ICML 2026 Papers — Page 35
International Conference on Machine Learning · 6554 papers
Memory-Efficient LLM Training with Dynamic Sparsity: From Stability to Practical Scaling
Qiao Xiao (Eindhoven University of Technology), Torsten Hoefler (ETH Zuerich)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a dynamic sparse training framework called SMET, which stabilizes training through optimizer-state warm-up and density-aware learning rate scaling, and reduces memory consumption by storing only gradients and optimizer states on sparse parameters.
MemoryBench: A Benchmark for Memory and Continual Learning in LLM Systems
Qingyao Ai (Tsinghua University), Yiqun LIU
Federated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the MemoryBench benchmark to evaluate the memory and continuous learning capabilities of LLM systems;
MemoryLLM: Plug-n-Play Interpretable Feed-Forward Memory for Transformers
AJAY KUMAR JAISWAL, Minsik Cho (Apple)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes MemoryLLM, a variant of the Transformer that decouples the feed-forward network (FFN) from self-attention and trains the FFN separately on word embeddings, allowing the FFN to function as an interpretable word-level key-value retrieval storage. It also achieves significant improvements in memory and computational efficiency by precomputing Token-wise Lookup (ToL), and designs Flex-MemoryLLM to reduce the performance gap with traditional LLMs while maintaining interpretability.
MentisOculi: Revealing the Limits of Reasoning with Mental Imagery
Jana Ricarda Zeller (Max-Planck-Institute for Intelligent Systems), Wieland Brendel (Max-Planck-Institute for Intelligent Systems)
TransformerLarge Language ModelVision Language ModelImageVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the MENTISOCULI benchmark to evaluate models' mental imagery capabilities in multi-step visual reasoning.
MEnvAgent: Scalable Polyglot Environment Construction for Verifiable Software Engineering
Chuanzhe Guo (Harbin Institute of Technology), Haifeng Wang (Baidu Inc.)
Computational EfficiencyData-Centric LearningAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A multi-agent based automated framework for verifiable software engineering environments, named MEnvAgent, was constructed, supporting the construction and reuse of executable environments for 10 mainstream programming languages;
MePo: Meta Post-Refinement for Rehearsal-Free General Continual Learning
Guanglong Sun (Tsinghua University), Yi Zhong (Tsinghua University)
ClassificationRepresentation LearningMeta LearningTransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningImage
🎯 What it does: Propose the MePo (Meta Post-Refinement) framework, which uses dual-level meta-learning with late-stage refinement of pre-trained models and pseudo-task sequences to enhance representation learning and output alignment in general continual learning (GCL) without replay.
MER-DG: Modality-Entropy Regularization for Multimodal Domain Generalization
Yavuz Yarici (Georgia Institute of Technology), Ghassan AlRegib (Georgia Institute of Technology)
Domain AdaptationContrastive LearningMultimodality
🎯 What it does: Proposes Modality-Entropy Regularization for Domain Generalization (MER-DG), which alleviates Fusion Overfitting in multi-modal domain generalization by maximizing the feature distribution entropy of each modality encoder, thereby improving cross-domain performance.
Merge to Remember: Sharpness-Aware Isotropic Merging for Continual Learning
Qun Yang (National University of Defense Technology), Long Lan (National University of Defense Technology)
ClassificationOptimizationFederated LearningKnowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningMixture of ExpertsAuto EncoderContrastive LearningImageTextBenchmark
🎯 What it does: Proposes the Sharpness-Aware Isotropic Merging (SAIM) framework, unifying fine-tuning and model merging in continual learning. It employs Sharpness-Aware Block Coordinate Descent (SA-BCD) during the fine-tuning phase to preemptively defend against parameter drift caused by merging, and enhances subspace alignment by dynamically balancing the singular value spectrum through adaptive isotropic merging, thereby reducing catastrophic forgetting.
MergeMix: Optimizing Mid-Training Data Mixtures via Learnable Model Merging
jiapeng wang, JUN ZHOU
OptimizationFederated LearningKnowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsContrastive LearningTextTabularSequentialBenchmark
🎯 What it does: Propose the MergeMix method, which uses expert model weight interpolation as a low-cost proxy, directly optimizing the data mixing ratio during the mid-training phase on downstream tasks;
MESA: Improving MoE Safety Alignment via Decentralized Expertise
Yitong Sun (Beihang University), Xingxing Wei (Beihang University)
Federated LearningSafty and PrivacyComputational EfficiencyReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: To address the safety alignment issue in MoE models, this paper proposes the MESA framework, which enhances safety while maintaining practical performance through decentralized expert allocation and dynamic routing refinement.
