NeurIPS 2025 Papers — Page 29
Conference on Neural Information Processing Systems · 5275 papers
MindOmni: Unleashing Reasoning Generation in Vision Language Models with RGPO
Yicheng Xiao (Tsinghua University), Ying Shan (The University of Hong Kong)
GenerationTransformerLarge Language ModelReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: MindOmni is proposed, a unified multimodal large language model that combines reinforcement learning to achieve reasoning generation capabilities, and has advantages in tasks such as image understanding, text generation, and visual editing.
MINGLE: Mixture of Null-Space Gated Low-Rank Experts for Test-Time Continual Model Merging
Zihuan Qiu (University of Electronic Science and Technology of China), Hongliang Li (University of Electronic Science and Technology of China)
Mixture of ExpertsMultimodality
🎯 What it does: A framework for continuous model merging using a small number of unlabeled samples during the inference phase (TTCMM) is proposed, along with the implementation scheme MINGLE.
miniF2F-Lean Revisited: Reviewing Limitations and Charting a Path Forward
Azim Ospanov (Huawei Hong Kong Research Center), Roozbeh Yousefzadeh (Chinese University of Hong Kong)
Large Language ModelTextBenchmark
🎯 What it does: This paper presents miniF2F-v2, correcting and validating over 300 errors and simplifications in the original miniF2F dataset, and constructing a complete automated reasoning pipeline from natural language to Lean proofs, along with an evaluation of its performance.
Minimal Semantic Sufficiency Meets Unsupervised Domain Generalization
Tan Pan (Fudan University), Mahsa Baktashmotlagh (Shanghai Academy of Artificial Intelligence for Science)
Domain AdaptationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes a domain generalization method called MS-UDG that does not require domain labels, capable of separating semantic information from variation factors and obtaining a minimal sufficient semantic representation.
Minimax Adaptive Online Nonparametric Regression over Besov spaces
Paul Liautaud (Sorbonne Université), Olivier Wintenberger (Sorbonne Université)
🎯 What it does: In the adversarial online regression framework, a wavelet-based adaptive algorithm is proposed, which can achieve competitive learning for any competitive function in the Besov space B^{s}_{p,q}(X) without prior knowledge of the Besov space parameters (s, p, q), and further constructs a locally adaptive version in space.
Minimax-Optimal Univariate Function Selection in Sparse Additive Models: Rates, Adaptation, and the Estimation-Selection Gap
Shixiang Liu (Renmin University of China)
Tabular
🎯 What it does: This paper studies the univariate function selection problem in Sparse Additive Models (SpAM), establishing the minimum maximum separation rate from the perspective of sparse multiple testing and support recovery, and proposes an adaptive selection procedure.
MiniMax-Remover: Taming Bad Noise Helps Video Object Removal
Bojia Zi (Chinese University of Hong Kong), Kam-Fai Wong (Chinese University of Hong Kong)
Object DetectionSegmentationOptimizationTransformerFlow-based ModelRectified FlowContrastive LearningVideo
🎯 What it does: Proposes the MiniMax-Remover two-stage video object removal framework, where the first stage uses a lightweight DiT model to remove text input and employs contrastive condition tokens; the second stage enhances robustness against adversarial noise through min-max optimization, ultimately generating high-quality videos in just 6 steps without CFG.
Minimizing False-Positive Attributions in Explanations of Non-Linear Models
Anders Gjølbye (Technical University of Denmark), Lars Kai Hansen (Technical University of Denmark)
Explainability and InterpretabilityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A method named PatternLocal is proposed, which can suppress erroneous positive attribution caused by confounding variables in local explanations of nonlinear models.
Minimum Width for Deep, Narrow MLP: A Diffeomorphism Approach
Geonho Hwang (Gwangju Institute for Science and Technology)
🎯 What it does: This study investigates the minimum width of deep narrow MLPs and provides optimal upper and lower bounds using differential homeomorphism methods.
MINT-CoT: Enabling Interleaved Visual Tokens in Mathematical Chain-of-Thought Reasoning
Xinyan Chen (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
Reinforcement LearningVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: The MINT-CoT method is proposed, which dynamically inserts fine-grained visual tokens during the chain of thought (CoT) process to enhance multimodal mathematical reasoning capabilities.
Mint: A Simple Test-Time Adaptation of Vision-Language Models against Common Corruptions
Wenxuan Bao (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
Domain AdaptationOptimizationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: We propose Mint, an algorithm that adapts only during testing on LayerNorm parameters, maximizing the inter-class variance of pseudo-labels using cumulative means and gradient accumulators to enhance the robustness of CLIP on corrupted data.
MIP against Agent: Malicious Image Patches Hijacking Multimodal OS Agents
Lukas Aichberger (Johannes Kepler University Linz), Adel Bibi (University of Oxford)
Adversarial AttackVision Language ModelImageMultimodality
🎯 What it does: This paper studies attacks on multimodal operating system agents and proposes a novel attack vector that hijacks OS agents by embedding malicious image patches (MIPs) in screenshots. The transferability and robustness of this method are validated under different prompts, screen layouts, agent components, and execution steps.
