NeurIPS 2025 Papers — Page 28
Conference on Neural Information Processing Systems · 5275 papers
Manipulating 3D Molecules in a Fixed-Dimensional E(3)-Equivariant Latent Space
Zitao Chen (Tsinghua University), Yanyan Lan (Tsinghua University)
OptimizationDrug DiscoveryGraph Neural NetworkFlow-based ModelAuto EncoderGraph
🎯 What it does: We propose MolFLAE, a VAE that can map 3D molecules to a fixed-dimensional, E(3) invariant latent space, and utilizes a Bayesian Flow Network for decoding, achieving zero-shot molecular editing and optimization.
Manipulating Feature Visualizations with Gradient Slingshots
Dilyara Bareeva (Fraunhofer Heinrich Hertz Institute), Kirill Bykov
Adversarial AttackImageOrdinary Differential Equation
🎯 What it does: This paper studies an attack method called Gradient Slingshots, which can manipulate the feature visualization results of deep neural networks to create arbitrary target images without changing the network structure or significantly degrading performance.
Many LLMs Are More Utilitarian Than One
Anita Keshmirian (Forward College), Lav R. Varshney (University of Illinois at Urbana-Champaign)
TransformerLarge Language ModelText
🎯 What it does: This study investigates whether large language models (LLMs) exhibit 'utilitarian enhancement' during moral reasoning in multi-agent (two or three person) discussions, comparing it to individual reasoning.
Many Minds, One Goal: Time Series Forecasting via Sub-task Specialization and Inter-agent Cooperation
Qihe Huang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
Graph Neural NetworkTransformerAgentic AITime SeriesFinance Related
🎯 What it does: A multi-agent time series forecasting framework called MAFS is proposed, utilizing multi-agent collaboration to address forecasting tasks with different time scales and signal characteristics.
MAP Estimation with Denoisers: Convergence Rates and Guarantees
Scott Pesme (Univ. Grenoble Alpes), Julien Mairal (Univ. Grenoble Alpes)
RestorationOptimization
🎯 What it does: An iterative denoising algorithm based on MMSE mean is proposed, and it is proven that under the prior conditions of being log-concave and having bounded third derivatives, it can converge to the proximal operator of the negative log prior, thus providing theoretical support for MAP estimation.
MAPLE: Multi-scale Attribute-enhanced Prompt Learning for Few-shot Whole Slide Image Classification
Junjie Zhou (Nanjing University of Aeronautics and Astronautics), Daoqiang Zhang (Nanjing University of Aeronautics and Astronautics)
ClassificationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImage
🎯 What it does: A multi-scale attribute-enhanced prompt learning framework called MAPLE is proposed for few-shot whole slide image classification.
Marginal-Nonuniform PAC Learnability
Steve Hanneke (Purdue University), Maximilian Thiessen (TU Wien)
🎯 What it does: This paper re-examines the classic non-uniform PAC learning model and proposes a variant of marginal non-uniform learning, investigating the complete characterization of learning rates.
Markov Persuasion Processes: Learning to Persuade From Scratch
Francesco Bacchiocchi (Politecnico di Milano), Nicola Gatti (Politecnico di Milano)
Reinforcement Learning
🎯 What it does: A learning algorithm OPPS based on the Markov Persuasion Process (MPP) is designed and analyzed in an unknown environment, capable of gradually learning information disclosure strategies to maximize long-term sender benefits while controlling the degree of violations in persuading the receiver, without prior transition, prior distribution, and reward information.
MARS: A Malignity-Aware Backdoor Defense in Federated Learning
Wei Wan (City University of Macau), Leo Yu Zhang (Griffith University)
Federated LearningImage
🎯 What it does: A federated learning backdoor defense framework named MARS is proposed, which identifies and filters backdoored models by measuring the backdoor energy of neurons.
Martingale Posterior Neural Networks for Fast Sequential Decision Making
Gerardo Duran-Martin (Oxford-Man Institute of Quantitative Finance), Kevin Patrick Murphy
Recommendation SystemOptimizationTabularSequential
🎯 What it does: A neural network based on the Mahalanobis posterior is proposed, utilizing first-order posterior prediction instead of parametric posterior for online learning and decision-making;
Martingale Score: An Unsupervised Metric for Bayesian Rationality in LLM Reasoning
Zhonghao He (University of Cambridge), Maarten Sap (Carnegie Mellon University)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes an unsupervised metric called Martingale Score to quantify whether there is belief entrenchment in large language models (LLMs) during the reasoning process, and uses this metric to evaluate the reasoning quality of different models, prompts, and reasoning techniques in real-world tasks.
Mask Image Watermarking
Runyi Hu (Nanyang Technological University), Tianwei Zhang (Nanyang Technological University)
Image TranslationRestorationData SynthesisConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: The MaskWM framework is proposed, supporting full-image and local watermark embedding, localization, and extraction, and achieving multi-watermark embedding.
Masked Diffusion Models as Energy Minimization
Sitong Chen (Renmin University of China), Chongxuan Li (Renmin University of China)
GenerationOptimizationLarge Language ModelDiffusion modelText
🎯 What it does: A theoretical framework is proposed that views the Masked Diffusion Model (MDM) as a discrete optimal transport problem, providing an equivalent formulation for energy minimization.
