arXivSub Start free trial

ICML 2025 Papers — Page 19

International Conference on Machine Learning · 3257 papers

Measuring Diversity: Axioms and Challenges

Mikhail Mironov (Yandex Research), Liudmila Prokhorenkova (Yandex Research)

Review/Survey Paper

🎯 What it does: This paper systematically evaluates existing diversity measurement methods and proposes three feasibility axioms: monotonicity, uniqueness, and continuity; it then constructs two metrics that satisfy all axioms but have high computational complexity, proving that the three axioms are not in conflict.

Measuring In-Context Computation Complexity via Hidden State Prediction

Vincent Herrmann (IDSIA USI SUPSI), Jürgen Schmidhuber (KAUST)

Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAuto EncoderTextSequential

🎯 What it does: This paper proposes the PHi layer and PHi loss, which measure the complexity of 'interesting' computations performed by large-scale language models in context by predicting the model's own hidden states.

Measuring Representational Shifts in Continual Learning: A Linear Transformation Perspective

Joonkyu Kim (Yonsei University), Jy-yong Sohn (Yonsei University)

Representation LearningConvolutional Neural NetworkTransformerImage

🎯 What it does: Theoretical research on the changes in hidden layer representations during the continual learning process is conducted, proposing an analyzable representation difference metric and deriving its upper bound.

Measuring Variable Importance in Heterogeneous Treatment Effects with Confidence

Joseph Paillard (Roche Innovation Center Basel), Denis-Alexander Engemann

TabularBiomedical Data

🎯 What it does: This paper proposes PermuCATE, an algorithm based on Conditional Permutation Importance (CPI) for interpretable and statistically confident global variable importance assessment of Conditional Average Treatment Effect (CATE) models in causal machine learning.

Mechanisms of Projective Composition of Diffusion Models

Arwen Bradley (Apple), Joshua M. Susskind (Apple)

GenerationData SynthesisDiffusion modelScore-based ModelImage

🎯 What it does: This paper studies the combination mechanism of diffusion models, proposing a formal definition of 'Projective Composition' and theoretically analyzing when linear score combinations can achieve correct combinations, with a particular focus on length generalization and OOD generation.

Mechanistic PDE Networks for Discovery of Governing Equations

Adeel Pervez (Institute of Science and Technology), Francesco Locatello (Informatics Institute)

OptimizationTime SeriesPhysics RelatedOrdinary Differential Equation

🎯 What it does: This paper proposes Mechanistic PDE Networks (MechNN-PDE), a framework that embeds differentiable linear PDE solvers into neural networks to automatically discover governing PDEs from spatiotemporal data; and implements a GPU-parallel sparse multigrid solver NeuRLP-PDE.

Mechanistic Unlearning: Robust Knowledge Unlearning and Editing via Mechanistic Localization

Phillip Huang Guo (University of Maryland), Gintare Karolina Dziugaite (Google DeepMind)

Federated LearningExplainability and InterpretabilityTransformerLarge Language ModelText

🎯 What it does: Exploring and proving that using mechanism localization (FLU) to edit and unlearn factual information in large language models can significantly enhance robustness and reduce side effects.

MedRAX: Medical Reasoning Agent for Chest X-ray

Adibvafa Fallahpour (University of Toronto), BO WANG

ClassificationSegmentationGenerationTransformerLarge Language ModelAgentic AIVision Language ModelImageMultimodalityBiomedical DataBenchmark

🎯 What it does: MedRAX is proposed, a medical reasoning agent based on LLM that can dynamically invoke various chest X-ray-specific tools to complete complex diagnostic and report generation tasks.

MedXpertQA: Benchmarking Expert-Level Medical Reasoning and Understanding

Yuxin Zuo (Tsinghua University), Bowen Zhou (Tsinghua University)

TransformerLarge Language ModelTextMultimodalityBenchmark

🎯 What it does: Proposes MedXpertQA, an expert-level medical reasoning and understanding benchmark covering 17 departments, 11 systems, and 4,460 questions, including text and multimodal subsets;

MELON: Provable Defense Against Indirect Prompt Injection Attacks in AI Agents

Kaijie Zhu (University of California), William Yang Wang

Adversarial AttackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark

🎯 What it does: A new Indirect Prompt Injection (IIP) defense method called MELON is proposed, which detects attacks by performing a parallel 'mask' re-execution during execution and comparing the tool calls of the two execution paths.

MemFreezing: A Novel Adversarial Attack on Temporal Graph Neural Networks under Limited Future Knowledge

Yue Dai (University of Pittsburgh), Jun Yang (University of Pittsburgh)

Adversarial AttackGraph Neural NetworkGraphTime Series

🎯 What it does: This paper proposes MemFreezing, a framework for adversarial attacks on Temporal Graph Neural Networks (TGNN) that can only access information from the graph prior to the attack;

Memorization Sinks: Isolating Memorization during LLM Training

Gaurav Rohit Ghosal, Aditi Raghunathan (Carnegie Mellon University)

TransformerLarge Language ModelText

🎯 What it does: Proposes the MemSinks training framework, which divides generalized neurons and memory sink neurons in the Transformer MLP layer, and uses sequence ID-dependent dropout to isolate the memory of repeated sequences, thereby achieving controllable memory and de-memory of repeatable sequences in large language models.