MIRA: Medical Time Series Foundation Model for Real-World Health Data
Hao Li (Peking University), Jiang Bian (Peking University)
TransformerMixture of ExpertsTime SeriesBiomedical DataElectronic Health RecordsOrdinary Differential Equation
🎯 What it does: MIRA is proposed, a foundational model specifically designed for medical time series, capable of continuous time prediction under irregular sampling and frequency diversity conditions;
MIRAGE: Assessing Hallucination in Multimodal Reasoning Chains of MLLM
Bowen Dong (Harbin Institute of Technology), Lei Zhang (Hong Kong Polytechnic University)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: Proposes the MIRAGE benchmark, specifically designed to evaluate hallucinations in multimodal reasoning chains, and provides multi-level evaluation metrics; simultaneously introduces the Logos method, which reduces reasoning hallucinations and improves accuracy through curriculum-based reinforcement training and collaborative prompt reasoning.
MISA: Memory-Efficient LLMs Optimization with Module-wise Importance Sampling
Yuxi Liu (Peking University), Kun Yuan (Peking University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A modular importance sampling method MISA has been developed for efficiently training large language models in memory-constrained environments.
MisoDICE: Multi-Agent Imitation from Mixed-Quality Demonstrations
The Viet Bui (Singapore Management University), Thanh Hong Nguyen (University of Oregon)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningSequential
🎯 What it does: A two-stage framework is proposed, where LLM and O-MAPL are first used for stepwise labeling of unlabeled mixed-quality demonstrations, followed by the implementation of MisoDICE for multi-agent offline imitation learning.
Miss-ReID: Delivering Robust Multi-Modality Object Re-Identification Despite Missing Modalities
Ruida Xi (Xidian University)
RecognitionRetrievalTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: The Miss-ReID framework is proposed, achieving robust re-identification of targets in the case of multimodal missing (where certain modalities may be missing during both training and inference).
Missing Data Imputation by Reducing Mutual Information with Rectified Flows
Jiahao Yu (University of Cambridge), Song Liu (University of Bristol)
Flow-based ModelRectified FlowImageTabularOrdinary Differential Equation
🎯 What it does: A missing value imputation method called MIRI is proposed, which iteratively reduces the mutual information between the filled data and the missing mask.
Mitigating Forgetting in LLM Fine-Tuning via Low-Perplexity Token Learning
Chao-Chung Wu (Appier AI Research), Hung-yi Lee (National Taiwan University)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper studies how using data generated by LLM during fine-tuning can effectively reduce forgetting of non-target tasks, and proposes a Selective Token Masking (STM) method to achieve the same effect by masking high perplexity tokens.
Mitigating Hallucination in VideoLLMs via Temporal-Aware Activation Engineering
Jianfeng Cai (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
Vision Language ModelVideo
🎯 What it does: To address the hallucination problem in VideoLLMs, an activation engineering framework based on video temporal variation features is proposed, which reduces the hallucination rate during the inference phase without additional fine-tuning.
Mitigating Hallucination Through Theory-Consistent Symmetric Multimodal Preference Optimization
Wenqi Liu (Shandong University), Liqiang Nie (Harbin Institute of Technology)
Recommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Proposes Symmetric Multimodal Preference Optimization (SymMPO), which reduces hallucinations in multimodal large language models (MLLMs) through symmetric preference learning and preference margin consistency regularization;
Mitigating Instability in High Residual Adaptive Sampling for PINNs via Langevin Dynamics
Minseok Jeong (Gwangju Institute of Science and Technology), Euiseok Hwang (Gwangju Institute of Science and Technology)
TabularSequentialPhysics RelatedStochastic Differential Equation
🎯 What it does: This paper proposes an adaptive sampling method based on Langevin dynamics (LAS) to improve the training stability and convergence speed of Physics-Informed Neural Networks (PINNs).
Mitigating Intra- and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models
Xiwen Wei (University of Texas at Austin), Radu Marculescu (University of Texas at Austin)
GenerationKnowledge DistillationRepresentation LearningMixture of ExpertsImageTextMultimodality
🎯 What it does: The Modality-Decoupled Experts (MoDE) framework is proposed to address the intra-modal and inter-modal catastrophic forgetting issues in Unified Multimodal Generative Models (UMGM) during continual instruction fine-tuning.
Mitigating Occlusions in Virtual Try-On via A Simple-Yet-Effective Mask-Free Framework
Chenghu Du (Wuhan University of Technology), Shili Xiong (Shanghai Artificial Intelligence Laboratory)
Image TranslationGenerationDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: A simple and effective mask-free virtual try-on framework is proposed to address inherent and acquired occlusion issues.
Mitigating Overthinking in Large Reasoning Models via Manifold Steering
Yao Huang (Beihang University), Yinpeng Dong (Tsinghua University)
Large Language ModelText
🎯 What it does: This paper studies the phenomenon of overthinking in large reasoning models (LRM) during the reasoning process through mechanism interpretability methods, and proposes an intervention strategy based on low-dimensional activation subspace—Manifold Steering—to suppress overthinking.
Mitigating Reward Over-optimization in Direct Alignment Algorithms with Importance Sampling
Nguyen Minh Phuc, Khoa D Doan
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: This paper proposes an Importance Sampling Direct Alignment Algorithm (IS-DAAs), which approximates the KL regularization in online RLHF by multiplying the importance ratio in the objective of direct alignment algorithms (such as DPO), thereby alleviating the issue of reward over-optimization that arises during offline alignment.