Masked Gated Linear Unit
Yukito Tajima (Institute of Science), Rio Yokota (Institute of Science)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Proposes Masked Gated Linear Units (MGLUs) and an efficient CUDA kernel FlashMGLU, utilizing shared weights and learnable binary masks to achieve separation of gates and value flows, reducing memory read and write.
MASTER: Enhancing Large Language Model via Multi-Agent Simulated Teaching
Liang Yue (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A MASTER multi-agent simulation teaching framework is proposed, utilizing interactions between teacher and student agents in three teaching scenarios: error correction, debate, and analogy reasoning, to generate a high-quality BOOST-QA instruction dataset, which is then used to fine-tune large language models.
MAT-Agent: Adaptive Multi-Agent Training Optimization
Jusheng Zhang (Sun Yat-sen University), Keze Wang (Sun Yat-sen University)
ClassificationOptimizationReinforcement LearningAgentic AIImage
🎯 What it does: A multi-agent framework called MAT-Agent is designed to dynamically adjust data augmentation, optimizers, learning rate scheduling, and loss functions, achieving real-time adaptive optimization for multi-label image classification training.
MATCH: Multi-faceted Adaptive Topo-Consistency for Semi-Supervised Histopathology Segmentation
Meilong Xu (Stony Brook University), Chao Chen (Stony Brook University)
SegmentationContrastive LearningImageBiomedical Data
🎯 What it does: A semi-supervised histopathological image segmentation framework is proposed, utilizing topological consistency in multiple perturbation predictions to enhance the topological accuracy of segmentation results.
Matching Markets Meet LLMs: Algorithmic Reasoning with Ranked Preferences
Hadi Hosseini (Penn State University), Ronak Singh (Penn State University)
Recommendation SystemOptimizationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper constructs a matching market benchmark to systematically evaluate the reasoning and algorithm execution capabilities of seven large language models across four types of tasks: stable matching generation, instability detection, instability repair, and preference querying.
Matchings Under Biased and Correlated Evaluations
Amit Kumar (Indian Institute of Technology Delhi), Nisheeth K. Vishnoi (Yale University)
🎯 What it does: This paper constructs a stable matching model with two institutions, studying how the representation ratio changes with the bias parameter β and the correlation parameter γ in the presence of group-level bias and partial correlation between institutions; it derives closed-form expressions for the matching threshold, representation ratio, and its normalized indicators in the large market limit.
MaterialRefGS: Reflective Gaussian Splatting with Multi-view Consistent Material Inference
Wenyuan Zhang (Tsinghua University), Zhizhong Han
RestorationGenerationData SynthesisNeural Radiance FieldGaussian SplattingImage
🎯 What it does: Learning multi-view illumination decomposition, utilizing 2D Gaussian scattering for reflection modeling and generating realistic views.
Matryoshka Pilot: Learning to Drive Black-Box LLMs with LLMs
ChangHao Li, Bo Dai (Georgia Institute of Technology)
Autonomous DrivingOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Designed and implemented a white-box LLM controller M-Pilot, which uses multi-round intermediate guidance to drive a black-box LLM to complete complex tasks;
Max Entropy Moment Kalman Filter for Polynomial Systems with Arbitrary Noise
Sangli Teng (University of Michigan), Luca Carlone (Massachusetts Institute of Technology)
OptimizationTime SeriesStochastic Differential Equation
🎯 What it does: A Kalman filter based on maximum entropy moment inference is proposed for state estimation under polynomial systems with arbitrary noise.
Maximizing the Value of Predictions in Control: Accuracy Is Not Enough
Yiheng Lin (California Institute of Technology), Adam Wierman (California Institute of Technology)
OptimizationReinforcement LearningSequential
🎯 What it does: A prediction power metric is introduced in the online optimal control problem with discrete time and random disturbances to quantify the improvement in control costs that can be achieved by utilizing predictive information.
MaxSup: Overcoming Representation Collapse in Label Smoothing
Yuxuan Zhou (University of Mannheim), Margret Keuper (CISPA Helmholtz Center for Information Security)
ClassificationRepresentation LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: Proposes the MaxSup method to address the issues of overconfidence in incorrect predictions and feature clustering compression in Label Smoothing, by penalizing the highest Logit instead.
MDNS: Masked Diffusion Neural Sampler via Stochastic Optimal Control
Yuchen Zhu (Georgia Institute of Technology), Molei Tao (FAIR at Meta)
OptimizationDiffusion modelGraphPhysics Related
🎯 What it does: A discrete neural sampler MDNS based on mask diffusion and stochastic optimal control has been developed for efficient sampling of high-dimensional multimodal target distributions.
MDReID: Modality-Decoupled Learning for Any-to-Any Multi-Modal Object Re-Identification
Yingying Feng (Northeastern University), Jiayi Ji (National University of Singapore)
RecognitionRetrievalTransformerImageMultimodality
🎯 What it does: A framework for image-level object re-identification across arbitrary modalities, MDReID, is proposed, which can handle both modality-matching and modality-mismatching retrieval tasks simultaneously.