Memory Layers at Scale

Vincent-Pierre Berges (Meta Platforms), Gargi Ghosh (Meta Platforms)

RetrievalOptimizationTransformerLarge Language ModelTextRetrieval-Augmented Generation

🎯 What it does: A trainable sparse memory layer is proposed and implemented on a large scale, replacing the feedforward network of the Transformer with a key-value lookup mechanism, demonstrating significant improvements in factual knowledge retrieval.

MENTOR: Mixture-of-Experts Network with Task-Oriented Perturbation for Visual Reinforcement Learning

Suning Huang (Stanford University), Huazhe Xu (Tsinghua University)

Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningMixture of ExpertsImage

🎯 What it does: This paper proposes a vision-based model-free reinforcement learning framework called MENTOR, which improves sample efficiency and achieves more robust policies by introducing a mixture of experts network (MoE) and a task-oriented perturbation mechanism.

Merge-Friendly Post-Training Quantization for Multi-Target Domain Adaptation

Juncheol Shin (Pohang University of Science and Technology), Eunhyeok Park (Pohang University of Science and Technology)

SegmentationDomain AdaptationAutonomous DrivingImage

🎯 What it does: This paper proposes a mergeable post-training quantization method for multi-target domain adaptation, HDRQ, to address the issue of performance degradation in model merging caused by quantization.

MERGE$^3$: Efficient Evolutionary Merging on Consumer-grade GPUs

Tommaso Mencattini (Ecole Polytechnique Federale de Lausanne), Emanuele Rodolà (Sapienza University of Rome)

OptimizationComputational EfficiencyLarge Language ModelText

🎯 What it does: This study proposes MERGE 3, an efficient framework for evolutionary merging of large language models implemented on a single consumer-grade GPU.

MERIT: Maximum-normalized Element-wise Ratio for Language Model Large-batch Training

Yang Luo (National University of Singapore), Yang You (National University of Singapore)

OptimizationTransformerLarge Language ModelText

🎯 What it does: The MERIT optimizer is proposed, which enhances the stability and effectiveness of training language models at large batch sizes by utilizing maximum norm and element-wise trust ratios.

Meta Optimality for Demographic Parity Constrained Regression via Post-Processing

Kazuto Fukuchi (University of Tsukuba)

OptimizationTabular

🎯 What it does: This paper addresses the regression problem under the constraint of population equality and proposes a post-processing method to achieve fair minimax optimal regression.

Meta-Black-Box-Optimization through Offline Q-function Learning

Zeyuan Ma (South China University of Technology), Yue-Jiao Gong (Singapore Management University)

OptimizationMeta LearningReinforcement LearningTabular

🎯 What it does: A MetaBBO framework Q-Mamba based on offline Q-learning is proposed, which learns dynamic algorithm configuration strategies on offline datasets.

Meta-Reinforcement Learning with Adaptation from Human Feedback via Preference-Order-Preserving Task Embedding

Siyuan Xu (Pennsylvania State University), Minghui Zhu (Pennsylvania State University)

Meta LearningReinforcement Learning from Human FeedbackReinforcement LearningAuto EncoderSequential

🎯 What it does: This paper proposes a meta-reinforcement learning framework POEM that is oriented towards human feedback, capable of quickly adapting to new tasks with only preference sequences instead of reward signals.

MetaAgent: Automatically Constructing Multi-Agent Systems Based on Finite State Machines

Yaolun Zhang (University of Wisconsin Madison), Chaowei Xiao (University of Wisconsin Madison)

Large Language ModelPrompt EngineeringTextBenchmark

🎯 What it does: Automatically construct a multi-agent system framework called MetaAgent, which uses finite state machines to automatically generate and optimize multi-agent processes based on task descriptions.

Metadata Conditioning Accelerates Language Model Pre-training

Tianyu Gao (Princeton University), Danqi Chen (Princeton University)

TransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: During the pre-training process, the authors provide additional contextual signals to the model by appending metadata (such as URLs) at the beginning of each document, and in the later stages of training, they remove the metadata in a 'cooling' phase, forming a new pre-training paradigm (MeCo).

MetaOptimize: A Framework for Optimizing Step Sizes and Other Meta-parameters

Arsalan Sharifnassab (Openmind Research Institute), Richard S. Sutton (University of Alberta)

OptimizationHyperparameter SearchConvolutional Neural NetworkTransformerImageText

🎯 What it does: A general framework called MetaOptimize is proposed, which can optimize hyperparameters such as learning rate in real-time during the training process to minimize the regret of discounted future loss.

Metastable Dynamics of Chain-of-Thought Reasoning: Provable Benefits of Search, RL and Distillation

Juno Kim (University of Tokyo and RIKEN AIP), Taiji Suzuki (University of Tokyo and RIKEN AIP)

OptimizationComputational EfficiencyKnowledge DistillationReinforcement LearningChain-of-Thought

🎯 What it does: This paper views Chain-of-Thought (CoT) as a regressive Markov chain with a dense-sparse structure, analyzing the theoretical improvements in reasoning efficiency brought by search, reinforcement learning, and distillation;

MetricEmbedding: Accelerate Metric Nearness by Tropical Inner Product

Muyang Cao (Zhejiang University), Wei Wang (Hong Kong University of Science and Technology)

OptimizationGraph Neural NetworkGraph

🎯 What it does: This paper proposes a novel network structure called MetricEmbedding based on tropical inner product (max-plus operation) to address the Metric Nearness problem, which aims to recover any non-metric matrix to the closest metric matrix while ensuring the triangle inequality.

