R1, achieving strong inference capabilities through efficient data curation and adaptive sampling.
MobileRL: Online Agentic Reinforcement Learning for Mobile GUI Agents
Yifan Xu (Tsinghua University), Yuxiao Dong (Z.AI)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelMultimodality
🎯 What it does: Proposed an online agent-based reinforcement learning framework called MOBILERL, specifically designed for training mobile GUI agents, and designed the difficulty-adaptive GRPO (ADAGRPO) algorithm based on this framework.
MOBODY: Model-Based Off-Dynamics Offline Reinforcement Learning
Yihong Guo (Johns Hopkins University), Anqi Liu (Johns Hopkins University)
Reinforcement LearningAuto EncoderBenchmark
🎯 What it does: In an offline, dynamics-mismatched setting, propose the MOBODY algorithm by learning the target domain dynamics using a model and exploring policies through simulation rollouts;
MoCa: Modeling Object Consistency for 3D Camera Control in Video Generation
Chengzhijing (University Of Science And Technology Of China), Xun Yang (University Of Science And Technology Of China)
GenerationTransformerDiffusion modelVideoText
🎯 What it does: This paper proposes a dual-branch MoCa framework for achieving 3D camera control in text-driven video generation, enhancing video quality by learning view, appearance, and motion consistency.
Mod-Adapter: Tuning-Free and Versatile Multi-concept Personalization via Modulation Adapter
Weizhi Zhong (Sun Yat-sen University), Guanbin Li (Kuaishou Technology)
GenerationTransformerMixture of ExpertsVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Developed a Mod-Adapter-based multi-concept personalized image generation method without fine-tuning, capable of simultaneously customizing objects and abstract concepts (pose, lighting, material, etc.)
Modal Aphasia: Can Unified Multimodal Models Describe Images From Memory?
Michael Aerni (ETH Zurich), Florian Tramèr (ETH Zurich)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningImageTextMultimodality
🎯 What it does: The study reveals that unified multimodal models excel in visual generation but suffer significant failures in textual descriptions, introducing and validating the 'modal aphasia' phenomenon and demonstrating its potential threat to AI safety.
Modality Alignment across Trees on Heterogeneous Hyperbolic Manifolds
Wu Wei (Beijing Institute of Technology), Mehrtash Harandi (Monash University)
Representation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed the 'Alignment across Trees' method, constructing hierarchical feature trees for images and text and aligning them on heterogeneous hyperbolic manifolds to address asymmetry in audio-visual alignment.
Modality-free Graph In-context Alignment
Wei Zhuo (Nanyang Technological University), Siqiang Luo (Nanyang Technological University)
ClassificationDomain AdaptationMeta LearningGraph Neural NetworkPrompt EngineeringGraph
🎯 What it does: Proposes the MF-GIA framework, which leverages pre-trained graph neural networks to achieve cross-domain, modality-agnostic few-shot context learning, enabling parameter-free updates for node/edge classification on unseen domains.
Mode-conditioning unlocks superior test-time compute scaling
Chen Henry Wu (Carnegie Mellon University), Aditi Raghunathan (Carnegie Mellon University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelReinforcement LearningMixture of ExpertsTextChain-of-Thought
🎯 What it does: Propose the Mode-Conditioning (ModC) framework, which explicitly distinguishes different reasoning modes (e.g., specialized models or prefixes) during training and allocates sampling budgets according to modes during inference, thereby eliminating diversity collapse and improving test-time performance with parallel sampling.
Model Already Knows the Best Noise: Bayesian Active Noise Selection via Attention in Video Diffusion Model
Kwanyoung Kim (Gwangju Institute of Science and Technology), Sanghyun Kim (Samsung Research)
GenerationDiffusion modelVideo
🎯 What it does: Propose an active noise selection framework ANSE based on model attention uncertainty for video diffusion model generation
Model Collapse Is Not a Bug but a Feature in Machine Unlearning for LLMs
Yan Scholten (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a novel LLM unlearning method based on model collapse called Partial Model Collapse (PMC), which iteratively fine-tunes the model using its own generated answers to progressively 'forget' answers to specific sensitive questions;
Model Predictive Adversarial Imitation Learning for Planning from Observation
Tyler Han (University of Washington), Byron Boots (University of Washington)
Autonomous DrivingOptimizationExplainability and InterpretabilityReinforcement LearningGenerative Adversarial NetworkTabularTime Series
🎯 What it does: Propose an end-to-end planning adversarial imitation learning framework (MPAIL), which learns the reward function by observing only demonstration states and uses the MPPI planner to online optimize the policy at each step, achieving complete closed-loop learning from observation to control.