Mitigating Semantic Collapse in Partially Relevant Video Retrieval
WonJun Moon (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)
RetrievalKnowledge DistillationTransformerContrastive LearningVideoText
🎯 What it does: To address the semantic collapse problem that occurs in some relevant video retrieval tasks, two techniques are proposed: Text-Relation Preserving Learning (TCPL) and Cross-Branch Video Alignment (CBVA), which utilize token merging to further enhance semantic distinguishability.
Mitigating Sexual Content Generation via Embedding Distortion in Text-conditioned Diffusion Models
Jaesin Ahn (Kyungpook National University), Heechul Jung (Kyungpook National University)
GenerationDiffusion modelText
🎯 What it does: A defense framework based on a text encoder is proposed—Distorting Embedding Space (DES), which prevents diffusion models from generating pornographic content by distorting the unsafe embedding space.
Mitigating Spurious Features in Contrastive Learning with Spectral Regularization
Naghmeh Ghanooni (RPTU Kaiserslautern), Marius Kloft (RPTU Kaiserslautern)
Anomaly DetectionRepresentation LearningContrastive LearningImage
🎯 What it does: A spectral regularization framework is proposed to suppress reliance on spurious features during the unlabeled self-supervised pre-training phase, enhancing the diversity and robustness of representations.
Mitigating the Privacy–Utility Trade-off in Decentralized Federated Learning via f-Differential Privacy
Xiang Li (University of Pennsylvania), Weijie J Su
Federated LearningSafty and PrivacyImageTabular
🎯 What it does: Two privacy frameworks based on f-DP are proposed in decentralized federated learning (FL) — Pairwise Network f-DP (PNf-DP) and Secret-based f-Local DP (Secf-LDP). Precise privacy budgets and privacy-performance trade-off analyses are provided for two typical algorithms (random walk DP-SGD and DecoR with correlated noise).
Mitra: Mixed Synthetic Priors for Enhancing Tabular Foundation Models
Xiyuan Zhang (Amazon), Bernie Wang (Amazon)
ClassificationData SynthesisTransformerTabular
🎯 What it does: This paper systematically studies synthetic priors and proposes and constructs a multivariate prior mixture that includes SCM and tree models for pre-training the Tabular Foundation Model (TFM) MITRA, demonstrating its state-of-the-art (SOTA) performance on multiple classification and regression benchmarks.
Mix Data or Merge Models? Balancing the Helpfulness, Honesty, and Harmlessness of Large Language Model via Model Merging
Jinluan Yang (Zhejiang University), Kun Kuang (Zhejiang University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: The first 3H (Helpfulness, Honesty, Harmlessness) balanced optimization benchmark is proposed, and a systematic comparison of data mixing and model fusion methods is conducted.
MIX: A Multi-view Time-Frequency Interactive Explanation Framework for Time Series Classification
Viet-Hung Tran (Queen's University Belfast), Son T. Mai
ClassificationExplainability and InterpretabilityRecurrent Neural NetworkTransformerTime SeriesElectrocardiogram
🎯 What it does: A multi-perspective time-frequency interaction interpretation framework, MIX, is proposed, which utilizes Haar wavelet decomposition to generate multi-resolution perspectives and explains the decisions of deep time series classification models through interactive cross-perspective refinement, keystone-first IG, and greedy feature selection.
MixAT: Combining Continuous and Discrete Adversarial Training for LLMs
Csaba Dékány (INSAIT), Martin Vechev (ETH Zurich)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper proposes a novel adversarial training method called MIXAT, which combines continuous and discrete perturbations to enhance the robustness of large language models.
Mixed-Sample SGD: an End-to-end Analysis of Supervised Transfer Learning
Yuyang Deng (Columbia University), Samory Kpotufe (Columbia University)
Domain AdaptationOptimizationSupervised Fine-TuningTabular
🎯 What it does: Designed and analyzed a hybrid sampling SGD algorithm for supervised transfer learning, which can adaptively select the source/target sample ratio while ensuring convergence and generalization.
Mixing Expert Knowledge: Bring Human Thoughts Back To the Game of Go
Yichuan Ma (Fudan University), Kai Chen (Shanghai AI Laboratory)
Large Language ModelReinforcement LearningMixture of ExpertsTextChain-of-Thought
🎯 What it does: By mixing general LLMs with structured Go expert data, combined with long-chain thinking data for cold start fine-tuning, and then using GRPO reinforcement learning for self-exploration, the LoGos model is ultimately launched, which possesses professional Go level while maintaining general reasoning abilities.
MixPrompt: Efficient Mixed Prompting for Multimodal Semantic Segmentation
Zhiwei Hao (Beijing Institute of Technology), Dan Zeng (Shanghai University)
SegmentationPrompt EngineeringMultimodality
🎯 What it does: The MixPrompt framework is proposed, which integrates auxiliary modalities into a pre-trained RGB segmentation model through a lightweight prompting module, achieving efficient multi-modal semantic segmentation.