MF-LAL: Drug Compound Generation Using Multi-Fidelity Latent Space Active Learning

Peter Eckmann (University of California San Diego), Rose Yu (University of California San Diego)

GenerationDrug DiscoveryTabularBiomedical Data

🎯 What it does: This paper proposes a multi-precision latent space active learning framework MF-LAL that integrates multi-level precision simulators with generative models to generate drug molecules with high true binding free energy at different precision levels.

MGD$^3$ : Mode-Guided Dataset Distillation using Diffusion Models

Jeffrey A Chan Santiago, Mubarak Shah (University of Central Florida)

GenerationData SynthesisKnowledge DistillationDiffusion modelImage

🎯 What it does: By introducing a three-stage process of pattern discovery, pattern guidance, and stop guidance on a pre-trained diffusion model, we achieve the extraction of diverse and representative synthetic samples from the data distribution without the need for fine-tuning, completing dataset distillation.

MIB: A Mechanistic Interpretability Benchmark

Aaron Mueller (Boston University), Yonatan Belinkov

Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark

🎯 What it does: This paper proposes and implements the MIB (Mechanistic Interpretability Benchmark), aimed at unifying the evaluation of circuit localization and causal variable localization methods in neural language models.

MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance

Yuang Zhang (Tencent), FangYuan Zou

GenerationData SynthesisPose EstimationDiffusion modelVideo

🎯 What it does: Generate long-duration, high-quality human action videos based on a single reference image and a sequence of poses.

Mind the Gap: A Practical Attack on GGUF Quantization

Kazuki Egashira (ETH Zurich), Martin Vechev (ETH Zurich)

OptimizationAdversarial AttackText

🎯 What it does: An attack framework for the GGUF quantization method (k-quants) is proposed, which can inject malicious behavior after model quantization while the full-precision version remains normal.

Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Attention Layers

Thiziri Nait Saada (Mathematical Institute, University of Oxford), Jared Tanner (Mathematical Institute, University of Oxford)

Transformer

🎯 What it does: Analyzed the spectral properties of the softmax self-attention layer in the Transformer during random initialization, revealing rank collapse in the width direction (tokens converge to a single representation within a single layer) and issues leading to gradient explosion.

Mind Your Step (by Step): Chain-of-Thought can Reduce Performance on Tasks where Thinking Makes Humans Worse

Ryan Liu (Princeton University), Thomas L. Griffiths (Princeton University)

ClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringTextMultimodalityChain-of-Thought

🎯 What it does: This study investigates the limitations of Chain of Thought (CoT) prompts in large language and multimodal models, drawing on human overthinking tasks to identify scenarios where CoT performance declines.

MindAligner: Explicit Brain Functional Alignment for Cross-Subject Visual Decoding from Limited fMRI Data

Yuqin Dai (Nanjing University of Science and Technology), Jiamin Wu (Chinese University of Hong Kong)

RetrievalExplainability and InterpretabilityContrastive LearningBiomedical DataMagnetic Resonance Imaging

🎯 What it does: We propose MindAligner, an explicit brain function alignment framework that maps the fMRI signals of new subjects to the known subject space through a brain transfer matrix, enabling cross-subject visual decoding.

MindCustomer: Multi-Context Image Generation Blended with Brain Signal

Muzhou Yu (Xi'an Jiaotong University), Kaisheng Ma (Tsinghua University)

GenerationData SynthesisRecommendation SystemDiffusion modelImageMultimodalityMagnetic Resonance Imaging

🎯 What it does: The MindCustomer method is proposed, which utilizes diffusion models to integrate brain signals, images, and optional text context to generate personalized images without masks and from a single image.

MindLLM: A Subject-Agnostic and Versatile Model for fMRI-to-text Decoding

Weikang Qiu (Yale University), Rex Ying (Yale University)

TransformerLarge Language ModelTextBiomedical DataMagnetic Resonance Imaging

🎯 What it does: The MindLLM model is proposed, achieving theme-independent and general text decoding of fMRI signals.

Minerva: A Programmable Memory Test Benchmark for Language Models

Menglin Xia (Microsoft), Reza Shokri (National University of Singapore)

TransformerLarge Language ModelTextBenchmark

🎯 What it does: Minerva proposes a programmable benchmark framework that can automatically generate and evaluate the memory usage capabilities of large language models at different levels, from basic search, recall, and editing to state tracking and combinatorial tasks across multiple dimensions.

Minimalist Concept Erasure in Generative Models

Yang Zhang (National University of Singapore), Kenji Kawaguchi (National University of Singapore)

GenerationData SynthesisAdversarial AttackFlow-based ModelRectified FlowImageText

🎯 What it does: A minimal concept elimination method is proposed for text-to-image generation models, aiming to eliminate specified concepts by fine-tuning the distribution distance of the final generated results without significantly affecting the overall performance of the model.

