ICLR 2025 Papers — Page 21
International Conference on Learning Representations · 3704 papers
MAP: Multi-Human-Value Alignment Palette
Xinran Wang (University of Minnesota), Ali Anwar (University of Minnesota)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A Multi-Human-Value Alignment Palette (MAP) method is proposed, allowing for a one-time alignment of generative AI according to user-specified multi-dimensional value objectives while maintaining the model's original distribution.
MAPS: Advancing Multi-Modal Reasoning in Expert-Level Physical Science
Erle Zhu (Tsinghua University), Hongning Wang (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodalityPhysics RelatedChain-of-Thought
🎯 What it does: The MAPS framework is proposed, which combines multimodal large language models with physical perception models and simulators, using chain simulation to improve the reasoning accuracy of circuit analysis problems.
MaRS: A Fast Sampler for Mean Reverting Diffusion based on ODE and SDE Solvers
Ao Li (University of Chinese Academy of Sciences), Minfeng Xu (Alibaba Group)
RestorationComputational EfficiencyDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A fast sampler MRSampler based on semi-analytical solutions is proposed to accelerate the sampling process of Mean Reverting Diffusion.
MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model
Junjie Li (Microsoft Research Asia), Jiang Bian (Microsoft Research Asia)
GenerationTransformerLarge Language ModelReinforcement LearningTime SeriesSequentialFinance Related
🎯 What it does: This paper proposes a large-scale order-level financial market simulation foundational model (LMM) and a financial market simulation engine (MarS) based on LMM, capable of generating realistic order flows across three dimensions: high resolution, controllability, and interactivity. It supports various downstream financial tasks such as forecasting, risk detection, impact analysis, and reinforcement learning.
Mask in the Mirror: Implicit Sparsification
Tom Jacobs (CISPA Helmholtz Center for Information Security), Rebekka Burkholz (CISPA Helmholtz Center for Information Security)
Image
🎯 What it does: This study investigates implicit regularization in continuous sparsification and proposes the PILoT algorithm for dynamically controlling the Bregman potential to achieve a transition from L2 to L1.
Mask-DPO: Generalizable Fine-grained Factuality Alignment of LLMs
Yuzhe Gu (Shanghai Jiao Tong University), Kai Chen (Shanghai Jiao Tong University)
OptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes and implements Mask-DPO, a direct preference optimization method that combines sentence-level factual masking to reduce hallucinations in large language models.
Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling
Kaiwen Zheng (Tsinghua University), Qinsheng Zhang (NVIDIA)
GenerationData SynthesisDiffusion modelText
🎯 What it does: Analyze and redefine the theory and training objectives of Masked Diffusion Models (MDMs), proving their equivalence to masked models; propose the First Hit Sampler (FHS) to accelerate sampling and reveal numerical errors caused by low-precision Gumbel sampling.
Masked Temporal Interpolation Diffusion for Procedure Planning in Instructional Videos
Yufan Zhou (Harbin Institute of Technology), Weigang Zhang (Harbin Institute of Technology)
GenerationData SynthesisRobotic IntelligenceTransformerDiffusion modelVideo
🎯 What it does: A program planning framework based on diffusion models, MTID, is proposed, which utilizes intermediate latent visual features for temporal interpolation and restricts the action space through a masking mechanism, ultimately generating a coherent action sequence that aligns with task objectives from initial and final visual observations.
MaskGCT: Zero-Shot Text-to-Speech with Masked Generative Codec Transformer
Yuancheng Wang (Chinese University of Hong Kong), Zhizheng Wu (Chinese University of Hong Kong)
GenerationData SynthesisTransformerAudio
🎯 What it does: A completely non-autoregressive zero-shot text-to-speech (TTS) system called MaskGCT is constructed, which uses a masked generative transformer to first predict semantic codes and then acoustic codes, achieving end-to-end speech synthesis.
MAST: model-agnostic sparsified training
Yury Demidovich (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationConvolutional Neural NetworkImageTabular
🎯 What it does: A model-agnostic sparse training (MAST) framework is proposed, which combines pre-trained models with random sketches to sparsify the model and gradients, forming a new optimization objective.