Mean Flows for One-step Generative Modeling
Zhengyang Geng (Carnegie Mellon University), Kaiming He (Massachusetts Institute of Technology)
GenerationData SynthesisTransformerDiffusion modelFlow-based ModelImage
🎯 What it does: Proposes the MeanFlow framework, which utilizes the average velocity field to implement a one-stage generative model.
Mean-Field Sampling for Cooperative Multi-Agent Reinforcement Learning
Emile Timothy Anand (Georgia Institute of Technology), Guannan Qu (Carnegie Mellon University)
Reinforcement Learning
🎯 What it does: This paper proposes the SUBSAMPLE-MFQ algorithm, which utilizes the subsample mean field method to achieve scalable policy learning in cooperative multi-agent reinforcement learning. It can solve in polynomial time at a subsample size k and converges to the optimal policy at a rate of ˜O(1/√k) as k approaches n.
Measure gradients, not activations! Enhancing neuronal activity in deep reinforcement learning
Jiashun Liu (Hong Kong University of Science and Technology), Ling Pan (Hong Kong University of Science and Technology)
Reinforcement LearningDiffusion modelSequential
🎯 What it does: This paper studies the problem of neuron activity decay in deep reinforcement learning, proposing a gradient magnitude-based neuron activity metric called GraMa, and implementing a neuron reset method called ReGraMa based on this metric;
Measure-Theoretic Anti-Causal Representation Learning
Arman Behnam (Illinois Institute of Technology), Binghui Wang (Illinois Institute of Technology)
Representation LearningBiomedical Data
🎯 What it does: This paper proposes a measure-theory-based counterfactual representation learning framework called ACIA, which learns how observations are generated from labels at a low level and removes environmental noise at a high level.
Measuring AI Ability to Complete Long Software Tasks
Thomas Kwa (Model Evaluation and Threat Research), Lawrence Chan (Model Evaluation and Threat Research)
AI Code AssistantTabularBenchmark
🎯 What it does: This paper proposes and quantifies the '50% task completion time critical point' metric for AI completing software tasks, measuring the duration at which AI can achieve a 50% success rate, and tracks the trends of cutting-edge models from 2019 to 2025.
Measuring and Controlling Solution Degeneracy across Task-Trained Recurrent Neural Networks
Ann Huang (Harvard University), Kanaka Rajan (Harvard University)
Recurrent Neural NetworkSequential
🎯 What it does: A unified framework was constructed to quantify the overlap of the solution space of task training RNNs at three levels: behavior, neural dynamics, and weight space. Subsequently, the complexity of tasks, learning methods, network size, and structural regularization were systematically varied across 3,400 RNNs (for 4 neuroscience-related tasks) to quantify their impact on the diversity of the solution space.
Measuring and Guiding Monosemanticity
Ruben Härle (TU Darmstadt), Kristian Kersting (TU Darmstadt)
Safty and PrivacyRepresentation LearningAuto EncoderText
🎯 What it does: This paper proposes the Feature Monosemanticity Score (FMS) to measure the unambiguity of latent representations, and based on this, introduces Guided Sparse Autoencoders (G‑SAE), which enforce feature unambiguity and local decoupling in the latent space during training through conditional loss. The detection and guidance capabilities of this method are then evaluated in tasks such as toxicity detection, Shakespearean writing style identification, and privacy attribute recognition.
Measuring the Faithfulness of Thinking Drafts in Large Reasoning Models
Zidi Xiong (Harvard University), Himabindu Lakkaraju (Harvard University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This study investigates the authenticity of thinking drafts in large reasoning models (LRM) and proposes a two-dimensional measurement framework: Intra-Draft Faithfulness and Draft-to-Answer Faithfulness.
MeCeFO: Enhancing LLM Training Robustness via Fault-Tolerant Optimization
Rizhen Hu (Peking University), Kun Yuan (Peking University)
OptimizationTransformerLarge Language ModelText
🎯 What it does: A fault-tolerant optimization algorithm for large language model training, MeCeFO, is proposed, which enables rapid and low-overhead recovery during node failures while maintaining training stability.
Mechanism Design for LLM Fine-tuning with Multiple Reward Models
Haoran Sun (Peking University), Xiaotie Deng (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: The study investigates how to prevent agents from strategically misreporting their preferences during the fine-tuning process of large language models (LLMs) through mechanism design, thereby ensuring the achievement of the social welfare maximization goal.
Mechanism Design via the Interim Relaxation
Kshipra Bhawalkar (Google Research), Alexandros Psomas (Purdue University)
Optimization
🎯 What it does: This paper proposes a general framework for designing revenue-maximizing mechanisms for multi-party participants with additive preferences under downward closed constraints; the framework utilizes interim relaxation and a two-level online conflict resolution solution (two-level OCRS) to round the expected feasible interim rules, resulting in a sequence mechanism that is Bayesian incentive compatible (BIC) and has non-negative individual expected utility (BIR).
Mechanistic Interpretability of RNNs emulating Hidden Markov Models
Elia Torre (Institute of Neuroinformatics, University of Zurich and ETH Zurich), Valerio Mante (Institute of Neuroinformatics, University of Zurich and ETH Zurich)
Explainability and InterpretabilityRecurrent Neural NetworkSequential
🎯 What it does: Train standard RNNs to fit three types of HMMs (linear chain, fully connected, cyclic), and then reveal their internal implementations through reverse engineering: noise-maintained orbital dynamics, slow noise integration groups, and transfer mechanisms of fast kick-neuron interactions;
Median Selection with Noisy and Structural Information
Chenglin Fan (Seoul National University), Mingyu Kang (Seoul National University)
🎯 What it does: The study efficiently computes the exact median using side information through weak/strong comparisons or DAG structures in the presence of noise weak comparisons or known partial order structures.