Mesh Based Simulations with Spatial and Temporal awareness
Paul Garnier (CEMEF Mines Paris PSL), Elie Hachem (CEMEF Mines Paris PSL)
OptimizationComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerMeshGraphTime SeriesPhysics Related
🎯 What it does: Proposes a unified framework to address the bottleneck of node supervision and explicit Euler time steps in traditional graph neural networks for grid simulations;
Mesh Field Theory: Port–Hamiltonian Formulation of Mesh-Based Physics
Satoshi Noguchi (JAMSTEC), Yoshinobu Kawahara (University of Osaka)
Graph Neural NetworkReinforcement LearningDiffusion modelScore-based ModelContrastive LearningMeshGraphPhysics Related
🎯 What it does: Propose Mesh Field Theory (MeshFT) and its neural implementation MeshFT-Net, achieving local port-Hamiltonian decomposition of mesh physical dynamics;
MeshTok: Efficient Multi-Scale Tokenization for Scalable PDE Transformers
Yanshun Zhao (University of Science and Technology of China), Jingrun Chen (University of Science and Technology of China)
OptimizationComputational EfficiencyTransformerMixture of ExpertsContrastive LearningMeshTabularTime SeriesBenchmarkPhysics Related
🎯 What it does: Propose MeshTok, a Transformer-based adaptive multi-scale tokenization framework for PDE solvers;
Message Passing on the Edge: Towards Scalable and Expressive GNNs
Pablo Barcelo (Pontificia Universidad Católica de Chile), Cristobal Rojas (Pontificia Universidad Católica de Chile)
Computational EfficiencyRepresentation LearningGraph Neural NetworkContrastive LearningGraphTabularBiomedical Data
🎯 What it does: This paper proposes an edge-based message passing framework EB-1WL and the corresponding GNN structure EB-GNN, aiming to enhance the expressiveness of graph neural networks.
Message Tuning Outshines Graph Prompt Tuning: A Prismatic Space Perspective
Yancheng Chen (Academy of Mathematics and Systems Science, Chinese Academy of Sciences), Chuan Zhou (Academy of Mathematics and Systems Science, Chinese Academy of Sciences)
ClassificationGraph Neural NetworkSupervised Fine-TuningPrompt EngineeringGraph
🎯 What it does: This paper proposes a framework for measuring the adaptability of graph models based on Prismatic Space Theory, and designs the Message Tuning method on this basis to enhance the adaptability of graph foundational models in downstream tasks.
MET-Bench: Multimodal Entity Tracking for Evaluating the Limitations of Vision-Language and Reasoning Models
Vanya Cohen (University of Texas at Austin), Ray Mooney
Object TrackingTransformerReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed and evaluated the multi-modal entity state tracking benchmark MET-Bench, exploring the tracking capabilities of vision-language models under text and image updates.
Meta Context Engineering via Agentic Skill Evolution
Haoran Ye (Peking University), Guojie Song (Peking University)
OptimizationFederated LearningMeta LearningDrug DiscoveryAI Code AssistantTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextFinance Related
🎯 What it does: Propose the Meta Context Engineering (MCE) framework, which jointly optimizes context engineering skills and context content through a two-layer evolutionary approach to enhance the reasoning performance of large language models across multiple domains.
Meta Flow Maps enable scalable reward alignment
Peter Potaptchik (University of Oxford), Yee Whye Teh (University of Oxford)
GenerationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerDiffusion modelScore-based ModelFlow-based ModelImageMultimodalityStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose Meta Flow Maps (MFM), a stochastic flow mapping, for achieving scalable reward alignment in generative models, supporting steering during inference and offline fine-tuning.
Meta-Black-Box Optimization Can Do Search Guidance for Expensive Constrained Multi-Objective Optimization
Yukun Du (National University of Defense Technology), Shengkun Chang (National University of Defense Technology)
OptimizationMeta LearningTransformerReinforcement LearningDiffusion modelTabularBenchmark
🎯 What it does: A MetaSG-SAEA framework based on meta black-box optimization is proposed, which uses a meta policy to guide the search area for costly constrained multi-objective optimization.
Meta-iLaD: Identifiable Latent Dynamics via Meta-Learning of Dynamics Environments
Yubo Ye (Rochester Institute of Technology), Linwei Wang (Rochester Institute of Technology)
Meta LearningRecurrent Neural NetworkTransformerFlow-based ModelTime SeriesSequentialPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes the Meta-iLaD framework, which achieves identifiable latent dynamics models by performing meta-learning on dynamic environments;
Meta-learning Structure-Preserving Dynamics
Cheng Jing (Arizona State University), Kookjin Lee (Arizona State University)
Meta LearningTabularTime SeriesBenchmarkPhysics Related
🎯 What it does: This paper proposes a meta-learning framework based on modulation for learning dynamic models that preserve structural properties (such as energy conservation and dissipation).
MetaDNS: Enhancing Exploration in Discrete Neural Samplers via Metadynamics
Xiaochen Du (Massachusetts Institute of Technology), Rafael Gomez-Bombarelli
OptimizationComputational EfficiencyRecurrent Neural NetworkTransformerDiffusion modelScore-based ModelContrastive LearningGraphTabularSequentialPhysics Related
🎯 What it does: Propose the MetaDNS framework, combining well-tempered metadynamics with a discrete neural sampler, significantly enhancing sampling exploration capabilities for low-temperature multi-modal distributions;
MetaMoE: Diversity-Aware Proxy Selection for Privacy-Preserving Mixture-of-Experts Unification
Weisen Jiang (Chinese University of Hong Kong), Sinno Jialin Pan (Chinese University of Hong Kong)
Federated LearningSafty and PrivacyRepresentation LearningTransformerSupervised Fine-TuningMixture of ExpertsContrastive LearningImageTextMultimodality
🎯 What it does: Propose a privacy-preserving Mixture-of-Experts unified framework called MetaMoE, which can merge individually trained expert models into a single deployable MoE model without sharing private data.
MetaOthello: A Controlled Study of Multiple World Models in Transformers
Aviral Chawla (University of Vermont), Juniper L Lovato (University of Michigan)
Explainability and InterpretabilityRepresentation LearningMeta LearningTransformerPrompt EngineeringMixture of ExpertsContrastive LearningWorld ModelText
🎯 What it does: Construct the MetaOthello framework to generate multiple variants of Othello with the same rules but different update logic, train a small GPT model on mixed data, and analyze the organization of its internal world models through linear probes and alignment/intervention analysis.