Model-based Offline RL via Robust Value-Aware Model Learning with Implicitly Differentiable Adaptive Weighting
Zhongjian Qiao (City University Of Hong Kong), Shuang Qiu (University Of Chicago)
Reinforcement Learning
🎯 What it does: Propose the ROMI method to enhance the robustness and stability of RAMBO;
Model-Guided Microstimulation Steers Primate Visual Behavior
Johannes Mehrer (EPFL), Martin Schrimpf (EPFL)
Convolutional Neural NetworkDiffusion modelGenerative Adversarial NetworkImageText
🎯 What it does: Designed and validated a model-driven micro-stimulation framework based on a top-level deep network to regulate visual behavior in the higher visual cortex (IT area) of primates, combining electrode-model mapping, stimulation parameter optimization, and experimental validation to demonstrate that model predictions correlate with animal behavior.
Modeling Interference for Treatment Effect Estimation in Network Dynamic Environment
Qiang Huang (Mohamed bin Zayed University of Artificial Intelligence), Jin Tian (Mohamed bin Zayed University of Artificial Intelligence)
Recurrent Neural NetworkGraph Neural NetworkGenerative Adversarial NetworkGraph
🎯 What it does: Proposed a framework named DSPNET for estimating intervention effects in dynamic network environments, capable of simultaneously modeling time-evolving hidden covariates and network interventions (interference effects).
Modeling Others' Minds as Code
Kunal Jha, Max Kleiman-Weiner (University Of Washington)
AI Code AssistantTransformerLarge Language ModelTextSequential
🎯 What it does: This paper proposes a framework called ROTE, which models others' behaviors as executable programs. It generates script-like code using an LLM and predicts the next actions of humans and AI through Bayesian inference with weighted probabilities.
Modeling the Density of Pixel-level Self-supervised Embeddings for Unsupervised Pathology Segmentation in Medical CT
Mikhail Goncharov, Maxim Panov
SegmentationAnomaly DetectionKnowledge DistillationConvolutional Neural NetworkFlow-based ModelContrastive LearningBiomedical DataComputed Tomography
🎯 What it does: Designed and implemented a completely unsupervised CT image pathology detection framework called Screener, which estimates pathological abnormalities using feature maps obtained through dense self-supervised learning and obscured-invariant conditional variables.
MoDr: Mixture-of-Depth-Recurrent Transformers for Test-Time Reasoning
Xiaojing Zhang (DataCanvas), Zhanxing Zhu (University of Southampton)
Computational EfficiencyRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose MoDr Transformer, which dynamically selects the most suitable branch for the next step generation by introducing multi-branch LoRA and hard gate routing in Huginn's recurrent module, achieving multi-path 'deep reasoning',
MoE-GS: Mixture of Experts for Dynamic Gaussian Splatting
In-Hwan Jin (Pusan National University), Kyeongbo Kong (Pusan National University)
GenerationKnowledge DistillationMixture of ExpertsGaussian SplattingVideoBenchmark
🎯 What it does: Proposed the MoE-GS framework, which achieves adaptive reconstruction by combining multiple dynamic Gaussian rasterization experts and using a volume-aware pixel router.