MEGADance: Mixture-of-Experts Architecture for Genre-Aware 3D Dance Generation
kaixing yang, Hongyan Liu (Tsinghua University)
GenerationTransformerMixture of ExpertsVideoMultimodality
🎯 What it does: This paper proposes a music-driven 3D dance generation framework based on Mixture-of-Experts called MEGADance, which is divided into two stages: high-fidelity dance quantization (using FSQ and motion constraints) and music-based dance generation (utilizing MoE and a Mamba-Transformer hybrid backbone).
MEgoHand: Multimodal Egocentric Hand-Object Interaction Motion Generation
Bohan Zhou (Peking University), Zongqing Lu (Peking University)
GenerationPose EstimationDepth EstimationRobotic IntelligenceTransformerVision Language ModelImageVideoMultimodality
🎯 What it does: This paper presents MEgoHand, a multimodal top-down hand-object interaction motion generation framework that utilizes visual-language models and monocular depth to achieve spatial reasoning without object priors, and generates continuous, physically feasible hand trajectories through a closed-loop DiT flow matching generator and Temporal Orthogonal Filtering.
MEIcoder: Decoding Visual Stimuli from Neural Activity by Leveraging Most Exciting Inputs
Jan Sobotka (École Polytechnique Fédérale de Lausanne), Ján Antolík (Charles University)
RecognitionData SynthesisConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes MEIcoder, a decoding method based on the most excitatory input (MEI) of neurons, to reconstruct visual stimuli from V1 neuronal activity.
Mellow: a small audio language model for reasoning
Soham Deshmukh (Carnegie Mellon University), Bhiksha Raj (Carnegie Mellon University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningMultimodalityAudio
🎯 What it does: This paper designs and trains a small audio language model called Mellow, focusing on audio and text reasoning tasks, and constructs the ReasonAQA dataset to enhance the model's reasoning capabilities.
MemEIC: A Step Toward Continual and Compositional Knowledge Editing
Jin Seong (Electronics and Telecommunications Research Institute), Namhoon Lee (POSTECH)
TransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the MemEIC framework and the CCKEB benchmark for achieving continuous and combinatorial knowledge editing in large audiovisual language models.
Memo: Training Memory-Efficient Embodied Agents with Reinforcement Learning
Gunshi Gupta (University of Oxford), Rahaf Aljundi (Toyota Motor Europe)
Robotic IntelligenceTransformerReinforcement LearningSequential
🎯 What it does: Proposes the Memo framework, training a Transformer to compress and retrieve historical information in RL tasks by periodically generating summary tokens, thereby achieving more efficient long-term memory.
MEMOIR: Lifelong Model Editing with Minimal Overwrite and Informed Retention for LLMs
Ke Wang (École Polytechnique Fédérale de Lausanne), Pascal Frossard (École Polytechnique Fédérale de Lausanne)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes MEMOIR, a lifelong knowledge editing framework for large language models, which utilizes residual memory layers and sparse activation to achieve efficient and sustainable knowledge injection.
Memorization in Graph Neural Networks
Adarsh Jamadandi (CNRS IRISA), Franziska Boenisch (CISPA Helmholtz Center for Information Security)
ClassificationSafty and PrivacyGraph Neural NetworkGraph
🎯 What it does: The NCMemo framework is proposed, which quantifies the label memory of GNN in semi-supervised node classification using the leave-one-out method, and analyzes its relationship with graph homogeneity, training dynamics, and label inconsistency; further evaluates the impact of memory on privacy leakage and demonstrates that graph reconnection can reduce memory rates and MIA risks.
Memory by accident: a theory of learning as a byproduct of network stabilization
Basile Confavreux (Gatsby Computational Neuroscience Unit University College London), Andrew M Saxe
Meta LearningSpiking Neural NetworkSequential
🎯 What it does: Through meta-learning, a large class of plasticity rules that can maintain stability in large-scale recurrent synaptic networks is discovered, and it is observed that these rules can almost always produce memory functions such as familiarity/novelty detection without being specifically optimized for memory.
Memory Decoder: A Pretrained, Plug-and-Play Memory for Large Language Models
Jiaqi Cao (Shanghai Jiao Tong University), Zhouhan Lin (Shanghai Jiao Tong University)
RetrievalDomain AdaptationTransformerLarge Language ModelTextBiomedical DataFinance RelatedRetrieval-Augmented Generation
🎯 What it does: A pre-trained, pluggable Memory Decoder is proposed, which achieves non-parametric domain adaptation for large language models by learning to mimic the distribution of non-parametric retrievers within a small Transformer decoder.
Memory Injection Attacks on LLM Agents via Query-Only Interaction
Shen Dong (Michigan State University), Zhen Xiang (University of Georgia)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: This paper proposes an attack method (MINJA) that can inject malicious records into the LLM agent's memory bank solely through query interactions, achieving memory injection and malicious reasoning through bridging steps, indication prompts, and a progressive shortening strategy.