MetaPerch: Learning from metadata for bioacoustics foundation models
Mustafa Chasmai (University of Massachusetts Amherst), Jenny Hamer (Google DeepMind)
RecognitionHyperparameter SearchConvolutional Neural NetworkTransformerContrastive LearningAudio
🎯 What it does: Propose METAPERCH, a foundational model, which enhances the performance of acoustic species identification by incorporating various audio metadata (such as location, season, background species, etc.) as auxiliary supervision tasks during training.
MetaphorVU: Towards Metaphorical Video Understanding
Zhuoqun Li (Chinese Academy of Sciences), Le Sun (Chinese Academy of Sciences)
Explainability and InterpretabilityKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the MetaphorVU-Bench benchmark specifically for metaphor video understanding, systematically evaluating the capabilities of existing MLLMs in metaphor video reasoning, and introduced the MetaphorBoost approach based on a metaphor knowledge graph, enhancing cross-domain mapping during reasoning.
MetaStreet: Semi-Supervised Multimodal Learning for Street-Level Socioeconomic Prediction
Meng Chen (Shandong University), Weiming Huang (University of Leeds)
ClassificationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityGraph
🎯 What it does: Propose MetaStreet, a semi-supervised multi-modal framework for predicting street-level socio-economic indicators based on street view images.
Metis: Learning to Jailbreak LLMs via Self-Evolving Metacognitive Policy Optimization
Huilin Zhou (University of Science and Technology of China), Xuelong Li (China Telecom)
Explainability and InterpretabilityComputational EfficiencyAdversarial AttackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the Metis framework, which transforms jailbreak attacks on large language models into a strategy optimization problem during inference by utilizing a self-evolving meta-cognitive loop, enabling multi-round adaptive attacks.
Metric—Phase Fields: Decoupling Distance and Sign for Thin-Structure Reconstruction from Unoriented Point Clouds
Jiayi Kong (Nanyang Technological University), Ying He (Nanyang Technological University)
GenerationRepresentation LearningConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelNeural Radiance FieldContrastive LearningPoint CloudMesh
🎯 What it does: Propose a novel implicit representation called Metric-Phase Fields (MPF), which can reconstruct thin structures, open boundaries, and non-manifold geometries from unoriented point clouds.
MFH-NAS:A Hybrid Neural Architecture Search Framework for Multimodal Fusion Object Detection
Quanwei Gao (Northwest A&F University), Zirui Luo (Northwest A&F University)
Object DetectionNeural Architecture SearchConvolutional Neural NetworkTransformerContrastive LearningImageMultimodality
🎯 What it does: Propose a hybrid neural architecture search framework, MFH-NAS, for automatically discovering fusion structures in multi-modal fusion object detection, which optimizes fusion operators and explores cross-stage connections.
MGAL: A Multilingual Granularity-Aware Long-Context Benchmark
Chunhan Li (Hong Kong University of Science and Technology Guangzhou), Chengwei Qin (Hong Kong University of Science and Technology Guangzhou)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed MGAL, a multilingual, multi-grained, and position-aware benchmark for long-text evaluation, covering four granularities: word, sentence, paragraph, and document, and dividing the start/middle/end positions at the document and paragraph levels across the six official languages of the UN.
mHC: Manifold-Constrained Hyper-Connections
Zhenda Xie (DeepSeekAI), Wenfeng Liang (DeepSeekAI)
OptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsScore-based ModelContrastive LearningText
🎯 What it does: Proposes a manifold constraint method for Hyper-Connections (HC), called mHC, which restores the identity mapping of residuals by projecting residual mappings onto a doubly stochastic matrix (Birkhoff polytope), thereby enhancing training stability and scalability.
MICE-Bench: A Challenging and Comprehensive Benchmark for Multi-Reference Image Creation and Editing
Siqi Luo (Shanghai Jiao Tong University), Xiaohong Liu (Shanghai Jiao Tong University)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed MICE-Bench, a unified evaluation benchmark for multi-reference image generation and editing, along with corresponding evaluation scripts.
MidSteer: Optimal Affine Framework for Steering Generative Models
Tatiana Gaintseva (Queen Mary University of London), Ismail Elezi (Huawei Noah's Ark)
GenerationOptimizationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Propose MidSteer — a general optimal affine framework for concept elimination, switching, and directional switching in the intermediate representations of generative models; theoretically derive its connection to the existing LEACE method and provide a closed-form solution.
Midtraining Bridges Pretraining and Posttraining Distributions
Emmy Liu (Carnegie Mellon University), Chenyan Xiong (Carnegie Mellon University)
Domain AdaptationComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: Studied the effect of introducing midtraining (inserting specialized data between pretraining and post-training fine-tuning) in large language model training, systematically evaluated its performance across different domains, mixing ratios, and introduction timings, and compared it with continuous pretraining.
Milestone-Guided Policy Learning for Long-Horizon Language Agents
Zixuan Wang (Zhejiang University), Yongliang Shen (Zhejiang University)
TransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Studied improving reinforcement learning performance in long-horizon language agent tasks through a milestone-guided policy learning framework (BEACON).
MIMO-LP: A Multi-Input Multi-Output Framework for Subgraph-based Link Prediction
Yixin Song (Southeast University), Yujun Cai (University of Queensland)
Recommendation SystemComputational EfficiencyDrug DiscoveryGraph Neural NetworkTransformerMixture of ExpertsAuto EncoderContrastive LearningGaussian SplattingGraphTabularBenchmark
🎯 What it does: Propose a MIMO-LP framework that achieves efficient inference for subgraph-based link prediction by multiplexing message passing at the subgraph level.