Mastering Task Arithmetic: $\tau$Jp as a Key Indicator for Weight Disentanglement
Kotaro Yoshida (Institute of Science Tokyo), Hiroki Naganuma (Mila)
ClassificationOptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: A task arithmetic regularization method based on τ–Jacobian product (τ Jp) is proposed and validated, significantly reducing task interference and enhancing model editing effects;
Matcha: Mitigating Graph Structure Shifts with Test-Time Adaptation
Wenxuan Bao (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
Domain AdaptationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: The study investigates Test-Time Adaptation (TTA) of graph neural networks in environments with structural shifts and proposes a new framework called Matcha.
Matérn Kernels for Tunable Implicit Surface Reconstruction
Maximilian Weiherer (Friedrich Alexander Universitat Erlangen Nurnberg), Bernhard Egger (Friedrich Alexander Universitat Erlangen Nurnberg)
Point CloudMeshBenchmark
🎯 What it does: This paper proposes the use of the Matérn kernel family for adjustable implicit surface reconstruction and compares it with the traditional first-order arc-cosine kernel.
MatExpert: Decomposing Materials Discovery By Mimicking Human Experts
Qianggang Ding (Universite de Montreal), Bang Liu (Universite de Montreal)
GenerationRetrievalOptimizationExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningTextChain-of-Thought
🎯 What it does: The MatExpert framework is proposed, breaking down material discovery into three stages: retrieval, transformation, and generation, using LLM and contrastive learning to generate solid materials that meet user attributes.
MathCoder2: Better Math Reasoning from Continued Pretraining on Model-translated Mathematical Code
Zimu Lu (Multimedia Laboratory), Hongsheng Li (Multimedia Laboratory)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: The MathCode-Pile dataset was constructed, and continuous pre-training was conducted by generating mathematical code with reasoning steps through model translation, significantly enhancing the mathematical reasoning ability of LLMs.
MathGAP: Out-of-Distribution Evaluation on Problems with Arbitrarily Complex Proofs
Andreas Opedal (ETH Zurich), Mrinmaya Sachan (ETH Zurich)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: The MathGAP framework is proposed, utilizing an automatic generation method for controllable proof tree structures to evaluate the OOD generalization ability of LLMs on arbitrarily complex arithmetic proofs.
Matrix Product Sketching via Coordinated Sampling
Majid Daliri (New York University), Christopher Musco (New York University)
CompressionOptimizationTransformerTextTabular
🎯 What it does: A matrix multiplication approximation sketch method is proposed in a distributed environment through Priority Sampling, which supports independent computation of sketches for A and B, significantly compressing the storage space of sparse matrices while maintaining a Frobenius error of ϵ‖A‖_F‖B‖_F.
Matryoshka Multimodal Models
Mu Cai (University of Wisconsin Madison), Yong Jae Lee (Microsoft Research)
Computational EfficiencyTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodality
🎯 What it does: Developed the Matryoshka multimodal model, which uses a visual encoder to generate multi-layer nested visual tokens, allowing for dynamic selection of tokens of different granularities during inference based on needs;
MatryoshkaKV: Adaptive KV Compression via Trainable Orthogonal Projection
Bokai Lin (Shanghai Jiao Tong University), Zhijie Deng (Shanghai Jiao Tong University)
CompressionKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper compresses the feature dimensions of the KV cache of large language models by training a learnable orthogonal projection matrix, utilizing PCA initialization followed by knowledge distillation and a Matryoshka training strategy, ultimately achieving an adaptive heterogeneous compression rate that significantly reduces KV cache usage.
MAVIS: Mathematical Visual Instruction Tuning with an Automatic Data Engine
Renrui Zhang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
Data SynthesisOptimizationTransformerLarge Language ModelContrastive LearningImageTextMultimodalityChain-of-Thought
🎯 What it does: This paper presents an automated data engine and a four-stage training pipeline named MAVIS, specifically designed to enhance the visual mathematical reasoning capabilities of multimodal large language models (MLLMs);
MaxCutPool: differentiable feature-aware Maxcut for pooling in graph neural networks
Carlo Abate (Alma Mater Studiorum University of Bologna), Filippo Maria Bianchi (UiT Arctic University of Norway)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes a differentiable, feature-based MaxCut pooling layer called MaxCutPool, aimed at achieving sparse and trainable graph pooling in graph neural networks.
Maximizing the Potential of Synthetic Data: Insights from Random Matrix Theory
Aymane El Firdoussi (Technology Innovation Institute), Hakim Hacid (Technology Innovation Institute)
ClassificationData SynthesisText
🎯 What it does: This study investigates the theoretical and experimental impact of using validation mechanisms for synthetic data on model training in high-dimensional environments, conducting a quantitative analysis of the performance of binary classifiers that mix real and synthetic data based on random matrix theory.