Memory Mosaics at scale
Jianyu Zhang (New York University), Leon Bottou
TransformerSupervised Fine-TuningText
🎯 What it does: This work extends the Memory Mosaics architecture to the 8B scale (Memory Mosaics v2) and conducts pre-training and fine-tuning on real-world data of 1 trillion tokens, validating its capabilities in new task learning, knowledge storage, and contextual learning.
Memory-Augmented Potential Field Theory: A Framework for Adaptive Control in Non-Convex Domains
Dongzhe Zheng (Princeton University), Wenjie Mei (Nanjing University)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper proposes the Memory-Augmented Potential Field Theory (MAPFT) and implements a Memory-Augmented MPPI (MA-MPPI) controller based on this theory to achieve online experiential learning and adaptive optimization in highly non-convex state spaces.
Memory-Efficient Training with In-Place FFT Implementation
Xinyu Ding (Sun Yat-sen University), Zhongfeng Wang (Sun Yat-sen University)
Computational EfficiencyRecurrent Neural NetworkSupervised Fine-TuningText
🎯 What it does: A real-domain fully in-place fast Fourier transform (rdFFT) is proposed, addressing the memory mismatch issues in traditional FFT and rFFT during training, and integrating it into circulant matrix networks to achieve forward/backward propagation without additional caching.
Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression
Kunjun Li (National University of Singapore), Jenq-Neng Hwang (University of Washington)
GenerationCompressionComputational EfficiencyTransformerImage
🎯 What it does: Designed and implemented a scale-aware KV cache compression framework called ScaleKV for visual autoregressive models, significantly reducing KV cache usage while maintaining pixel-level fidelity.
Memory-Enhanced Neural Solvers for Routing Problems
Felix Chalumeau (InstaDeep), Nathan Grinsztajn (InstaDeep)
OptimizationTransformerReinforcement LearningTabular
🎯 What it does: MEMENTO is proposed, a memory mechanism that utilizes online storage to dynamically adjust the action distribution of the neural routing solver during the inference phase, thereby improving search efficiency.
Memory-Integrated Reconfigurable Adapters: A Unified Framework for Settings with Multiple Tasks
Susmit Agrawal (IIT Hyderabad), Vineeth N. Balasubramanian (IIT Hyderabad)
Domain AdaptationTransformerSupervised Fine-TuningImage
🎯 What it does: The MIRA framework is proposed, utilizing Hopfield-style associative memory and low-rank adapters to achieve seamless switching of a single model in domain generalization, domain incremental learning, and class incremental learning.
MemSim: A Bayesian Simulator for Evaluating Memory of LLM-based Personal Assistants
Zeyu Zhang (Renmin University of China), Ji-Rong Wen (Renmin University of China)
TransformerLarge Language ModelText
🎯 What it does: A Bayesian simulator named MemSim has been developed to automatically generate reliable, scalable, and diverse user messages and question-answer datasets, thereby objectively assessing the memory capabilities of personal assistants based on large language models (LLMs).
Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging
Anke Tang (Wuhan University), Dacheng Tao (Nanyang Technological University)
ClassificationRecognitionTransformerContrastive LearningImage
🎯 What it does: An algorithm for training-free, sequential merging of models is proposed, achieving efficient fusion when models appear in sequence.
MERIT: Multilingual Semantic Retrieval with Interleaved Multi-Condition Query
Wei Chow (ByteDance Inc), Juncheng Li (Zhejiang University)
RetrievalTransformerLarge Language ModelContrastive LearningTextMultimodality
🎯 What it does: The first multilingual, interleaved multi-condition semantic retrieval dataset MERIT has been constructed, and a retrieval fine-tuning framework CORAL based on MLLM has been proposed.
MesaTask: Towards Task-Driven Tabletop Scene Generation via 3D Spatial Reasoning
Jinkun Hao (Shanghai Jiao Tong University), Jiangmiao Pang (Shanghai AI Laboratory)
GenerationData SynthesisOptimizationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: A task-driven desktop scene generation framework called MesaTask based on large language models is proposed and implemented, which can generate 3D desktop scenes with realistic physical layouts and complex object relationships based on high-level task instructions.
Mesh Interpolation Graph Network for Dynamic and Spatially Irregular Global Weather Forecasting
Zinan Zheng (Hong Kong University of Science and Technology), Jia Li (Hong Kong University of Science and Technology)
Graph Neural NetworkTabularTime Series
🎯 What it does: Designed and implemented the Mesh Interpolation Graph Network (MIGN) for global dynamic and spatially irregular weather station predictions.
Mesh-RFT: Enhancing Mesh Generation via Fine-grained Reinforcement Fine-Tuning
Jian Liu (Hong Kong University of Science and Technology), Chunchao Guo (Tencent Hunyuan)
GenerationOptimizationTransformerReinforcement LearningPoint CloudMesh
🎯 What it does: This paper proposes the Mesh-RFT framework, which fine-tunes a pre-trained 3D mesh generation model using fine-grained reinforcement learning (Masked Direct Preference Optimization, M-DPO) and implements local refinement with face-level quality masks, significantly improving mesh geometric integrity and topological consistency.