MoEEdit: Efficient and Routing-Stable Knowledge Editing for Mixture-of-Experts LLMs
Yupu Gu (Tsinghua University), Pan Li (Georgia Institute of Technology)
OptimizationComputational EfficiencyMixture of ExpertsTextBenchmark
🎯 What it does: Propose the MoEEdit framework to achieve stable routing and computationally efficient knowledge editing in Mixture-of-Experts LLMs;
MoGA: Mixture-of-Groups Attention for End-to-End Long Video Generation
Weinan Jia (University Of Science And Technology Of China), Zhendong Mao (University Of Science And Technology Of China)
GenerationTransformerVision Language ModelDiffusion modelVideoMultimodality
🎯 What it does: This paper proposes the Mixture-of-Groups Attention (MoGA) mechanism, achieving end-to-end long video generation, capable of handling context lengths of approximately 580,000 tokens on multi-shot videos with specifications of 1 minute, 480p, and 24fps;
MoGen: Detailed Neuronal Morphology Generation via Point Cloud Flow Matching
Franz Rieger (Google Research), Michal Januszewski (Google Research)
GenerationData SynthesisTransformerFlow-based ModelPoint CloudBiomedical Data
🎯 What it does: Constructed MoGen, a flow-matching-based point cloud generation model for generating high-resolution, controllable neuronal morphologies, and applied it to improve the morphological feasibility discriminator in connectomics, significantly reducing the cost of post-hoc manual correction.
MoL: Adaptive Mixture-of-Length Reasoning for Efficient Question Answering with Context
Guocong Li (Zhejiang University), Jian Wu (Zhejiang University)
Computational EfficiencyTransformerReinforcement LearningText
🎯 What it does: Propose the Mixture-of-Length (MoL) framework, which adaptively controls the answer length based on question difficulty to achieve efficient and accurate context-based question answering.
MolecularIQ: Characterizing Chemical Reasoning Capabilities Through Symbolic Verification on Molecular Graphs
Christoph Bartmann (Johannes Kepler University Linz), Sohvi Luukkonen (Johannes Kepler University Linz)
Drug DiscoveryLarge Language ModelMixture of ExpertsGraphBiomedical DataBenchmark
🎯 What it does: Designed and released a fully symbolically verifiable molecular structure reasoning benchmark called MOLECULARIQ, covering three task categories: counting, indexing, and constraint generation, while conducting fine-grained evaluation of the molecular reasoning capabilities of multiple LLMs.
MolEditRL: Structure-Preserving Molecular Editing via Discrete Diffusion and Reinforcement Learning
Yuanxin Zhuang (Hong Kong University of Science and Technology Guangzhou), Ying Sun (Nanjing University of Aeronautics and Astronautics)
Drug DiscoveryTransformerReinforcement LearningDiffusion modelBiomedical Data
🎯 What it does: Propose MolEditRL, a structure-preserving molecular editing framework that combines discrete graph diffusion pre-training with reinforcement learning fine-tuning.
MolLangBench: A Comprehensive Benchmark for Language-Prompted Molecular Structure Recognition, Editing, and Generation
Feiyang Cai (Clemson University), Feng Luo (University of Delaware)
RecognitionGenerationDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityGraphBenchmark
🎯 What it does: Create the MolLangBench benchmark to evaluate models' cross-modal capabilities in molecular structure recognition, molecular editing under language prompts, and generation tasks.
MOLM: Mixture of LoRA Markers
Samar Fares (Mohamed bin Zayed University of Artificial Intelligence), Karthik Nandakumar (Mohamed bin Zayed University of Artificial Intelligence)
GenerationDiffusion modelImageText
🎯 What it does: Designed and implemented a trackable watermark framework (MOLM) for text-to-image diffusion models, which achieves watermark embedding and extraction by dynamically routing lightweight low-rank adapters (LoRA) with key-based parameter perturbation.
MoM: Linear Sequence Modeling with Mixture-of-Memories
Jusen Du (Tsinghua University), Yu Cheng (Chinese University of Hong Kong)
Text
🎯 What it does: Propose the Mixture-of-Memories (MoM) architecture, which uses multiple independent memory states and assigns inputs to specific memories via a router, thereby enhancing the memory capacity of linear sequence models and reducing memory interference.
MoMa: A Simple Modular Learning Framework for Material Property Prediction
Botian Wang (Tsinghua University), Hao Zhou (Tsinghua University)
Representation LearningGraph Neural NetworkSupervised Fine-TuningGraphPhysics Related
🎯 What it does: Proposes the MoMa framework, which first trains specialized modules on multiple tasks and centralizes them in a Hub, then uses Adaptive Module Combination (AMC) to weight and fine-tune these modules for each downstream material property prediction task.