MixSignGraph: A Sign Sequence is Worth Mixed Graphs of Nodes
Shiwei Gan (Nanjing University), Hongkai Wen (University of Warwick)
RecognitionImage TranslationGraph Neural NetworkVideoText
🎯 What it does: This paper proposes MixSignGraph, which enhances the performance of gesture recognition and translation models by combining cross-region graph convolution and hierarchical graph convolution. It also introduces the TCTC pre-training scheme to improve translation effectiveness in the absence of gloss annotations.
Mixture of Inputs: Text Generation Beyond Discrete Token Sampling
Yufan Zhuang (University of California San Diego), Jianfeng Gao (Microsoft Research)
GenerationTransformerLarge Language ModelText
🎯 What it does: The Mixture of Inputs (MOI) method is proposed, which mixes the sampled discrete tokens with their corresponding probability distributions into continuous inputs during autoregressive generation, preserving the rich information of the model's predictive distribution; it can be used directly in the existing LLM inference process without additional training.
Mixture of Noise for Pre-Trained Model-Based Class-Incremental Learning
Kai Jiang (Northwestern Polytechnical University), Xuelong Li (China Telecom)
ClassificationTransformerSupervised Fine-TuningImage
🎯 What it does: A forward noise mixing method (MIN) for class-incremental learning on pre-trained models is proposed, which alleviates forgetting caused by parameter drift by learning task-specific noise and mixing it in intermediate features.
Mixture of Scope Experts at Test: Generalizing Deeper Graph Neural Networks with Shallow Variants
Gangda Deng (University of Southern California), Viktor Prasanna (University of Southern California)
Graph Neural NetworkMixture of ExpertsGraph
🎯 What it does: A framework for mixing different depth GNN experts during the testing phase (Moscat) is proposed, which enhances the generalization ability of deep GNNs on heterogeneous graphs by decoupling expert training and gating learning.
Mixture-of-Experts Meets In-Context Reinforcement Learning
Wenhao Wu (Nanjing University), Zhi Wang (Nanjing University)
Robotic IntelligenceTransformerReinforcement LearningMixture of ExpertsContrastive LearningMultimodality
🎯 What it does: A framework is proposed that introduces Token-wise and Task-wise two-layer Mixture-of-Experts (MoE) in In-Context Reinforcement Learning (ICRL), enhancing the ability to handle multimodal inputs (state, action, reward) and task diversity.
Mixture-of-Experts Operator Transformer for Large-Scale PDE Pre-Training
Hong Wang (University of Science and Technology of China), Yan Jiang
TransformerMixture of ExpertsPhysics Related
🎯 What it does: A pre-training model for PDE based on a sparse Mixture-of-Experts (MoE) Transformer, named MoE-POT, is proposed, which can expand model capacity without increasing inference costs.
Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation
Sangmin Bae (KAIST), Se-Young Yun (KAIST)
Computational EfficiencyTransformerMixture of ExpertsText
🎯 What it does: A Mixture-of-Recursions (MoR) framework is proposed, which efficiently combines parameter sharing, token-level adaptive recursion depth, and KV caching to achieve a lighter Transformer;
Mixtures of Subspaces for Bandwidth Efficient Context Parallel Training
Sameera Ramasinghe (Pluralis Research), Alexander Long (Pluralis Research)
TransformerText
🎯 What it does: By jointly learning low-rank subspace projection and rotation parameters, efficient context-parallel Transformer training is achieved in low-bandwidth decentralized environments.
MJ-Video: Benchmarking and Rewarding Video Generation with Fine-Grained Video Preference
Haibo Tong (University of Illinois Urbana-Champaign), Huaxiu Yao (University of North Carolina at Chapel Hill)
GenerationRecommendation SystemReinforcement LearningMixture of ExpertsVideoBenchmark
🎯 What it does: A large video preference benchmark, MJ-BENCH-VIDEO, is proposed, covering 5 dimensions and 28 fine-grained evaluation criteria. Based on this, a video reward model, MJ-VIDEO, with a Mixture-of-Experts structure is developed for fine-grained and interpretable preference assessment.
MLE-STAR: Machine Learning Engineering Agent via Search and Targeted Refinement
Jaehyun Nam (KAIST), Tomas Pfister (Google Cloud)
OptimizationAI Code AssistantTransformerLarge Language ModelAgentic AIImageTextMultimodalityTabularRetrieval-Augmented GenerationAudio
🎯 What it does: A machine learning engineering agent named MLE-STAR has been developed, which utilizes LLMs and search engines to first retrieve task-related models to generate initial code. It then locates key code blocks through ablation analysis for deep iterative refinement and introduces an automated integration strategy to further enhance performance, suitable for multimodal tasks (tables, images, text, audio).
MLEP: Multi-granularity Local Entropy Patterns for Generalized AI-generated Image Detection
Lin Yuan (Chongqing University of Posts and Telecommunications), Xinbo Gao (Chongqing University of Posts and Telecommunications)
Object DetectionConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A method for detecting AI-generated images based on Multi-Scale Local Entropy Patterns (MLEP) is proposed.
MLLM-For3D: Adapting Multimodal Large Language Model for 3D Reasoning Segmentation
Jiaxin Huang (MBZUAI), Tongliang Liu (MBZUAI)
SegmentationTransformerLarge Language ModelMultimodalityPoint Cloud
🎯 What it does: The MLLM-For3D framework is proposed, which combines a pre-trained two-dimensional multimodal large language model (MLLM) with SAM to generate multi-view pseudo-segmentation masks and project them into three-dimensional space, achieving label-free three-dimensional inference segmentation.