MeshCoder: LLM-Powered Structured Mesh Code Generation from Point Clouds
BingQuan Dai, Jiangmiao Pang (Shanghai Artificial Intelligence Laboratory)
GenerationAI Code AssistantTransformerLarge Language ModelPoint CloudMesh
🎯 What it does: Proposes the MeshCoder framework, which directly converts 3D point clouds into editable Blender Python scripts, enabling the structured reconstruction of complex objects.
MESS+: Dynamically Learned Inference-Time LLM Routing in Model Zoos with Service Level Guarantees
Herbert Woisetschläger (Technical University of Munich), Hans Arno Jacobsen
OptimizationTransformerLarge Language ModelText
🎯 What it does: Proposes MESS+, a dynamic learning inference LLM routing algorithm that ensures minimum SLA compliance while minimizing operational costs.
Meta CLIP 2: A Worldwide Scaling Recipe
Yung-Sung Chuang (Meta), Hu Xu (Meta)
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: Proposes Meta CLIP 2, a complete data planning and training scheme for training CLIP from scratch on a global scale;
Meta Guidance: Incorporating Inductive Biases into Deep Time Series Imputers
Jiacheng You (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
Anomaly DetectionData-Centric LearningTransformerTime Series
🎯 What it does: By designing learnable Non-Stationary Guidance (NSG) and Periodic Guidance (PG) matrices, and automatically assigning weights to each sequence using Meta Guidance, the performance of deep missing value imputation models is enhanced.
Meta-D2AG: Causal Graph Learning with Interventional Dynamic Data
Tian Gao (IBM Corporation), Miao Liu (Chinese University of Hong Kong)
Meta LearningGraph Neural NetworkReinforcement LearningTime SeriesSequential
🎯 What it does: A dynamic directed acyclic graph (DAG) learning framework based on meta-learning, called Meta-DAG, is proposed, which can quickly adapt to new distributions and graph structures in non-stationary time series.
Meta-Learning an In-Context Transformer Model of Human Higher Visual Cortex
Muquan Yu (Chinese University of Hong Kong), Andrew Luo
Meta LearningTransformerImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: We propose BraInCoRL, a Transformer-based meta-learning framework that can generate individualized voxel-level encoding models without fine-tuning, given only a small number of image-brain response examples for target individuals.
Meta-learning how to Share Credit among Macro-Actions
Ionel Hosu, Razvan Pascanu (Mila - Quebec Artificial Intelligence Institute)
Meta LearningReinforcement LearningVideo
🎯 What it does: This paper proposes a regularization method for credit allocation using macro action similarity to improve exploration and learning efficiency in reinforcement learning.
Meta-Learning Objectives for Preference Optimization
Carlo Alfano (University of Oxford), Yee Whye Teh (University of Oxford)
OptimizationMeta LearningLarge Language ModelReinforcement LearningSequential
🎯 What it does: The paper proposes and evaluates a novel preference optimization algorithm framework based on mirror descent in large language models and reinforcement learning environments.
MetaDefense: Defending Fine-tuning based Jailbreak Attack Before and During Generation
Weisen Jiang (Chinese University of Hong Kong), Sinno Jialin Pan (Chinese University of Hong Kong)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes MetaDefense, a two-stage defense framework that detects whether LLM responses are harmful before and after generation, preventing finetuning-based jailbreak attacks;
MetaFind: Scene-Aware 3D Asset Retrieval for Coherent Metaverse Scene Generation
Zhenyu Pan, Han Liu
GenerationRetrievalGraph Neural NetworkContrastive LearningMultimodalityPoint Cloud
🎯 What it does: MetaFind is proposed, a scene-aware multimodal 3D asset retrieval framework designed to generate coherent scenes in the metaverse;
MetaGS: A Meta-Learned Gaussian-Phong Model for Out-of-Distribution 3D Scene Relighting
Yumeng He (Shanghai Jiao Tong University), Yunbo Wang (Shanghai Jiao Tong University)
GenerationMeta LearningGaussian SplattingImage
🎯 What it does: The MetaGS model is proposed, which combines 3D Gaussian splatting with a differentiable Blinn-Phong reflection model for handling out-of-distribution (OOD) 3D scene relighting.
MetaKoopman: Bayesian Meta-Learning of Koopman Operators for Modeling Structured Dynamics under Distribution Shifts
Mahmoud Selim (TRATON), Karl Henrik Johansson
Autonomous DrivingMeta LearningReinforcement LearningTime Series
🎯 What it does: The MetaKoopman framework is proposed, which learns the Koopman operator as a closed-form update prior through Bayesian meta-learning, achieving rapid adaptation to structured dynamics under distribution drift.
MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems
Xuanming Zhang (University of Wisconsin-Madison), Sharon Li (University of Wisconsin-Madison)
TransformerLarge Language ModelAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: Proposes the MetaMind multi-agent framework, which achieves human social thinking modeling through a three-stage collaboration of theoretical mind, moral constraints, and response validation.