MaxInfoRL: Boosting exploration in reinforcement learning through information gain maximization
Bhavya Sukhija (ETH Zurich), Carmelo Sferrazza (UC Berkeley)
Reinforcement LearningSequential
🎯 What it does: A general offline model-agnostic reinforcement learning framework named MAXINFORL is proposed, which can adaptively incorporate information gain as intrinsic rewards into existing algorithms (such as SAC, REDQ, DrQ, DrQv2), thereby achieving a balance between guided exploration and reward maximization.
McEval: Massively Multilingual Code Evaluation
Linzheng Chai (Beihang University), Zhoujun Li (Beihang University)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: The first multilingual code evaluation benchmark MCEVAL covering 40 programming languages is proposed, which includes three main tasks: code generation, interpretation, and completion, and a corresponding multilingual instruction dataset MCEVAL-INSTRUCT is constructed.
MCNC: Manifold-Constrained Reparameterization for Neural Compression
Chayne Thrash (Vanderbilt University), Soheil Kolouri (Vanderbilt University)
CompressionTransformerLarge Language ModelImage
🎯 What it does: A model compression method through low-dimensional nonlinear manifold reparameterization is proposed—MCNC, aimed at significantly reducing the number of trainable parameters in large neural networks.
MDSGen: Fast and Efficient Masked Diffusion Temporal-Aware Transformers for Open-Domain Sound Generation
Trung X. Pham (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelVideoAudio
🎯 What it does: A video-guided audio generation framework named MDSGen is proposed, utilizing a transformer architecture to achieve an efficient diffusion model, addressing the bottlenecks of traditional U-Net models in terms of parameters, memory, and inference speed.
Measuring and Enhancing Trustworthiness of LLMs in RAG through Grounded Attributions and Learning to Refuse
Maojia Song (Singapore University of Technology and Design), Soujanya Poria (Singapore University of Technology and Design)
GenerationRetrievalOptimizationTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a comprehensive metric, TRUST-SCORE, to measure the credibility of large language models (LLMs) in retrieval-augmented generation (RAG) systems. Based on this, an adversarial dataset is constructed, and direct preference optimization (DPO) is used to achieve TRUST-ALIGN for aligning LLMs, significantly improving the model's citation accuracy, rejection capability, and overall credibility.
Measuring And Improving Engagement of Text-to-Image Generation Models
Varun Khurana (Adobe), Balaji Krishnamurthy (Adobe)
GenerationRecommendation SystemOptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: This paper presents a study on text-to-image generation aimed at user engagement, constructing the large-scale real marketing scenario dataset EngagingImageNet, training the EngageNet model for predicting image engagement, and exploring methods to enhance generated image engagement through prompt optimization, fine-tuning, and reinforcement learning, ultimately launching the automated evaluation platform Engagement Arena.
Measuring And Improving Persuasiveness Of Large Language Models
Somesh Kumar Singh (Adobe Media and Data Science Research), Balaji Krishnamurthy (Adobe Media and Data Science Research)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed two frameworks, PersuasionBench and PersuasionArena, for the automatic evaluation of the persuasive power of large language models, and introduced the Transsuasion task; constructed a dataset of approximately 1.57 million transsuasion pairs.
Measuring memorization in RLHF for code completion
Jamie Hayes (Google DeepMind), Aneesh Pappu (Google DeepMind)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper systematically evaluates the risks of training data memorization in code completion tasks between RLHF and direct preference learning (IPO), analyzing the memory propagation mechanisms of the three stages of RLHF.
Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
Michael Aerni (ETH Zurich), Florian Tramèr (ETH Zurich)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Evaluate the non-adversarial training data reproduction of large language models under natural prompts and quantify the overlap between generated text and internet content.
Mechanism and Emergence of Stacked Attention Heads in Multi-Layer Transformers
Tiberiu Mușat (ETH Zurich)
RetrievalTransformerLarge Language ModelText
🎯 What it does: This study investigates a class of retrieval problems (including conditional retrieval) that require a specific number of layers to solve. Through the minimization training of multi-layer Transformers and reverse engineering of attention maps, it reveals the implementation mechanism of retrieval problems—a series of stacked attention heads.