MLLMs Need 3D-Aware Representation Supervision for Scene Understanding
Xiaohu Huang (University of Hong Kong), Kai Han (University of Hong Kong)
Object DetectionKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningImageMultimodality
🎯 What it does: Proposes the 3DRS framework, which utilizes a pre-trained 3D foundational model to provide 3D perception supervision for multimodal large language models (MLLMs), enhancing their scene understanding capabilities.
MLZero: A Multi-Agent System for End-to-end Machine Learning Automation
Haoyang Fang (Amazon Web Services), George Karypis (Amazon Web Services)
Large Language ModelAgentic AIImageTextMultimodalityTabularBenchmark
🎯 What it does: A multi-agent framework called MLZero is proposed, achieving end-to-end automation from raw multimodal data to a complete machine learning model with zero human intervention.
MM-Agent: LLM as Agents for Real-world Mathematical Modeling Problem
Fan Liu (Hong Kong University of Science and Technology), Hao Liu (Hong Kong University of Science and Technology)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Proposes the MM-Bench benchmark and the MM-Agent framework to achieve end-to-end automation for open mathematical modeling problems.
MMaDA: Multimodal Large Diffusion Language Models
Ling Yang (Princeton University), Mengdi Wang (Princeton University)
GenerationTransformerLarge Language ModelReinforcement LearningDiffusion modelTextMultimodalityChain-of-Thought
🎯 What it does: MMaDA is proposed—a unified multimodal diffusion language model that can simultaneously perform text reasoning, multimodal understanding, and text-to-image generation;
MoBA: Mixture of Block Attention for Long-Context LLMs
Enzhe Lu (Moonshot AI), Jiezhong Qiu (Hangzhou Institute of Medicine)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Proposes Mixture of Block Attention (MoBA), which achieves sparsification of long sequence attention by dynamically routing to select historical KV blocks, and supports unstructured, switchable sparse and full attention.
MobileODE: An Extra Lightweight Network
Le Yu (Sichuan University), Tao He (Sichuan University)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerImageOrdinary Differential Equation
🎯 What it does: A lightweight network called MobileODE is proposed, which utilizes discrete neural ODEs to replace the 1×1 convolution of MobileNet with a learnable Channelwise ODE Solver (COS).
MobileUse: A Hierarchical Reflection-Driven GUI Agent for Autonomous Mobile Operation
Ning Li (Shanghai Jiao Tong University), Weinan Zhang (Shanghai Jiao Tong University)
Robotic IntelligenceTransformerLarge Language ModelAgentic AIMultimodality
🎯 What it does: A GUI mobile agent named MobileUse is proposed, achieving robust execution and error recovery for long-sequence tasks through hierarchical reflection and active exploration.
MOBO-OSD: Batch Multi-Objective Bayesian Optimization via Orthogonal Search Directions
Lam Ngo (RMIT University), Hongyu Zhang (Chongqing University)
OptimizationTabularBenchmark
🎯 What it does: A batch multi-objective Bayesian optimization algorithm MOBO-OSD based on orthogonal search directions is proposed to efficiently discover the multi-objective Pareto front under a limited evaluation budget.
MoCha: Towards Movie-Grade Talking Character Generation
Cong Wei (University of Waterloo), Wenhu Chen (University of Waterloo)
GenerationData SynthesisTransformerDiffusion modelVideoTextMultimodalityAudio
🎯 What it does: The MoCha model is proposed, achieving the generation of complete conversational character videos solely through voice and text.
Modality-Aware SAM: Sharpness-Aware-Minimization Driven Gradient Modulation for Harmonized Multimodal Learning
Hossein R. Nowdeh, Fatemeh Afghah (Clemson University)
OptimizationMultimodality
🎯 What it does: A modality-aware Sharpness-Aware Minimization (M-SAM) optimizer has been developed to dynamically balance the gradient updates of the dominant modality and other modalities in multimodal learning.
Model Editing for Vision Transformers
Xinyi Huang (Independent Researcher), Long-Kai Huang (Hong Kong Baptist University)
ClassificationRecognitionTransformerImage
🎯 What it does: A two-stage model editing framework called RefineViT is proposed for visual Transformers, which first identifies the attention heads that cause errors and then corrects the representations through projection matrices, thereby rectifying errors while maintaining predictions for other samples.
Model Inversion with Layer-Specific Modeling and Alignment for Data-Free Continual Learning
Ruilin Tong (University of New South Wales), Dong Gong (University of New South Wales)
GenerationData SynthesisOptimizationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A hierarchical model inversion (PMI) is proposed, combining feature distribution modeling and alignment to achieve synthetic data generation and replay in data-free continual learning, avoiding the storage of historical data.
Model Merging in Pre-training of Large Language Models
Yunshui Li (ByteDance), Yonghui Wu
OptimizationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes a framework for model merging during the pre-training phase of large language models—Pre-trained Model Averaging (PMA)—and experimentally verifies that it can significantly improve model performance and reduce training costs.