MIMOMamba: From Scalar Duality to Matrix-Valued Attention
Yanbo Li (Chinese University of Hong Kong), Wenqiang Pu (Shenzhen Research Institute of Big Data)
Computational EfficiencyRepresentation LearningTransformerMixture of ExpertsAuto EncoderContrastive LearningTabularTime SeriesBenchmark
🎯 What it does: Propose MIMOMamba, which generalizes the state space bidirectionality (SSD) from single-input single-output (SISO) to multi-input multi-output (MIMO), achieving matrix-valued attention through matrix polynomial parameterization while maintaining linear time training.
Mind Dreamer: Untethering Imagination via Active Causal Intervention on Latent Manifolds
Shaojun Xu (Tsinghua University), Rong Zhao (Tsinghua University)
Recurrent Neural NetworkTransformerReinforcement LearningGenerative Adversarial NetworkContrastive LearningWorld ModelImageVideo
🎯 What it does: Propose the Mind Dreamer framework, which decouples historical entanglements on the latent manifold through active causal intervention, achieving more efficient model-based reinforcement learning.
Mind the Budget: Accelerating Deep Reinforcement Learning using Constrained Early Exit Neural Networks
Julien Brosseit, Joschka Boedecker (University of Freiburg)
Computational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkTransformerReinforcement LearningMixture of ExpertsContrastive LearningTabularTime Series
🎯 What it does: Propose Budgeted EXit Actor (BEXA), a method that integrates early exit neural networks (Early Exit Neural Network, ENN) into an off-policy actor-critic architecture, using linear programming to control exit decisions, achieving a trade-off between inference cost and performance.
Mind the Gap: Catching Hallucinations via Evidence Drop on the Reasoning Manifold
Qunjie Chen (Tongji University), Linye Li (Tongji University)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes a dynamic method based on 'Evidence Drop' to detect hallucinations in large language models during multi-step reasoning and to locate the erroneous steps.
Mind the Gap: Mixtures of Gaussians in Approximate Differential Privacy
Huikang Liu (Shanghai Jiao Tong University), Wolfram Wiesemann (Imperial Business School)
Safty and PrivacyGaussian SplattingTabularBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose a noise injection mechanism under (ε,δ) approximate differential privacy by mixing multiple Gaussian distributions with the same variance but different means, significantly reducing the expected noise magnitude and variance;
Mind the Gap: Structure-Aware Consistency in Preference Learning
Mehryar Mohri (Google Research), Yutao Zhong (Google Research)
Computational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper analyzes the effectiveness of the accidental approximation loss in preference learning for LLMs, and proves that under common neural network hypothesis sets, the unconstrained dual approximation loss is inconsistent with the true ranking error. Subsequently, the paper proposes a Margin-Shifted surrogate based on confidence intervals and a structure-aware SA-DPO objective.
Mind the State: Towards Unified, Context-Aware EEG-to-fMRI Synthesis
Yamin Li (Vanderbilt University), Catie Chang (Vanderbilt University)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelAuto EncoderContrastive LearningMultimodalityTime SeriesBiomedical DataMagnetic Resonance ImagingElectrocardiogram
🎯 What it does: Propose a unified, context-aware EEG-to-fMRI generative model called UniEFS, capable of reconstructing whole-brain ROI-level fMRI signals in one go;
Mind Your Entropy: From Maximum Entropy to Trajectory Entropy-Constrained RL
Guojian Zhan (Tsinghua University), Shengbo Eben Li (Tsinghua University)
Recurrent Neural NetworkTransformerReinforcement LearningTabularTime Series
🎯 What it does: Proposed the Trajectory Entropy Constrained Reinforcement Learning (TECRL) framework and implemented a practical algorithm based on DSAC called DSAC-E, aimed at addressing the non-stationary Q-value estimation and single-step entropy modulation issues in maximum entropy RL.
Mind Your Margin and Boundary: Are Your Distilled Datasets Truly Robust?
Muquan Li (University of Electronic Science and Technology of China), Tao He (University of Electronic Science and Technology of China)
Knowledge DistillationAdversarial AttackContrastive LearningImage
🎯 What it does: Proposes a robust dataset distillation method called C<sup>2</sup>R that combines attack-aware curriculum with contrastive augmentation
Mind-Omni: A Unified Multi-Task Framework for Brain-Vision-Language Modeling via Discrete Diffusion
Yizhuo Lu (Chinese Academy of Sciences), Huiguang He (Chinese Academy of Sciences)
Representation LearningData-Centric LearningTransformerMixture of ExpertsVision Language ModelDiffusion modelAuto EncoderContrastive LearningImageTextBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose Mind-Omni, a unified multi-task framework that utilizes discrete diffusion models to simultaneously perform seven brain-visual-language encoding and decoding tasks.
MIND: Decoupling Model-Induced Label Noise via Latent Manifold Disentanglement
Dayong Ren (Nanjing University)
SegmentationScore-based ModelFlow-based ModelAuto EncoderContrastive LearningImagePoint Cloud
🎯 What it does: Propose the MIND framework, which utilizes a Latent Decoupling Estimator (LDE) to decompose model-induced instance-dependent label noise into identifiable low-rank subspace basis matrices, thereby achieving unsupervised noise removal.