Minimax Optimal Regret Bound for Reinforcement Learning with Trajectory Feedback

Zihan Zhang (University of Washington), Ruosong Wang (Peking University)

Reinforcement LearningSequential

🎯 What it does: This study explores reinforcement learning (RL) with trajectory feedback and proposes an algorithm that achieves an asymptotically near-optimal regret bound of O(√SAHK³) over K experiments.

Minimum Width for Universal Approximation using Squashable Activation Functions

Jonghyun Shin (Korea University), Sejun Park (Korea University)

🎯 What it does: This paper studies the minimum width required for deep neural networks using general activation functions to achieve Lp approximation.

MIPT: Multilevel Informed Prompt Tuning for Robust Molecular Property Prediction

yeyunchen, Jiangming Shi (Xiamen University)

Drug DiscoveryGraph Neural NetworkContrastive LearningGraph

🎯 What it does: A multi-layer information prompt tuning framework (MIPT) is proposed, which effectively transfers pre-trained graph neural networks to molecular property prediction tasks through lightweight prompt learning, LoRA low-rank adaptation, layer prompts, and a noise penalty mechanism.

MiraGe: Editable 2D Images using Gaussian Splatting

Joanna Waczynska, Przemysław Spurek (Jagiellonian University)

RestorationGenerationGaussian SplattingImage

🎯 What it does: Reconstructing and editing two-dimensional images in three-dimensional space using planar three-dimensional Gaussian distribution;

Mirror, Mirror of the Flow: How Does Regularization Shape Implicit Bias?

Tom Jacobs (CISPA Helmholtz Center for Information Security), Rebekka Burkholz (CISPA Helmholtz Center for Information Security)

TransformerFlow-based ModelImageText

🎯 What it does: This paper studies the interaction between explicit regularization (such as weight decay) and implicit bias, proposing to incorporate explicit regularization into the mirror flow framework and analyzing the geometric effects of training dynamics through time-dependent Legendre functions.

MIRROR: Make Your Object-Level Multi-View Generation More Consistent with Training-Free Rectification

Tianchi Xing (University of Chinese Academy of Sciences), Tiande Guo (University of Chinese Academy of Sciences)

GenerationData SynthesisDepth EstimationDiffusion modelImage

🎯 What it does: A training-free, pluggable multi-view generation consistency correction framework called MIRROR is proposed to address the Janus problem and inconsistencies arising from multi-view diffusion models.

MissScore: High-Order Score Estimation in the Presence of Missing Data

Wenqin Liu (University of Melbourne), Mingming Gong (University of Melbourne)

Score-based ModelTabular

🎯 What it does: Proposes the MissScore framework, which directly learns high-order (e.g., second-order) data distribution gradients (scores) under missing data conditions, avoiding the bias of traditional imputation followed by learning;

Mitigating Heterogeneous Token Overfitting in LLM Knowledge Editing

Tianci Liu (Purdue University), Jing Gao (Purdue University)

TransformerLarge Language ModelText

🎯 What it does: A knowledge editing method named OVERTONE is proposed, aimed at alleviating the heterogeneous token overfitting problem that large language models (LLMs) encounter during knowledge editing by adopting an adaptive smoothing target distribution for each token.

Mitigating Local Cohesion and Global Sparseness in Graph Contrastive Learning with Fuzzy Boundaries

Yuena Lin (Beijing University of Technology), Gengyu Lyu (Beijing University of Technology)

OptimizationRepresentation LearningGraph Neural NetworkAuto EncoderContrastive LearningGraph

🎯 What it does: Addressing the issues of local cohesion and global sparsity in graph contrastive learning, fuzzy boundaries are introduced to improve node representation.

Mitigating Object Hallucination in Large Vision-Language Models via Image-Grounded Guidance

Linxi Zhao (Cornell University), Quanquan Gu (University of California)

Object DetectionGenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality

🎯 What it does: The MARINE framework is proposed, which reduces the object hallucination problem of large visual language models by incorporating image-based guidance during the inference phase.

MITIGATING OVER-EXPLORATION IN LATENT SPACE OPTIMIZATION USING LES

Omer Ronen (University of California Berkeley), Bin Yu (University of California Berkeley)

OptimizationRecurrent Neural NetworkTransformerAuto EncoderSequential

🎯 What it does: Proposes the Latent Exploration Score (LES) as a constraint added to Latent Space Optimization (LSO) to reduce excessive exploration of invalid solutions in discrete black-box optimization.

Mitigating Over-Squashing in Graph Neural Networks by Spectrum-Preserving Sparsification

Langzhang Liang (Korea Advanced Institute of Science and Technology), Kijung Shin (Korea Advanced Institute of Science and Technology)

OptimizationGraph Neural NetworkNeural Radiance FieldGraph

🎯 What it does: This paper proposes a graph reconnection method called GOKU based on spectral-preserving sparsification to alleviate the over-compression problem in graph neural networks.

Mitigating Plasticity Loss in Continual Reinforcement Learning by Reducing Churn

Hongyao Tang (Mila Quebec AI Institute), Glen Berseth (Mila Quebec AI Institute)

Reinforcement LearningSequential

🎯 What it does: This paper explains and addresses the issue of 'plasticity loss' in long-term learning, which refers to the network's loss of adaptability to new tasks, by studying 'churn' (the fluctuation of network outputs on untrained samples) in deep continual reinforcement learning.