MetaSlot: Break Through the Fixed Number of Slots in Object-Centric Learning
Hongjia Liu (Aalto University), Joni Pajarinen (Aalto University)
Object DetectionSegmentationData-Centric LearningImageVideo
🎯 What it does: Proposes MetaSlot to address the fixed number of slots and random initialization issues in Slot Attention;
metaTextGrad: Automatically optimizing language model optimizers
Guowei Xu (Tsinghua University), James Zou (Stanford University)
OptimizationMeta LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the metaTextGrad framework, which enhances the task adaptability of LLM optimizers through meta-optimization.
Metis: A Foundation Speech Generation Model with Masked Generative Pre-training
Yuancheng Wang (Chinese University of Hong Kong), Zhizheng Wu (Chinese University of Hong Kong)
GenerationTransformerMultimodalityAudio
🎯 What it does: Metis is proposed, a unified speech generation framework based on large-scale unsupervised speech pre-training and fine-tuning, which performs excellently across various tasks such as zero-shot, voice conversion, target speaker extraction, speech enhancement, and lip reading generation.
Metric Automata Theory: A Unifying Theory of RNNs
Adam Dankowiakowski (University of Oxford), Alessandro Ronca (IRIS-AI)
Recurrent Neural Network
🎯 What it does: This paper proposes Metric Automata Theory (MAT), which extends traditional finite automata theory to continuous dynamical systems, thereby providing a unified framework for expressive power analysis of all RNNs (including xLSTM, SSM, Mamba, etc.) and presents a series of rigorous expressiveness theorems under finite precision and geometric constraints.
Metritocracy: Representative Metrics for Lite Benchmarks
Ariel D. Procaccia, Shirley Zhang
OptimizationTabularBenchmark
🎯 What it does: This paper studies how to select a small subset from a complete set of evaluation metrics that can represent the ranking information of the original metric set;
Metropolis Adjusted Microcanonical Hamiltonian Monte Carlo
Jakob Robnik (University of California), Uros Seljak
OptimizationHyperparameter SearchTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A new unbiased microcanonical Markov chain sampler (MAMS) is proposed and implemented, achieving high-dimensional unbiased sampling competitive with NUTS by incorporating Metropolis-Hastings steps into microcanonical dynamics.
Metropolis-Hastings Sampling for 3D Gaussian Reconstruction
Hyunjin Kim (University of California San Diego), Jaesik Park (Seoul National University)
RestorationGenerationGaussian SplattingPoint Cloud
🎯 What it does: An adaptive 3D Gaussian point cloud modeling framework based on Metropolis-Hastings sampling is proposed, replacing traditional threshold heuristic densification and pruning;
MF-LLM: Simulating Population Decision Dynamics via a Mean-Field Large Language Model Framework
Qirui Mi (Chinese Academy of Sciences), Jun Wang (University College London)
TransformerLarge Language ModelSupervised Fine-TuningTextTime SeriesFinance Related
🎯 What it does: By combining mean field theory with large language models (LLM), a scalable simulation of population-level decision dynamics has been achieved.
MGE-LDM: Joint Latent Diffusion for Simultaneous Music Generation and Source Extraction
Yunkee Chae (Seoul National University), Kyogu Lee (Seoul National University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderAudio
🎯 What it does: A unified latent diffusion framework (MGE-LDM) is proposed, capable of simultaneously performing overall music generation, partial generation (source filling), and arbitrary source extraction based on text.
MGUP: A Momentum-Gradient Alignment Update Policy for Stochastic Optimization
Da Chang (Shenzhen Institute of Advanced Technology), Ganzhao Yuan (Shenzhen University of Advanced Technology)
OptimizationTransformerSupervised Fine-TuningText
🎯 What it does: A selective update mechanism based on momentum-gradient consistency (MGUP) is proposed, which allows for using a larger learning rate for a portion of the parameters at each step while applying a smaller learning rate to the remaining parameters, thus achieving efficient and stable training.
MI-TRQR: Mutual Information-Based Temporal Redundancy Quantification and Reduction for Energy-Efficient Spiking Neural Networks
Dengfeng Xue (Xidian University), Zhetao Li (Jinan University)
Spiking Neural NetworkImageTime Series
🎯 What it does: A parameter-free, pluggable MI-TRQR module is proposed to quantify and eliminate redundant synapses across time steps in SNNs, thereby reducing energy consumption and improving accuracy.
MIBP-Cert: Certified Training against Data Perturbations with Mixed-Integer Bilinear Programs
Tobias Lorenz (CISPA Helmholtz Center for Information Security), Mario Fritz (CISPA Helmholtz Center for Information Security)
OptimizationTabular
🎯 What it does: A training-time robustness certification method based on Mixed Integer Bilinear Programming (MIBP), called MIBP-Cert, is proposed, which can accurately compute the upper and lower bounds of model parameters under constrained data perturbations and provide provable robustness.
MiCADangelo: Fine-Grained Reconstruction of Constrained CAD Models from 3D Scans
Ahmet Serdar Karadeniz (University of Luxembourg), Djamila Aouada (University of Luxembourg)
GenerationOptimizationConvolutional Neural NetworkTransformerPoint CloudMesh
🎯 What it does: MiCADangelo converts 3D scans into editable parametric CAD models;
MiCo: Multi-image Contrast for Reinforcement Visual Reasoning
Xi Chen (Hong Kong University), Hengshuang Zhao (Hong Kong University)
TransformerReinforcement LearningVision Language ModelContrastive LearningImageVideoBenchmark
🎯 What it does: Design a self-supervised multi-image comparison and chain reasoning framework (MiCo) that allows visual language models to learn fine-grained comparison and reasoning across multiple images without human annotations.