MIND: Multi-rationale INtegrated Discriminative Reasoning Framework for Multi-modal Large Models
Chuang Yu (Shenyang Institute of Automation, Chinese Academy of Sciences), Xiangyu Yue (Chinese University of Hong Kong)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningImageTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a multi-reason integration discriminative reasoning framework based on a multi-modal large language model (MIND), achieving the model's 'understand → re-examine → correct' reasoning ability through three major technologies: automatically constructing multi-reason data, two-stage evolutionary learning, and contrastive alignment.
MindFlow: Mind Supernet Powered Thinking Flows for Research Idea Innovation
Mengdi Liu (Chinese Academy of Sciences), Chenglin Wu (DeepWisdom)
Explainability and InterpretabilityNeural Architecture SearchGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the MindFlow framework, which explicitly represents research ideas as graph-structured thought processes, and achieves controllable, multi-objective research innovation generation through a mental hypernetwork and relative ranking optimization.
MindZero: Learning Online Mental Reasoning With Zero Annotations
Shunchi Zhang (Johns Hopkins University), Tianmin Shu (Johns Hopkins University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelGenerative Adversarial NetworkContrastive LearningTextMultimodalitySequentialChain-of-Thought
🎯 What it does: Train a multimodal language model called MindZero through self-supervised reinforcement learning, enabling online annotation-free theoretical mind reasoning, capable of real-time prediction of human goals and providing proactive assistance.
MineDraft: A Framework for Batch Parallel Speculative Decoding
Zhenwei Tang (Nanyang Technological University), Bryan Kian Hsiang Low (National University of Singapore)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Designed and implemented the MINEDRAFT framework, achieving batch parallel speculative decoding (PSD), which hides the draft latency and improves the throughput and latency of LLM inference by alternating between draft generation and verification between two request batches.
MiniAppBench: Evaluating the Shift from Text to Interactive HTML Responses in LLM-Powered Assistants
Zuhao zhang, Shuai Li (Shanghai Jiao Tong University)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the MINIAPPBENCH benchmark to evaluate the ability of LLMs to generate interactive HTML applications (MiniApp);
Minibatch Optimal Transport and Perplexity Bound Estimation in Discrete Flow Matching
Etrit Haxholli (Metadialog Research), Eli Waxman (Metadialog Research)
GenerationData SynthesisComputational EfficiencyTransformerLarge Language ModelDiffusion modelScore-based ModelFlow-based ModelText
🎯 What it does: Proposes a path length minimization objective based on dynamic optimal transport in discrete flow matching (DFM), derives its Kantorovich form, and provides the Benamou-Brenier formula for the discrete case; meanwhile, it presents two computable upper bounds on perplexity and introduces multimask flow to improve generation quality.
Minibatch selection for Language Models via Partition Matroid Constrained Gradient Matching
Prayas Agrawal (Microsoft Research India), Pratik Jawanpuria (Indian Institute of Technology Bombay)
Domain AdaptationOptimizationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a cross-domain small-batch selection method based on partition mother graph constraints (PARTITIONSEL), aimed at efficiently fine-tuning large language models on multi-domain data.
miniF2F-Dafny: LLM-Guided Mathematical Theorem Proving via Auto-Active Verification
Mantas Baksys, Sean B. Holden (University of Cambridge)
OptimizationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This study proposes MINIF2F-DAFNY, which migrates the miniF2F mathematical proof benchmark to the automated verifier Dafny, and utilizes LLMs to assist in generating proof prompts;
MINIM: Privacy-Aware Minimal View for Agents via Trusted Local Sanitization
Hexuan Yu (Virginia Tech), Wenjing Lou (Virginia Tech)
Federated LearningSafty and PrivacyGraph Neural NetworkTransformerPrompt EngineeringContrastive LearningTextGraph
🎯 What it does: Proposes MINIM, a trustworthy local agent that performs structured UI observation on the client side. It generates a ternary publication strategy by predicting sensitivity and task necessity for each UI element, thereby minimizing the observations sent to the remote inference server.
MiniMax Learning of Interpretable Factored Stochastic Policies from Conjoint Data, with Uncertainty Quantification
Connor Thomas Jerzak (University of Texas at Austin), Rishi Hazra (Harvard)
OptimizationFederated LearningExplainability and InterpretabilityReinforcement Learning from Human FeedbackReinforcement LearningContrastive LearningTabularReview/Survey Paper
🎯 What it does: Learn interpretable factorized stochastic policies on the offline federated (conjoint) data to maximize the probability of candidate selection, and achieve minimax strategies in two-stage elections with adversarial interactions;
Minimax Optimal Strategy for Delayed Observations in Online Reinforcement Learning
Harin Lee (University of Washington), Kevin Jamieson (University of Washington)
Reinforcement Learning
🎯 What it does: This paper studies the problem of state observation delay in online reinforcement learning, and proposes a learning algorithm based on an augmented state space combined with the UCBVI method.
Minimax-Optimal Policy Regret in Partially Observable Markov Games
Raman Arora (Johns Hopkins University)
Reinforcement Learning
🎯 What it does: Studied sequential decision-making in partially observable environments with strategically adaptive opponents, establishing a framework for partially observable Markov games (POMGs).
Minimizing Mismatch Risk: A Prototype-Based Routing Framework for Zero-shot LLM-generated Text Detection
Ke Sun (Westlake University), Yue Zhang (Westlake University)
ClassificationAnomaly DetectionFederated LearningExplainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsContrastive LearningTextBenchmark
🎯 What it does: This paper addresses the mismatch problem between the source model and the proxy model in zero-shot LLM text detection, proposing a prototype-based routing framework called DetectRouter, which can dynamically select the most suitable detector for each input.