MixBridge: Heterogeneous Image-to-Image Backdoor Attack through Mixture of Schrödinger Bridges

Shixi Qin (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)

GenerationData SynthesisAdversarial AttackMixture of ExpertsDiffusion modelImageStochastic Differential Equation

🎯 What it does: This paper proposes MixBridge, an image-to-image diffusion model based on the Schrödinger bridge, which simultaneously achieves clean generation and various hidden attacks within the same model.

Mixed-curvature decision trees and random forests

Philippe Chlenski (Columbia University), Itsik Pe'er (Columbia University)

ClassificationOptimizationTabularTime Series

🎯 What it does: A general framework for decision tree and random forest learning on mixed curvature product manifolds is proposed, utilizing angular projection in two-dimensional subspaces to achieve decision splits;

MixMin: Finding Data Mixtures via Convex Minimization

Anvith Thudi (University of Toronto), Chris J. Maddison (Stanford University)

OptimizationText

🎯 What it does: A mixed data source method called MixMin based on convex optimization is proposed, which learns the optimal data mixing ratio using surrogate models and downstream task performance.

Mixture of Experts Made Intrinsically Interpretable

Xingyi Yang (University of Oxford), Philip Torr (University of Oxford)

Explainability and InterpretabilityTransformerMixture of ExpertsText

🎯 What it does: An interpretable mixture of experts language model MoE-X is proposed, which eliminates multi-semantic neurons by using ReLU and sparse routing among experts.

Mixture of Experts Provably Detect and Learn the Latent Cluster Structure in Gradient-Based Learning

Ryotaro Kawata (University of Tokyo), Taiji Suzuki (University of Tokyo)

Mixture of Experts

🎯 What it does: It is proven that in regression tasks with potential clustering structures, a single neural network cannot learn this structure, while a mixture of experts network (MoE) can successfully learn it by routing samples to different experts; the sample and time complexity of MoE under gradient descent is also provided.

Mixture of Hidden-Dimensions: Not All Hidden-States’ Dimensions are Needed in Transformer

Yilong Chen (Institute of Information Engineering, Chinese Academy of Sciences), Haifeng Wang (Baidu Inc.)

TransformerMixture of ExpertsText

🎯 What it does: This study investigates the sparse activation features of Transformer hidden dimensions and proposes the MOHD (Mixture of Hidden Dimensions) architecture, which achieves sparse activation and scalability/compression of hidden dimensions through shared sub-dimensions and dedicated sub-dimensions with dynamic routing.

Mixture of Lookup Experts

Shibo Jie (Peking University), Yunhe Wang (Huawei)

Mixture of ExpertsText

🎯 What it does: This paper proposes the Mixture of Lookup Experts (MoLE), which reparameterizes expert parameters as a lookup table (LUT), allowing for inference without loading experts, thereby significantly reducing memory usage and latency.

ML$^2$-GCL: Manifold Learning Inspired Lightweight Graph Contrastive Learning

Jianqing Liang (Shanxi University), Zhiqiang Wang (Shanxi University)

ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph

🎯 What it does: A lightweight graph contrastive learning framework ML²-GCL based on manifold learning is proposed, which obtains positive and negative sample weights through local linear reconstruction under a single view, avoiding the overhead of graph augmentation and a large number of contrast pairs.

MM-RLHF: The Next Step Forward in Multimodal LLM Alignment

YiFan Zhang, Rong Jin

OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageVideoTextMultimodalityBenchmark

🎯 What it does: A dataset of 120k fine-grained human preference comparison pairs, termed MM-RLHF, was constructed, and based on this, a new reward model and alignment algorithm were proposed to comprehensively enhance the alignment effectiveness of multimodal large language models (MLLMs).

MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency

Dongzhi Jiang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)

Large Language ModelMultimodalityBenchmarkChain-of-Thought

🎯 What it does: This paper proposes the MME-CoT benchmark for systematically evaluating the chain-of-thought reasoning capabilities of large multimodal models.

MMedPO: Aligning Medical Vision-Language Models with Clinical-Aware Multimodal Preference Optimization

Kangyu Zhu (University of North Carolina at Chapel Hill), Huaxiu Yao (University of North Carolina at Chapel Hill)

OptimizationTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityBiomedical Data

🎯 What it does: Proposes MMedPO, a multimodal preference optimization method that uses clinical relevance as a weight to enhance the factuality of medical visual language models by generating preference samples with hallucinations and local lesion noise.

MMInference: Accelerating Pre-filling for Long-Context Visual Language Models via Modality-Aware Permutation Sparse Attention

Yucheng Li (University of Surrey), Lili Qiu (Microsoft Corporation)

RetrievalComputational EfficiencyTransformerVision Language ModelVideoTextMultimodality

🎯 What it does: The MMInference method is proposed, which accelerates the pre-fill stage of long-context visual language models through permutation-based dynamic sparse attention with multimodal perception.

MODA: MOdular Duplex Attention for Multimodal Perception, Cognition, and Emotion Understanding

Zhicheng Zhang (Nankai University), Jufeng Yang (Nankai University)

RecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningMultimodality

🎯 What it does: Proposes MODA (MOdular Duplex Attention) to address the issues of cross-modal attention inconsistency and hierarchical decay in multimodal large language models, enhancing fine-grained perception, cognition, and emotional understanding.