MoMaGen: Generating Demonstrations under Soft and Hard Constraints for Multi-Step Bimanual Mobile Manipulation
Chengshu Li (Stanford University), Li Fei-Fei (Stanford University)
OptimizationRobotic IntelligenceVideo
🎯 What it does: Propose the MOMAGEN framework, which generates diverse demonstrations for multi-step dual-arm mobile manipulation tasks through constrained optimization using single human demonstration data, and trains visual motion policies with these demonstrations.
MomaGraph: State-Aware Unified Scene Graphs with Vision-Language Models for Embodied Task Planning
Yuanchen Ju (University of California, Berkeley), Koushil Sreenath (University of California, Berkeley)
Robotic IntelligenceReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerReinforcement LearningVision Language ModelMultimodalityGraphBenchmark
🎯 What it does: Propose MomaGraph scene graph representation and its VLM prediction model MomaGraph-R1 for mobile manipulation task planning, and construct a large-scale multi-view task-oriented dataset MomaGraph-Scenes and a comprehensive evaluation benchmark MomaGraph-Bench.
MoNE: Replacing Redundant Experts with Lightweight Novices for Structured Pruning of MoE
Geng Zhang (National University of Singapore), Yang You (National University of Singapore)
Computational EfficiencyMixture of ExpertsTextBenchmark
🎯 What it does: This paper proposes a new Mixture-of-Novices-and-Experts (MoNE) expert pruning method, which replaces redundant experts with lightweight novices requiring no computational overhead using a small amount of calibration data, thereby significantly compressing the MoE model while maintaining high performance.
Monitoring Decomposition Attacks with Lightweight Sequential Monitors
Chen Yueh-Han (New York University), He He (New York University)
Anomaly DetectionSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextMultimodalityChain-of-Thought
🎯 What it does: A lightweight sequential monitor was researched and implemented to detect and prevent decomposition attacks on LLM agents, and the DecomposedHarm dataset containing 4,634 task pairs was constructed.
Mono4DGS-HDR: High Dynamic Range 4D Gaussian Splatting from Alternating-exposure Monocular Videos
Jinfeng Liu (Hong Kong University of Science and Technology), Dan Xu (Hong Kong University of Science and Technology)
GenerationGaussian SplattingOptical FlowVideoBenchmark
🎯 What it does: Reconstruct 4D HDR scenes from monocular alternating exposure LDR videos without pose information, supporting HDR and LDR video rendering from arbitrary viewpoints.
Monocular Normal Estimation via Shading Sequence Estimation
Zongrui Li (Nanyang Technological University), Song Bai (ByteDance)
GenerationConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelAuto EncoderImageVideo
🎯 What it does: Propose the RoSE method, which transforms monocular normal estimation into shadow sequence estimation, generating shadow sequences using a video generation model and solving for normals via OLS.
Monotone Near-Zero-Sum Games: A Generalization of Convex-Concave Minimax
Ruichen Luo (IST Austria), Krishnendu Chatterjee (IST Austria)
OptimizationTabular
🎯 What it does: Defined a new class of single-player monotone near-zero-sum games, proposed the Iterative Coupled Linearization (ICL) algorithm, and proved superior gradient complexity compared to traditional variational inequality methods in this class of games; subsequently applied this framework to regularized matrix games and competitive games with small incentives, demonstrating faster convergence speeds.
MoRA: Missing Modality Low-Rank Adaptation for Visual Recognition
Shu Zhao (Pennsylvania State University), Vijaykrishnan Narayanan (Pennsylvania State University)
RecognitionSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Propose a parameter-efficient fine-tuning method called MoRA for handling missing modalities in visual recognition tasks.
MoRA: Mobility as the Backbone for Geospatial Representation Learning at Scale
Ya Wen (University of Hong Kong), Yulun Zhou (University of Hong Kong)
Representation LearningGraph Neural NetworkContrastive LearningImageMultimodalityGraph
🎯 What it does: Construct a geospatial representation learning framework named MoRA centered on human mobility graphs, integrating multimodal data including POI, satellite imagery, and demographic statistics, and outputting a unified 128-dimensional regional embedding;
Mordal: Automated Pretrained Model Selection for Vision Language Models
Shiqi He (University of Michigan), Mosharaf Chowdhury (University of Michigan)
Representation LearningHyperparameter SearchVision Language ModelMultimodality
🎯 What it does: Under aligned data for downstream tasks, automatically identify the optimal combination of pre-trained vision encoders and language models to construct the best Vision-Language Model (VLM).