Minimizing Upper Confidence Bounds: A Data-Driven Framework for Stochastic Programming
Shixin Liu (University of Michigan-Dearborn), Jian Hu (University of Michigan-Dearborn)
OptimizationTabularBenchmark
🎯 What it does: Proposed a data-driven framework based on the Average Percentile Upper Bound (APUB), which utilizes bootstrap and L-shaped branch-and-bound to solve two-stage stochastic programming, thereby reducing the estimation error caused by the 'curse of the decision-maker' when the sample size is insufficient.
Minimum Distance Summaries for Robust Neural Posterior Estimation
Sherman Khoo (University of Bristol), Mark Beaumont (University of Bristol)
OptimizationFederated LearningExplainability and InterpretabilityRepresentation LearningData-Centric LearningRobotic IntelligenceDiffusion modelScore-based ModelFlow-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningTabularTime SeriesBiomedical DataBenchmark
🎯 What it does: Propose a post-hoc, test-time summary adaptation method (Minimum Distance Summaries, MDS), which enhances robustness by minimizing the distance between the observed data and the conditional predictive distribution given summaries, while keeping the pre-trained neural posterior estimator (NPE) unchanged.
Mining Tensor/Neuron-Level Sparsity to Maximize Mixture-of-Experts Potential in Post-Training and Inference
Weilin Cai (Hong Kong University of Science and Technology), Jiayi Huang (Hong Kong University of Science and Technology)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Utilize tensor-level and neuron-level sparsity to enhance MoE model performance through complete expert partitioning, threshold-gated expert dropping, and dual-threshold + load-aware strategies during pre-training and inference stages.
Mining Useful General Data for Low-Resource Domain Adaptation
Pingjie Wang (Shanghai Jiao Tong University), Yu Wang (Shanghai Jiao Tong University)
Domain AdaptationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextTabularFinance RelatedChain-of-Thought
🎯 What it does: Propose NTK-Selector, which improves the adaptation effect in low-resource domains by leveraging general domain data;
MINT: Minimal Information Neuro-Symbolic Tree for Objective-Driven Knowledge-Gap Reasoning and Active Elicitation
Zeyu Fang (George Washington University), Tian Lan (George Washington University)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose Minimal Information Neuro-Symbolic Tree (MINT), which uses symbolic tree reasoning to identify knowledge gaps and combines neural planning strategies to actively initiate binary queries to humans, thereby improving open-domain joint planning performance.
MIRA: A Score for Conditional Distribution Accuracy and Model Comparison
Sammy Nasser Sharief (Mila - Québec AI Institute), Laurence Perreault-Levasseur (Mila - Québec AI Institute)
GenerationData SynthesisOptimizationScore-based ModelImagePoint CloudGraphTabularTime SeriesBenchmark
🎯 What it does: Proposes a sample-based benchmark score called Mira, used to evaluate the similarity between candidate conditional distributions (or posterior distributions) and the true distribution, relying only on joint samples from the true data generation process.
MIRO: MultI-Reward cOnditioned pretraining improves T2I quality and efficiency
Nicolas Dufour (ENPC), David Picard (ENPC)
GenerationData SynthesisReinforcement Learning from Human FeedbackTransformerPrompt EngineeringDiffusion modelFlow-based ModelImageTextMultimodality
🎯 What it does: Propose the MIRO method, which injects multiple rewards (aesthetics, user preferences, text-image alignment, visual reasoning, etc.) as conditional vectors into the generative model during the pre-training phase, achieving multi-reward alignment.
Mirror Descent Actor Critic via Bounded Advantage Learning
Ryo Iwaki (IBM Research)
Reinforcement LearningScore-based ModelContrastive LearningTabularTime Series
🎯 What it does: Proposed the Mirror Descent Actor-Critic (MDAC) algorithm, and improved the performance in continuous action domains by bounding the log probability term in the critic loss.
Mirror Descent Under Generalized Smoothness
Dingzhi Yu (Nanjing University), Lijun Zhang (Nanjing University)
OptimizationConvolutional Neural NetworkTransformerLarge Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes a new concept of 'ℓ*‑smoothness' and proves the convergence of mirror descent (MD) and its variants, such as accelerated, optimistic, and mirror prox, in convex, stochastic, and non-convex composite problems. The obtained convergence rates are the same as the classical L-smooth case.
Mirror Mean-Field Langevin Dynamics
Anming Gu (University of Texas at Austin), Juno Kim (UC Berkeley)
OptimizationFederated LearningScore-based ModelContrastive LearningGraphTabularStochastic Differential Equation
🎯 What it does: Propose Mirror Mean-Field Langevin Dynamics (MMFLD) for minimizing entropy-regularized distributed objectives under convex domain constraints.
MIST: Moment-Aligned Invariant Stability Transform for Robust Flow Matching
Liang Peng (Alibaba Group), Lin Qu (Alibaba Group)
GenerationTransformerDiffusion modelFlow-based ModelImageVideoTextBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: To address the instability of Classifier-Free Guidance (CFG) in flow matching models, the MIST framework is proposed, which applies two training-free steps (Invariant Alignment and Stability Thresholding) during the sampling phase to impose geometric constraints and dynamic smoothing on the guidance vector.