Model Provenance Testing for Large Language Models
Ivica Nikolic, Prateek Saxena (National University of Singapore)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a method for testing the source of large language models that can determine whether a model is fine-tuned from a certain parent model using only black-box API queries.
Model Reconciliation via Cost-Optimal Explanations in Probabilistic Logic Programming
Yinxu Tang (Washington University in St. Louis), William Yeoh (Washington University in St. Louis)
OptimizationExplainability and InterpretabilityTabular
🎯 What it does: This paper proposes a model reconciliation method within the framework of the probabilistic logic program ProbLog, addressing the inconsistency between agent and human models in terms of the most probable explanation (MPE) probabilities by generating cost-optimal explanations.
Model Selection for Off-policy Evaluation: New Algorithms and Experimental Protocol
Pai Liu (University of Illinois), Nan Jiang (University of Illinois)
Reinforcement Learning
🎯 What it does: This paper studies how to perform model selection for different OPE (Offline Policy Evaluation) methods in offline reinforcement learning, proposing a new selector based on LSTD called LSTD-Tournament, as well as several model-based selection methods, and designing a controllable and stable experimental protocol.
MODEL SHAPLEY: Find Your Ideal Parameter Player via One Gradient Backpropagation
Xu Chu (Peking University), Junfeng Zhao (Big Data Technology Research Center)
CompressionOptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper proposes MODEL SHAPLEY, which uses Shapley values to measure the importance of parameters in large neural networks at the parameter level, and achieves a scalable closed-form approximation through a single gradient backpropagation.
Model-Based Policy Adaptation for Closed-Loop End-to-end Autonomous Driving
Haohong Lin (Carnegie Mellon University), Ding Zhao (Stanford University)
Autonomous DrivingDiffusion modelGenerative Adversarial NetworkGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes a Model-based Policy Adaptation (MPA) framework that enhances the robustness and safety of pre-trained end-to-end (E2E) driving models in closed-loop autonomous driving by generating adversarial trajectory data, a diffusion policy adapter, and a multi-principle Q-value model.
Model-Guided Dual-Role Alignment for High-Fidelity Open-Domain Video-to-Audio Generation
Kang Zhang (Korea Advanced Institute of Science and Technology), Joon Son Chung (Ulsan National Institute of Science and Technology)
GenerationTransformerFlow-based ModelVideoMultimodalityAudio
🎯 What it does: This paper presents MGAudio, a video-to-audio generation framework based on flow-matching Transformer, which enhances the quality and synchronization of audio through model guidance and dual-role alignment.
Model-Informed Flows for Bayesian Inference
Joohwan Ko (University of Massachusetts Amherst), Justin Domke (University of Massachusetts Amherst)
OptimizationFlow-based ModelTabular
🎯 What it does: The paper proposes a Model Information Flow (MIF) that combines Variationally Inferred Parameters (VIP) with Forward Autoregressive Flow (FAF) to construct a new variational family that better approximates the posterior distribution of hierarchical Bayesian models.
Model–Behavior Alignment under Flexible Evaluation: When the Best-Fitting Model Isn’t the Right One
Itamar Avitan (Ben-Gurion University of Negev), Tal Golan (Ben-Gurion University of Negev)
ClassificationOptimizationImage
🎯 What it does: Conducted linear re-weight alignment on 20 pre-trained visual networks, simulated millions of behavioral data using the THINGS odd-one-out task, and carried out large-scale model recovery experiments to test whether prediction accuracy under linear alignment can truly identify generative models.
Modeling Cell Dynamics and Interactions with Unbalanced Mean Field Schrödinger Bridge
Zhenyi Zhang (Peking University), Peijie Zhou (Peking University)
Biomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A framework for the Unbalanced Mean Field Schrödinger Bridge (UMFSB) has been constructed, and a deep learning method called CytoBridge has been implemented to simultaneously learn the stochastic dynamics, proliferation/death, and intercellular interactions of cells from sparse temporal snapshot data.
Modeling Dynamic Neural Activity by combining Naturalistic Video Stimuli and Stimulus-independent Latent Factors
Finn Schmidt (University of Goettingen), Fabian H. Sinz (University of Goettingen)
Recurrent Neural NetworkVideo
🎯 What it does: This study constructs a probabilistic neural activity prediction model that can simultaneously utilize natural video stimuli and inferred latent brain states from neuronal activity to predict the complete response distribution of V1 neurons.
Modeling Microenvironment Trajectories on Spatial Transcriptomics with NicheFlow
Kristiyan Sakalyan (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
GenerationTransformerFlow-based ModelPoint CloudBiomedical Data
🎯 What it does: A flow-matching-based generative model NicheFlow is proposed to learn the co-evolutionary trajectory of spatial and gene expression in time-resolved spatial transcriptomics of cellular microenvironments.
Modeling Neural Activity with Conditionally Linear Dynamical Systems
Victor Geadah (Princeton University), Alex H Williams
OptimizationComputational EfficiencyTime Series
🎯 What it does: A conditional linear dynamical system (CLDS) model is proposed and implemented, using Gaussian process priors to smoothly vary the parameters of the linear dynamical system (LDS) with experimental conditions. By combining Kalman filtering and EM inference, it can explain the nonlinear dynamics of neural population activity with very little data.