More Than What Was Chosen: LLM-based Explainable Recommendation Beyond Noisy User Preferences
Chung Park (SK Telecom), Jaegul Choo (Korea Advanced Institute of Science & Technology)
Recommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose the C-APO framework, which simultaneously models revealed preferences (RP) and coherent preferences (CP) in recommendation systems, and jointly optimizes recommendations and explanations through conflict-aware adaptive weights.
More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models
Xinyu Tian (Australian National University), Jing Zhang (Australian National University)
Reinforcement LearningVision Language ModelMultimodality
🎯 What it does: Investigate the dual effects of multi-modal reasoning, discovering that long reasoning leads to visual forgetting, which in turn causes perceptual errors, and propose the VAPO method that guides reasoning through visual anchors.
MoReBench: Evaluating Procedural and Pluralistic Moral Reasoning in Language Models, More than Outcomes
Yu Ying Chiu (University of Washington), Sydney Levine (New York University)
Large Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Constructed and publicly released two novel moral reasoning benchmarks—MORE BENCH (1,000 scenarios, 23,018 human-written evaluation criteria) and MORE BENCHTHEORY (150 scenarios, covering five ethical frameworks)—and evaluated the reasoning quality of large models' Chain of Thought (CoT) and final answers using an LLM judge.
MoSA: Mosaic Shared Adaptation of Large Language Models
Xiequn Wang (Southern University of Science and Technology), Yu Zhang (Southern University of Science and Technology)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a new parameter-efficient fine-tuning method called Mosaic Shared Adaptation (MoSA), which performs full-rank updates on pre-trained model weights through randomized shared scalars;
MoSA: Motion-Coherent Human Video Generation via Structure-Appearance Decoupling
Haoyu Wang (Peking University), Shiliang Zhang (Peking University)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelAuto EncoderVideoText
🎯 What it does: Propose a decoupled framework called MoSA that splits human video generation into structure generation and appearance generation. First, a 3D structure transformer generates human motion structure based on text, then high-fidelity videos are synthesized under the guidance of this structure;
MOSAIC: Multi-Subject Personalized Generation via Correspondence-Aware Alignment and Disentanglement
Dong She (ByteDance), Jidong Jiang (ByteDance)
GenerationTransformerDiffusion modelImageBenchmark
🎯 What it does: Proposed the MOSAIC framework for multi-agent personalized image generation, addressing the challenges of identity preservation and semantic consistency.
MOSS: Efficient and Accurate FP8 LLM Training with Microscaling and Automatic Scaling
Yu Zhang (Chinese University of Hong Kong), Bei Yu (Chinese University of Hong Kong)
OptimizationComputational EfficiencyTransformerText
🎯 What it does: Proposed an efficient and accurate FP8 training framework called MOSS, which employs two-level micro-scale quantization and automatic scaling technology, achieving FP8 training for a 7B parameter model on 8 Hopper GPUs while maintaining the same training effectiveness as BF16.
Motion Prior Distillation in Time Reversal Sampling for Generative Inbetweening
Wooseok Jeon (Yonsei University), Hae-Gon Jeon (Yonsei University)
GenerationKnowledge DistillationConvolutional Neural NetworkDiffusion modelVideo
🎯 What it does: This paper proposes Motion Prior Distillation (MPD), a no-training sampling technique used only during inference, to address the bidirectional path motion prior conflicts in time-reversed sampling, thereby enhancing the spatiotemporal coherence of generated interpolated videos.
Motion-Aligned Word Embeddings for Text-to-Motion Generation
Ke Han (University of Trento), Nicu Sebe (University of Trento)
GenerationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextMultimodality
🎯 What it does: Designed and implemented the MATE framework, aligning vocabulary with human motion semantics by fine-tuning the embedding layer of large language models, thereby improving word-level understanding and generation quality in text-to-motion tasks.