Modalities Contribute Unequally: Enhancing Medical Multi-modal Learning through Adaptive Modality Token Re-balancing

Jie Peng (University of Science and Technology of China), Tianlong Chen (University of North Carolina at Chapel Hill)

TransformerMixture of ExpertsContrastive LearningMultimodalityBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease

🎯 What it does: An Adaptive Modal Marking Rebalancing (AMC) method is proposed to dynamically assess and fuse the importance of different modalities in medical multimodal learning.

Model Immunization from a Condition Number Perspective

Amber Yijia Zheng (Purdue University), Raymond A. Yeh (Purdue University)

OptimizationTransformerImageTabular

🎯 What it does: A theoretical framework for model immunity is proposed from the perspective of condition number analysis and implementation, aiming to make the model difficult to fine-tune for harmful tasks while remaining unaffected by normal tasks after pre-training.

Model Steering: Learning with a Reference Model Improves Generalization Bounds and Scaling Laws

Xiyuan Wei (Texas A&M University), Tianbao Yang (Texas A&M University)

OptimizationRepresentation LearningTransformerContrastive LearningImageText

🎯 What it does: A model steering framework based on a reference model is proposed and applied to CLIP pre-training.

Model Swarms: Collaborative Search to Adapt LLM Experts via Swarm Intelligence

Shangbin Feng (University of Washington), Tomas Pfister (Google)

OptimizationTransformerLarge Language ModelText

🎯 What it does: Proposes MODEL SWARMS, a collaborative search algorithm based on particle swarm optimization, designed for adaptive large language models in low data scenarios;

Model Uncertainty Quantification by Conformal Prediction in Continual Learning

Rui Gao (Wuhan University), Weiwei Liu (Wuhan University)

Convolutional Neural NetworkImage

🎯 What it does: A continuous learning uncertainty quantification method based on conformal prediction (CPCL) is proposed, which constructs a calibration set by replaying past task samples and uses Quantile Regression Forests (QRF) to estimate conditional quantiles, generating prediction intervals with asymptotic coverage guarantees.

Model-Based Exploration in Monitored Markov Decision Processes

Alireza Kazemipour (University of Alberta), Michael Bowling (University of Alberta)

OptimizationReinforcement LearningTabularBenchmark

🎯 What it does: This paper proposes a new model-based algorithm Monitored MBIE-EB to address the partially observable reward Monitored Markov Decision Process (Mon-MDP) problem, and provides the first theoretical guarantee for finite sample complexity.

Modeling All-Atom Glycan Structures via Hierarchical Message Passing and Multi-Scale Pre-training

Minghao Xu (Peking University), Wentao Zhang (Peking University)

Drug DiscoveryGraph Neural NetworkSupervised Fine-TuningGraph

🎯 What it does: A full-atomic-level glycan graph model GlycanAA was developed, and through multi-scale masked pre-training, PreGlycanAA was obtained, enhancing the representation of glycan structures and the performance of downstream tasks.

Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent

Yongxian Wei (Tsinghua University), Xiaochun Cao (Shenzhen Campus of Sun Yat-sen University)

OptimizationComputational EfficiencyTransformerSupervised Fine-TuningImageText

🎯 What it does: This paper proposes a framework that treats model merging as a constrained optimization problem. It optimizes the correction term of the task vector in a data-independent manner within a shared subspace using Adaptive Projection Gradient Descent (DOGE), aiming to make the merged model's performance on each task as close as possible to that of individually fine-tuned models while retaining shared knowledge.

Models of Heavy-Tailed Mechanistic Universality

Liam Hodgkinson (University of Melbourne), Michael W. Mahoney (University of California)

Image

🎯 What it does: This paper studies the heavy-tailed spectral distribution that arises in deep learning and proposes the High-Temperature Markov-Papoul (HTMP) random matrix family to describe the spectral density of the trained feature matrix, NTK, and weight matrix.

Modified K-means Algorithm with Local Optimality Guarantees

Mingyi Li (University of Tokyo), Akiko Takeda (University of Tokyo)

OptimizationTabular

🎯 What it does: The theoretical analysis of the local optimality of the K-means algorithm is conducted, providing counterexamples to prove that traditional assumptions do not hold, and a lightweight variant called LO-K-means is proposed, which guarantees convergence to continuous or discrete local optimal solutions while maintaining the same time complexity.

Modular Duality in Deep Learning

Jeremy Bernstein (Massachusetts Institute of Technology), Laker Newhouse (Massachusetts Institute of Technology)

OptimizationTransformerImage

🎯 What it does: This paper proposes a modular dualization method for recursively constructing the gradient dual mapping of neural networks to unify and improve existing training optimizers.

Modularized Self-Reflected Video Reasoner for Multimodal LLM with Application to Video Question Answering

Zihan Song (Tsinghua University), Wenwu Zhu (Tsinghua University)

OptimizationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoMultimodality

🎯 What it does: A modular self-reflective video reasoning framework (MSR-ViR) has been designed and implemented, combining the spatiotemporal grounding module (MoST-Grounding) with a multimodal LLM to achieve interpretable video question answering.