Modeling the Economic Impacts of AI Openness Regulation
Tori Qiu (Carnegie Mellon University), Hoda Heidari (Carnegie Mellon University)
🎯 What it does: A game model has been constructed that includes general model developers and niche domain refiners to analyze the impact of AI openness regulation on openness and performance decisions.
Modelling the control of offline processing with reinforcement learning
Eleanor Spens (University of Oxford), Timothy Edward John Behrens (University of Oxford)
OptimizationMeta LearningReinforcement LearningImage
🎯 What it does: A meta-controller model is constructed to optimize task performance during the waking phase by selecting different offline learning actions (such as replaying memories, building world models, and generating simulated data) during the sleep phase.
MODEM: A Morton-Order Degradation Estimation Mechanism for Adverse Weather Image Recovery
Hainuo Wang (Tianjin University), Xiaojie Guo (Tianjin University)
RestorationImage
🎯 What it does: A degradation estimation mechanism based on Morton order (MODEM) is proposed for image restoration under various severe weather conditions.
ModHiFi: Identifying High Fidelity predictive components for Model Modification
Dhruva Kashyap (Indian Institute of Science), Chiranjib Bhattacharyya (Indian Institute of Science)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerImageText
🎯 What it does: The ModHiFi algorithm is proposed, which utilizes Subset Fidelity to identify high-fidelity (HiFi) components in models through synthetic data without training data or loss functions, thereby achieving structured pruning and class-level forgetting.
MoE-Gyro: Self-Supervised Over-Range Reconstruction and Denoising for MEMS Gyroscopes
Feiyang Pan (Southeast University), Guangbin Dou (Southeast University)
RestorationOptimizationMixture of ExpertsAuto EncoderTime SeriesBenchmark
🎯 What it does: Designed and implemented a self-supervised Mixture-of-Experts framework MoE-Gyro for over-range reconstruction and noise suppression of MEMS gyroscopes, addressing the challenges of traditional hardware upgrades.
MoEMeta: Mixture-of-Experts Meta Learning for Few-Shot Relational Learning
Han Wu (University of Sydney), Jie Yin (University of Sydney)
Meta LearningMixture of ExpertsGraph
🎯 What it does: Proposes the MoEMeta framework, which combines Mixture-of-Experts to learn global relationship prototypes and achieves local adaptation through task-specific projections, enhancing few-shot knowledge graph relation learning effectiveness.
MoESD: Unveil Speculative Decoding's Potential for Accelerating Sparse MoE
Zongle Huang (Tsinghua University), Tianyu Zhang (Huawei Noah's Ark Lab)
OptimizationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Explores the potential of sparse MoE models in Speculative Decoding (SD) acceleration at medium batch sizes;
MOF-BFN: Metal-Organic Frameworks Structure Prediction via Bayesian Flow Networks
Rui Jiao (Alibaba Group), Yang Liu (Tsinghua University)
GenerationOptimizationGraph Neural NetworkFlow-based ModelGraph
🎯 What it does: A framework for predicting MOF structures based on Bayesian Flow Networks (BFN), named MOF-BFN, is proposed, which can jointly predict lattice parameters, block-level fractional coordinates, and orientations (quaternions) in a fractional coordinate system, achieving end-to-end MOF structure generation and prediction.
MoFo: Empowering Long-term Time Series Forecasting with Periodic Pattern Modeling
Jiaming Ma (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
TransformerTime Series
🎯 What it does: A novel long-period time series forecasting framework MoFo is proposed, which explicitly models periodic correlations and trends using a periodic structure, significantly improving long-term forecasting performance.
MoGe-2: Accurate Monocular Geometry with Metric Scale and Sharp Details
Ruicheng Wang (University of Science and Technology of China), Jiaolong Yang (Microsoft Research)
RestorationDepth EstimationTransformerContrastive LearningImagePoint Cloud
🎯 What it does: Predicting geometric information for 3D reconstruction from monocular images, MoGe-2 is proposed, which can simultaneously achieve accurate relative geometry, true scale, and fine-grained details.
MokA: Multimodal Low-Rank Adaptation for MLLMs
Yake Wei (Renmin University of China), Di Hu (Renmin University of China)
CompressionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageVideoTextMultimodalityAudio
🎯 What it does: This paper studies a parameter-efficient fine-tuning method for multimodal large language models called MokA, aimed at improving multimodal task performance while keeping the model parameters unchanged.
Mol-LLaMA: Towards General Understanding of Molecules in Large Molecular Language Model
Dongki Kim (KAIST), Sung Ju Hwang (KAIST)
Explainability and InterpretabilityDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningTextPoint CloudGraph
🎯 What it does: Mol-LLaMA is proposed, a general molecular language model capable of understanding molecular structures, chemical and biological features, and possessing interpretability and reasoning abilities.
MoleBridge: Synthetic Space Projecting with Discrete Markov Bridges
Rongchao Zhang (Peking University), Hanpin Wang (Peking University)
GenerationDrug DiscoveryGraph Neural NetworkTransformerGraph
🎯 What it does: A semi-autoregressive model called MoleculeBridge based on Markov bridges is proposed to project non-synthesizable molecules into synthesizable suffix expressions, generating complete synthesis pathways.