Mitigating Bias in Locally Constrained Decoding via Tractable Proposals
Meihua Dang (Stanford University), Stefano Ermon (Stanford University)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelReinforcement LearningDiffusion modelScore-based ModelAuto EncoderContrastive LearningTextSequentialStochastic Differential Equation
🎯 What it does: In constraint generation tasks for large language models, we propose Global Constraint Decoding (GCD) and Probabilistic Global Constraint Decoding (P-GCD), and use them as the proposal and potential distributions in sequential Monte Carlo (SMC) sampling, thereby reducing the bias of local constraint decoding (LCD) and ensuring that constraints are satisfied within the given token limit.
Mitigating Conversational Inertia in Multi-Turn Agents
Yang Wan (Zhejiang University), Linchao Zhu (Zhejiang University)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes and verifies the 'dialogue inertia' problem in multi-turn dialogue agents, and mitigates this issue through two techniques: Context Preference Learning and Clip Context for managing context.
Mitigating Error Accumulation in Continuous Navigation via Memory-Augmented Kalman Filtering
Yin Tang (Central South University), Deyu Zhang (Central South University)
Autonomous DrivingOptimizationComputational EfficiencyReinforcement Learning from Human FeedbackRecurrent Neural NetworkTransformerVision-Language-Action ModelContrastive LearningGaussian SplattingSimultaneous Localization and MappingMultimodalityTime SeriesSequentialRetrieval-Augmented Generation
🎯 What it does: Propose the NeuroKalman framework, which utilizes memory-enhanced Kalman filtering to correct state drift in continuous navigation of drones, divided into two stages: prediction prior and update posterior;
Mitigating Error Propagation in Low-Rank Approximation of Large Models via Distribution-Aware Whitening
Yao Zhu (Zhejiang University), Yunjian Zhang (University of Chinese Academy of Sciences)
CompressionComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a distribution-aware whitening framework, which performs low-rank approximation after dynamically whitening the input features at each layer, achieving model compression and parameter-efficient fine-tuning.
Mitigating Gradient Pathology in PINNs through Aligned Constraint
Yichen Luo (KTH Royal Institute of Technology), Zhibo Pang (Peking University)
OptimizationBenchmarkPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose CAML, a loss function that alleviates PINN gradient pathologies by introducing alignment constraints and residual delays;
Mitigating Hallucinations in Large Vision-Language Models via Causal Route Gating
Zhe Cheng (Southwestern University of Finance and Economics), Dehuan Shen (National University of Singapore)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed a gating mechanism that applies causal interventions to visual and textual routing during the inference phase to reduce hallucinations in large-scale vision-language models.
Mitigating Label Shift in Tabular In-Context Learning via Test-Time Posterior Adjustment
Seunghan Lee (LG AI Research)
ClassificationDomain AdaptationData-Centric LearningTransformerPrompt EngineeringContrastive LearningTabularBenchmark
🎯 What it does: This paper proposes a post-processing technique for Tabular Foundation Models during inference, aiming to alleviate majority class bias caused by label distribution drift;
Mitigating Manifold Departure: Uncertainty-Aware Subspace Rectification for Trustworthy MLLM Decoding
Yingxuan Zhuang (Zhejiang University), Jintao Chen (Zhejiang University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Proposes a training-agnostic decoding method called MGAP, which suppresses hallucinations in multimodal large language models using subspace projection and adaptive gating.
Mitigating Mask Prior Drift and Positional Attention Collapse in Large Diffusion Vision-Language Models
Sujung Hong (Yonsei University), Seong Jae Hwang (Yonsei University)
RecognitionImage TranslationRestorationSegmentationGenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: This paper addresses the issues of repetitive generation and visual mismatch in large-scale diffusion vision-language models during long-text generation and visual localization, proposing two training-free intervention methods during inference to alleviate these defects.
Mitigating Noise-Induced Layout Priors for Object Counting in Diffusion Models
Xiaoling Gu, Mohan Kankanhalli
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelImageTextMultimodality
🎯 What it does: Propose a training-agnostic framework that utilizes initial noise adjustment and attention consistency strategies to precisely control object counting in text-to-image diffusion models;
Mitigating Perceptual Judgment Bias in Multimodal LLM-as-a-Judge via Perceptual Perturbation and Reward Modeling
Seojeong Park (KAIST), Hyunjung Shim (KAIST)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper aims to eliminate perceptual judgment bias in multi-modal LLM discriminators under scenarios of visual and textual conflicts by constructing a Perceptual Perturbation Judgment Dataset (PPJD) and performing reinforcement learning on the discriminators within this dataset.
Mitigating Plasticity Loss through Architectural Design in Continual Learning
Niklas Koeppe (KAIST), Sang Wan Lee (KAIST)
OptimizationReinforcement LearningContrastive LearningImageVideo
🎯 What it does: Designed and evaluated a network layer called InterpLayer to alleviate plasticity loss in continual reinforcement learning.
Mitigating Premature Exploitation in Particle-based Monte Carlo for Inference-Time Scaling
Giorgio Giannone (AI Innovation, Red Hat), Akash Srivastava (Core AI, IBM)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose Entropic Particle Filtering (ePF), which alleviates the particle depletion problem caused by early overutilization in particle filtering through two techniques: entropic annealing and look-ahead modulation, and enhances the inference performance of large language models during scaling.
Mitigating Reward Hacking in LLM-based Recommendation: A Preference Optimization Approach
Heyu Chen (University of Science and Technology of China), Jiancan Wu (University of Science and Technology of China)
Recommendation SystemOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose the SIRIUS method, which mitigates the reward hacking problem in LLM recommendation systems by introducing pseudo-negative samples.