MIDAS: Misalignment-based Data Augmentation Strategy for Imbalanced Multimodal Learning
Seong-Hyeon Hwang (Korea Advanced Institute of Science and Technology), Steven Euijong Whang (Korea Advanced Institute of Science and Technology)
ClassificationData-Centric LearningVideoMultimodalityAudio
🎯 What it does: MIDAS is proposed, a data augmentation strategy based on misaligned samples, which utilizes semantic conflicts in multimodal information to alleviate the modality imbalance problem.
MigGPT: Harnessing Large Language Models for Automated Migration of Out-of-Tree Linux Kernel Patches Across Versions
Pucheng Dang (Institute of Computing Technology, Chinese Academy of Sciences), Xing Hu (Institute of Computing Technology, Chinese Academy of Sciences)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the MIGGPT framework for the automatic migration of offline patches (out-of-tree patches) in the Linux kernel, significantly reducing manual maintenance costs.
MIHC: Multi-View Interpretable Hypergraph Neural Networks with Information Bottleneck for Chip Congestion Prediction
Zeyue Zhang (Renmin University of China), Xiongye Xiao (University of Tennessee)
OptimizationExplainability and InterpretabilityGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A multi-view interpretable hypergraph neural network MIHC is proposed for chip congestion prediction.
Mind the Gap: Bridging Thought Leap for Improved Chain-of-Thought Tuning
Haolei Xu (Zhejiang University), Yueting Zhuang (Zhejiang University)
OptimizationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: This study addresses the issue of missing 'Thought Leap' in Chain of Thought (CoT) reasoning, proposing the CoT Thought Leap Bridge task and training the CoT-Bridge model, which can automatically detect and complete leap steps, thereby enhancing the integrity of CoT and the model's reasoning performance.
Mind the Gap: Removing the Discretization Gap in Differentiable Logic Gate Networks
Shakir Yousefi, Roger Wattenhofer
ClassificationOptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes Gumbel Logic Gate Networks (GLGNs), which address the slow training and high discretization error of Differentiable Logic Gate Networks (DLGNs) by injecting Gumbel noise into gate selection and using a straight-through estimator;
Mind the GAP! The Challenges of Scale in Pixel-based Deep Reinforcement Learning
Ghada Sokar (Google DeepMind), Pablo Samuel Castro (Google DeepMind)
Convolutional Neural NetworkReinforcement LearningMixture of ExpertsImage
🎯 What it does: The study investigates the reasons for performance degradation in pixel input reinforcement learning networks during scaling and proposes alleviating the bottleneck by using global average pooling (GAP) on the encoder output.
Mind the Quote: Enabling Quotation-Aware Dialogue in LLMs via Plug-and-Play Modules
Yueqi Zhang (Beijing Institute of Technology), Kan Li (Xiaohongshu Inc)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposes a span-conditioned generation task that allows LLMs to accurately locate and use referenced text in conversations, and builds an automated data generation and validation pipeline, releasing a benchmark that includes five sub-scenes.
Mind-the-Glitch: Visual Correspondence for Detecting Inconsistencies in Subject-Driven Generation
Abdelrahman Eldesokey (King Abdullah University of Science and Technology), Peter Wonka (King Abdullah University of Science and Technology)
SegmentationGenerationAnomaly DetectionDiffusion modelContrastive LearningImage
🎯 What it does: A method is proposed to decompose the features of a pre-trained diffusion model into semantic and visual sub-features, and based on this, a quantifiable and locatable visual inconsistency assessment metric, VSM, is designed.
MIND: Material Interface Generation from UDFs for Non-Manifold Surface Reconstruction
Xuhui Chen (Institute of Software, Chinese Academy of Sciences), Ying He (Nanyang Technological University)
SegmentationGenerationAutonomous DrivingPoint CloudMesh
🎯 What it does: Proposes the MIND algorithm, which directly generates material interfaces (MI) from unsigned distance fields (UDF), achieving non-manifold mesh reconstruction.
MindForge: Empowering Embodied Agents with Theory of Mind for Lifelong Cultural Learning
Mircea Tudor Lică (Delft University of Technology), Chirag Raman (Delft University of Technology)
Knowledge DistillationRobotic IntelligenceTransformerLarge Language ModelAgentic AIText
🎯 What it does: The MINDFORGE framework is proposed, enabling embodied agents based on large language models to achieve cultural lifelong learning through structured theory of mind, natural language communication, and multi-component memory.
MindJourney: Test-Time Scaling with World Models for Spatial Reasoning
Yuncong Yang (University of Massachusetts Amherst), Chuang Gan (University of Massachusetts Amherst)
Reinforcement LearningVision Language ModelDiffusion modelWorld ModelImageVideoBenchmark
🎯 What it does: During testing, a controllable video diffusion model is used as a world model, allowing the visual language model (VLM) to interactively explore implicit 3D scenes from a single image through spatial beam search, thereby enhancing spatial reasoning capabilities.