Modulated Diffusion: Accelerating Generative Modeling with Modulated Quantization

Weizhi Gao (North Carolina State University), Xiaorui Liu (North Carolina State University)

GenerationData SynthesisComputational EfficiencyDiffusion modelImage

🎯 What it does: This paper proposes a general framework named MoDiff, which accelerates the sampling process of diffusion models using modulated quantization and error compensation techniques, and is compatible with existing caching and post-training quantization methods.

MODULI: Unlocking Preference Generalization via Diffusion Models for Offline Multi-Objective Reinforcement Learning

Yifu Yuan (Tianjin University), Jianye HAO

Reinforcement LearningDiffusion modelTabularBenchmark

🎯 What it does: We propose MODULI, an offline multi-objective reinforcement learning framework based on conditional diffusion models, which generates trajectories consistent with given preferences and extracts actions to achieve policy planning for any preference.

MoE-SVD: Structured Mixture-of-Experts LLMs Compression via Singular Value Decomposition

Wei Li (University of Birmingham), Yike Guo (Hong Kong University of Science and Technology)

CompressionComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText

🎯 What it does: Proposes MoE-SVD, a compression framework that does not require additional training, utilizing SVD for low-rank decomposition of the expert layers in Mixture-of-Experts LLMs to achieve parameter compression and inference acceleration.

MoEQuant: Enhancing Quantization for Mixture-of-Experts Large Language Models via Expert-Balanced Sampling and Affinity Guidance

Zhixuan Chen (Houmo AI), JiangyongYu

Large Language ModelMixture of ExpertsText

🎯 What it does: This paper proposes the MoEQuant framework, which implements low-bit post-training quantization for Mixture-of-Experts large models, addressing the error issues caused by expert imbalance.

MOGIC: Metadata-infused Oracle Guidance for Improved Extreme Classification

Suchith Chidananda Prabhu (Indian Institute of Technology Delhi), Manik Varma (Microsoft)

ClassificationTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: The MOGIC framework is proposed, employing a two-stage training approach that utilizes an oracle containing metadata from both the query side and the label side to guide a low-latency extreme classification model, achieving higher prediction accuracy.

MoH: Multi-Head Attention as Mixture-of-Head Attention

Peng Jin (Peking University), Shuicheng YAN

ClassificationGenerationComputational EfficiencyTransformerMixture of ExpertsImage

🎯 What it does: Proposes Mixture-of-Head Attention (MoH), which dynamically routes each token to activate only a portion of the heads in the multi-head attention of the Transformer and replaces standard summation with weighted summation, significantly reducing computational load;

MoHAVE: Mixture of Hierarchical Audio-Visual Experts for Robust Speech Recognition

Sungnyun Kim (KAIST), Se-Young Yun (KAIST)

RecognitionTransformerMixture of ExpertsMultimodalityAudio

🎯 What it does: Proposes MoHAVE - a hierarchical sparse Mixture-of-Experts structure designed to achieve scalable and robust audio-visual speech recognition.

Moirai-MoE: Empowering Time Series Foundation Models with Sparse Mixture of Experts

Xu Liu (Salesforce AI Research), Doyen Sahoo (Salesforce AI Research)

TransformerMixture of ExpertsTime Series

🎯 What it does: This paper presents MOIRAI-MOE, a time series pre-training foundation model that utilizes sparse mixture of experts (MoE) technology.

MoMa: Modulating Mamba for Adapting Image Foundation Models to Video Recognition

Yuhuan Yang (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)

RecognitionOptimizationComputational EfficiencyTransformerSupervised Fine-TuningVideo

🎯 What it does: The pre-trained image foundation model (such as CLIP) is adapted into a video recognition model through parameter-efficient fine-tuning, utilizing Mamba's state space model and the SeqMod mechanism.

Momentum-Driven Adaptivity: Towards Tuning-Free Asynchronous Federated Learning

Wenjing Yan (Chinese University of Hong Kong), Ying-Jun Angela Zhang (Chinese University of Hong Kong)

OptimizationFederated LearningImage

🎯 What it does: Two asynchronous federated learning algorithms, MasFL and AdaMasFL, are proposed to address data heterogeneity and hyperparameter tuning issues.

MONA: Myopic Optimization with Non-myopic Approval Can Mitigate Multi-step Reward Hacking

Sebastian Farquhar (Google DeepMind), Rohin Shah (Google DeepMind)

OptimizationReinforcement LearningTabular

🎯 What it does: This paper proposes and validates a method that combines myopic optimization with foresight approval (MONA) to prevent reinforcement learning agents from performing multi-step reward hacking without being detected by supervisors.

Monte Carlo Tree Diffusion for System 2 Planning

Jaesik Yoon (KAIST), Sungjin Ahn (KAIST)

Diffusion modelImage

🎯 What it does: A scalable system 2 planning framework (MCTD) is proposed by combining diffusion models with Monte Carlo tree search.