Mitigating Reward Hacking in RLHF via Bayesian Non-negative Reward Modeling
Zhibin Duan (Xi'an Jiaotong University), Dandan Guo (Jilin University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose the Bayesian Non-Negative Reward Model (BNRM), which learns the reward function in RLHF through a sparse non-negative factorization Bayesian framework, reducing reward over-optimization;
Mitigating Staleness in Asynchronous Pipeline Parallelism via Basis Rotation
Hyunji Jung (Pohang University of Science and Technology), Namhoon Lee (Pohang University of Science and Technology)
OptimizationFederated LearningComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: The study investigates the gradient delay problem in asynchronous pipeline parallel training and proposes a basis vector rotation method to mitigate oscillations caused by gradient lag.
Mitigating Surgical Data Imbalance with Dual-Prediction Video Diffusion Model
Danush Kumar Venkatesh (NCT/UCC Dresden), Omid Mohareri (Intuitive Surgical)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelContrastive LearningOptical FlowVideoBiomedical Data
🎯 What it does: Propose SurgiFlowVid, which utilizes a dual-prediction video diffusion model to generate rare category surgical videos, addressing the issue of data imbalance.
Mitigating the Contractivity Trap in Diffusion ODEs via Stein Stabilization
Shigui Li (South China University of Technology), Delu zeng
GenerationDiffusion modelScore-based ModelImageOrdinary Differential Equation
🎯 What it does: Propose a training-free, plug-and-play ODE inference stabilization framework called SteinDiff, based on Stein correction, to address convergence traps in large-step inference.
Mitigating the Modality Gap in Vision–Language Models with Fractal Spectral Geometry
Zihan Zhou (Auburn University), Patrick Emami (National Laboratory of the Rockies)
ClassificationRetrievalRepresentation LearningVision Language ModelDiffusion modelScore-based ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose a multi-scale alignment method based on fractal spectral analysis (FSAlign), which reduces the modality gap in vision-text models by constructing a shared fractal geometric structure and matching local heat diffusion spectra.
Mitigating the Safety–Utility Trade-off in LLM Alignment via Adaptive Safe Context Learning
Yanbo Wang (University of Chinese Academy of Sciences), Ran He (University of Chinese Academy of Sciences)
Safty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose Adaptive Safe Context Learning (ASCL), which decouples model inference from safety rules by dynamically retrieving safety rules, thereby reducing the safety-utility trade-off in large inference models.
Mitigating Translationese Bias in Multilingual LLM-as-a-Judge via Disentangled Information Bottleneck
Hongbin Zhang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextMultimodalityBenchmark
🎯 What it does: This paper studies and mitigates the 'translationese bias' in multilingual LLM judges, proposing the DIBJUDGE framework.
Mitigating Visual Hallucinations via Semantic Curriculum Preference Optimization in MLLMs
Yuanshuai li, Yaochu Jin (Westlake University)
OptimizationExplainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose a new multimodal large language model alignment framework called SCPO to reduce visual hallucinations.
MiVE: Multiscale Vision-language features for reference-guided video Editing
Tong Wang (Meitu Inc.), Ting Liu (Meitu Inc.)
Image TranslationRestorationGenerationTransformerVision Language ModelDiffusion modelFlow-based ModelAuto EncoderImageVideoText
🎯 What it does: Propose a reference-guided video editing framework called MiVE based on multi-scale vision-language features, achieving synchronized propagation of attributes from the source video to the reference image while preserving the integrity of the original motion and unedited regions.
MixFP4: Enhancing NVFP4 with Adaptive FP4/INT4 Block Representations
Jiaxiang Zou (Hong Kong University of Science and Technology), Xinyu Chen (Hong Kong University of Science and Technology)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningText
🎯 What it does: Propose MixFP4, a block-level hybrid microformat quantization scheme implemented on the basis of NVFP4, which can dynamically select E2M1 (FP4) or E1M2 (INT-style) in each 4-bit block to better match local tensor statistics;
Mixing Configurations for Downstream Prediction
Juntang Wang (Duke Kunshan University), Shixin Xu (Duke Kunshan University)
ClassificationRepresentation LearningData-Centric LearningTransformerMixture of ExpertsAuto EncoderContrastive LearningImageTextTabularBiomedical Data
🎯 What it does: Propose the MixConfig module, which can extract limited stable clustering configurations in any embedding space, and use an energy-aware selector to learn sample-level weighted mixing for downstream prediction.
Mixing Expertise with Confidence: A Mixture of Experts Framework for Robust Multi-Modal Continual Learning
Md Abdullah Al Forhad (University of North Texas), Weishi Shi (University of North Texas)
ClassificationRecognitionDomain AdaptationComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerMixture of ExpertsVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a hybrid expert framework that does not require shared parameters or task ID predictors, leveraging the open-set learning capability of multi-modal pre-trained models to achieve dynamic communication and knowledge sharing among experts;
MixReasoning: Switching Modes to Think
Haiquan Lu (National University Of Singapore), Xinchao Wang (National University Of Singapore)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the MixReasoning framework, allowing large reasoning models to adaptively switch between detailed reasoning and concise reasoning modes within a single response, dynamically adjusting the depth of reasoning.
Mixture of Concept Bottleneck Experts
Francesco De Santis (Polytechnic of Torino), Pietro Barbiero (IBM Research)
Explainability and InterpretabilityRepresentation LearningMixture of ExpertsAuto EncoderImageTextTabular
🎯 What it does: Propose the Mixture of Concept Bottleneck Experts (M-CBE) framework, extending traditional Concept Bottleneck Models (CBM) into a configurable multi-expert model, and provide two instances: Linear M-CBE and Symbolic M-CBE.