MoME: Mixture of Matryoshka Experts for Audio-Visual Speech Recognition
Umberto Cappellazzo, Maja Pantic
RecognitionRepresentation LearningLarge Language ModelMixture of ExpertsVideoMultimodalityAudio
🎯 What it does: Designed the MoME framework, combining Matryoshka representation learning with sparse Mixture-of-Experts to achieve resource-adaptive inference for audio-visual speech recognition;
Moment- and Power-Spectrum-Based Gaussianity Regularization for Text-to-Image Models
Jisung Hwang (Korea Advanced Institute of Science and Technology), Minhyuk Sung (Korea Advanced Institute of Science and Technology)
GenerationOptimizationImageText
🎯 What it does: A unified Gaussian regularization framework is proposed, combining moment matching in the time domain with power spectrum matching in the frequency domain for latent space optimization in text-to-image models.
Momentum Multi-Marginal Schrödinger Bridge Matching
Panagiotis Theodoropoulos (Georgia Institute of Technology), Guan-Horng Liu (FAIR at Meta)
OptimizationTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes Momentum Multi-Marginal Schrödinger Bridge Matching (3MSBM), a trainable matching framework for learning smooth and consistent trajectories under sparse samples at multiple time points.
Momentum-SAM: Sharpness Aware Minimization without Computational Overhead
Marlon Becker (University of Muenster), Benjamin Risse (University of Muenster)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposes the Momentum-SAM (MSAM) optimizer, which imposes sharpness constraints on deep networks without increasing additional forward/backward propagation, achieving better generalization performance.
MonarchAttention: Zero-Shot Conversion to Fast, Hardware-Aware Structured Attention
Can Yaras (University of Michigan), Laura Balzano (University of Michigan)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelImageText
🎯 What it does: This paper proposes MonarchAttention, a zero-shot sub-quadratic complexity attention approximation method based on the Monarch structured matrix, and implements an efficient hardware-friendly version aimed at GPUs.
MoniTor: Exploiting Large Language Models with Instruction for Online Video Anomaly Detection
Shengtian Yang (Nanjing University of Aeronautics and Astronautics), Jie Qin (Nanjing University of Aeronautics and Astronautics)
Anomaly DetectionRecurrent Neural NetworkTransformerLarge Language ModelVision Language ModelVideo
🎯 What it does: This paper proposes MoniTor, an online, training-free video anomaly detection framework that utilizes large language models (LLM) and visual language models (VLM) to detect and locate anomalous events in real-time streaming video.
Monitoring Risks in Test-Time Adaptation
Mona Schirmer (University of Amsterdam), Eric Nalisnick (Johns Hopkins University)
Domain AdaptationAnomaly DetectionImageTime Series
🎯 What it does: A novel unsupervised risk monitoring framework is proposed to track and detect the decline in model performance during test-time adaptation (TTA), achieving online monitoring of risks during model adaptation by combining sequential testing and confidence sequences.
Monoculture or Multiplicity: Which Is It?
Mila Gorecki (Max Planck Institute for Intelligent Systems), Moritz Hardt (Max Planck Institute for Intelligent Systems)
Large Language ModelPrompt EngineeringText
🎯 What it does: This paper conducts a systematic empirical evaluation of whether large language models in decision ecosystems tend towards a monoculture or multiplicity.
MonoLift: Learning 3D Manipulation Policies from Monocular RGB via Distillation
Ziru Wang (State Grid Corporation of China), Jingdong Wang (Baidu)
Knowledge DistillationRobotic IntelligenceTransformerImage
🎯 What it does: During training, a three-layer knowledge distillation is used to transfer pseudo-depth information to a monocular RGB student model, enabling it to achieve 3D perception and precise control without using depth input.
Monotone and Separable Set Functions: Characterizations and Neural Models
SOUTRIK SARANGI, Abir De (Indian Institute of Technology Bombay)
TextPoint Cloud
🎯 What it does: This paper proposes a mapping function from sets to vectors that can simultaneously satisfy monotonicity and separability (MAS) and its neural network implementation MASNET, used for set inclusion detection and generalizing monotonic set functions.
MoodAngels: A Retrieval-augmented Multi-agent Framework for Psychiatry Diagnosis
Mengxi Xiao (Wuhan University), Min Peng (Wuhan University)
ClassificationData SynthesisRetrievalTransformerLarge Language ModelDiffusion modelTextBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: The MoodAngels multi-agent framework and the MoodSyn synthetic dataset are proposed to assist in the diagnosis of mood disorders (depression, bipolar disorder).
MoonCast: High-Quality Zero-Shot Podcast Generation
Zeqian Ju (University of Science and Technology of China), Xu Tan (Microsoft Research)
GenerationTransformerLarge Language ModelTextAudio
🎯 What it does: MoonCast provides a zero-shot, long-duration, two-speaker podcast generation system that can automatically generate podcast audio with a natural improvisational dialogue style from non-speech sources such as text, web pages, and PDFs.
MoORE: SVD-based Model MoE-ization for Conflict- and Oblivion-Resistant Multi-Task Adaptation
Shen Yuan (Renmin University of China), Hongteng Xu
Large Language ModelMixture of ExpertsText
🎯 What it does: A model MoE strategy based on Singular Value Decomposition (SVD) called MoORE is proposed for parameter-efficient adaptation of large-scale pre-trained models in multi-task scenarios.