Monte Carlo Tree Search for Comprehensive Exploration in LLM-Based Automatic Heuristic Design

Zhi Zheng (National University of Singapore), Bryan Hooi (National University of Singapore)

OptimizationTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: Using Monte Carlo Tree Search (MCTS) to replace traditional population evolution, a complete heuristic function tree is constructed, utilizing LLM to generate and improve heuristics, and achieving a more comprehensive search through the tree structure;

Monte-Carlo Tree Search with Uncertainty Propagation via Optimal Transport

Tuan Quang Dam (Hanoi University of Science and Technology), Odalric-Ambrym Maillard (University of Lille)

Reinforcement LearningTabular

🎯 What it does: A Monte-Carlo Tree Search algorithm based on L1-Wasserstein barycenter and α-divergence for distributed backup strategy—Wasserstein MCTS, utilizes a Gaussian node model to propagate uncertainty in the tree.

MoRAgent: Parameter Efficient Agent Tuning with Mixture-of-Roles

Jing Han (Beijing University of Posts and Telecommunications), Ying Nie (Huawei Noah's Ark Lab)

Supervised Fine-TuningAgentic AIText

🎯 What it does: Proposes MoRAgent, which implements parameter-efficient agent task fine-tuning using a low-rank adapter with Mixture-of-Roles.

More Than Meets the Eye: Enhancing Multi-Object Tracking Even with Prolonged Occlusions

Bishoy Galoaa (Northeastern University), Sarah Ostadabbas (Northeastern University)

Object TrackingTransformerGaussian SplattingOptical FlowVideo

🎯 What it does: The MOTE (MOre Than meets the Eye) framework is proposed to specifically address the long-term occlusion problem in multi-object tracking, maintaining trajectory consistency and reducing identity switches.

Morse: Dual-Sampling for Lossless Acceleration of Diffusion Models

Chao Li (Intel Labs China), Anbang Yao (Intel Labs China)

GenerationData SynthesisComputational EfficiencyDiffusion modelImage

🎯 What it does: In this work, the authors propose a dual-sampling framework named Morse, which achieves lossless acceleration of the generative process of diffusion models by combining jump sampling with residual feedback.

MP-Nav: Enhancing Data Poisoning Attacks against Multimodal Learning

Jingfeng Zhang (New York University), Farshad Khorrami (New York University)

Adversarial AttackData-Centric LearningVision Language ModelImageTextMultimodality

🎯 What it does: This paper proposes the MP-Nav module, which systematically selects concept pairs and instances in multimodal learning to enhance the effectiveness of data poisoning attacks.

MPO: An Efficient Post-Processing Framework for Mixing Diverse Preference Alignment

Tianze Wang (Rutgers University), Linjun Zhang (Rutgers University)

OptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningText

🎯 What it does: A post-processing framework MPO is proposed, which aggregates existing single-objective policies to achieve multi-preference alignment.

MTL-UE: Learning to Learn Nothing for Multi-Task Learning

Yi Yu (Nanyang Technological University), Alex Kot

SegmentationAdversarial AttackMeta LearningTransformerAuto EncoderImage

🎯 What it does: This paper proposes and implements MTL-UE, a framework for generating unlearnable samples for multi-task learning, which significantly degrades the performance of MTL and STL models under unauthorized training.

MTSTRec: Multimodal Time-Aligned Shared Token Recommender

Ming-Yi Hong (National Taiwan University), Che Lin (National Taiwan University)

Recommendation SystemTransformerLarge Language ModelMultimodalitySequential

🎯 What it does: This paper proposes MTSTRec, a multi-modal time-aligned shared token recommendation framework that utilizes multi-modal information such as product ID, images, text, and price for sequential recommendations.

MUDDFormer: Breaking Residual Bottlenecks in Transformers via Multiway Dynamic Dense Connections

Da Xiao (Beijing University of Posts and Telecommunications), Xingyuan Yuan (ColorfulClouds Technology Company Limited)

TransformerLarge Language ModelText

🎯 What it does: Proposed a Multi-path Dynamic Dense Connection (MUDD) and integrated it into the Transformer to construct the MUDDFormer model;

MuLan: Adapting Multilingual Diffusion Models for Hundreds of Languages with Negligible Cost

Sen Xing (Tsinghua University), Wenhai Wang (OpenGVLab Shanghai AI Laboratory)

GenerationData SynthesisTransformerVision Language ModelDiffusion modelImageText

🎯 What it does: This paper presents MuLan, a lightweight language adapter that enables the model to support image generation in over 110 languages using only a small amount of English text-image data, while keeping the text encoder and diffusion model frozen.

Multi-agent Architecture Search via Agentic Supernet

Guibin Zhang (National University of Singapore), Xiang Wang (University of Science and Technology of China)

OptimizationComputational EfficiencyNeural Architecture SearchLarge Language ModelAgentic AIMixture of ExpertsText

🎯 What it does: An automated multi-agent architecture search framework called MaAS based on agentic supernet is proposed, which can dynamically sample the most suitable multi-agent systems for queries of varying difficulty and domains, significantly improving performance and reducing inference costs.

Multi-Armed Bandits with Interference: Bridging Causal Inference and Adversarial Bandits

Su Jia (Cornell University), Nathan Kallus (Cornell University)

OptimizationReinforcement Learning

🎯 What it does: This paper proposes a framework for Multi-Armed Bandits with Interference (MABI) and designs a clustering randomization strategy based on EXP3-IX and robust random partitioning, which can achieve a high probability loss upper bound that vanishes with the experimental scale N under optimal expected loss;