Mechanistic Permutability: Match Features Across Layers
Nikita Balagansky (Moscow Institute of Physics and Technologies), Daniil Gavrilov (T-Tech)
Large Language ModelAuto EncoderText
🎯 What it does: Design and implement the SAE Match method to achieve data-independent feature alignment between different layers of the Sparse Autoencoder (SAE), thereby studying the evolution and persistence of internal features across layers.
MediConfusion: Can you trust your AI radiologist? Probing the reliability of multimodal medical foundation models
Mohammad Shahab Sepehri (University of Southern California), Mahdi Soltanolkotabi (University of Southern California)
RecognitionData-Centric LearningTransformerLarge Language ModelContrastive LearningImageTextMultimodalityBiomedical DataBenchmark
🎯 What it does: A medical visual question-answering benchmark named MediConfusion has been created to systematically evaluate the reliability of multimodal large language models in medical image reasoning.
Medium-Difficulty Samples Constitute Smoothed Decision Boundary for Knowledge Distillation on Pruned Datasets
Yudong Chen (University of Queensland), Sen Wang (University of Queensland)
Computational EfficiencyKnowledge DistillationImage
🎯 What it does: In knowledge distillation, static pruning of the training set is performed to retain only moderately difficult samples, and the logits of the retained samples are reshaped, thereby achieving training acceleration while maintaining or improving model accuracy.
MedTrinity-25M: A Large-scale Multimodal Dataset with Multigranular Annotations for Medicine
Yunfei Xie (Huazhong University of Science and Technology), Yuyin Zhou (UC Santa Cruz)
Data SynthesisRetrievalTransformerLarge Language ModelImageMultimodalityBiomedical DataRetrieval-Augmented Generation
🎯 What it does: A large-scale multimodal medical dataset called MedTrinity-25M is proposed, containing 25 million images and ROI-description triplets, covering 10 imaging modalities and over 65 diagnoses, supporting multi-granularity annotations.
MEGA-Bench: Scaling Multimodal Evaluation to over 500 Real-World Tasks
Jiacheng Chen (Simon Fraser University), Wenhu Chen (University of Waterloo)
TransformerLarge Language ModelVision Language ModelImageVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Designed the MEGA-BENCH benchmark, covering over 500 real multimodal tasks, supporting various input/output formats, and providing over 40 custom evaluation metrics;
Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis
Jinbin Bai (National University of Singapore), Shuicheng YAN
GenerationData SynthesisTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Proposed the Meissonic model, improving Mask Image Modeling (MIM) to achieve 1024×1024 high-resolution text-to-image generation and enabling zero-shot image editing.
MELODI: Exploring Memory Compression for Long Contexts
Yinpeng Chen (Google DeepMind), Jesper Sparre Andersen (Google DeepMind)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A hierarchical compressed memory architecture named MELODI is proposed, which utilizes short-term multi-layer recursive compression and single-layer long-term compression to efficiently process long texts with only a short context window.
Memory Efficient Transformer Adapter for Dense Predictions
Dong Zhang (Hong Kong University of Science and Technology), Kwang-Ting Cheng (Hong Kong University of Science and Technology)
Object DetectionSegmentationTransformerImage
🎯 What it does: A memory-efficient Transformer adapter (META) is proposed, which significantly reduces the memory access overhead and inference time of ViT in dense prediction tasks through shared layer normalization, cross-shaped self-attention, lightweight convolutional branches, and a cascading mechanism.
Memory Mosaics
Jianyu Zhang (Meta), Leon Bottou (Meta)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The Memory Mosaics architecture is proposed, utilizing multiple learnable associative memory units to achieve autoregressive language modeling and context learning similar to transformers.
Merging LoRAs like Playing LEGO: Pushing the Modularity of LoRA to Extremes Through Rank-Wise Clustering
Ziyu Zhao (Shanghai Innovation Institute), Fei Wu (Zhejiang University)
TransformerLarge Language ModelText
🎯 What it does: Decomposing, clustering, and aggregating multiple LoRA adapters by the minimum semantic unit (MSU) to form a scalable, training-free fused LoRA.
MeshAnything: Artist-Created Mesh Generation with Autoregressive Transformers
Yiwen Chen (Nanyang Technological University), Chi Zhang (Westlake University)
GenerationData SynthesisTransformerAuto EncoderPoint CloudMesh
🎯 What it does: The MeshAnything model is proposed to convert any 3D representation asset into a mesh created by artists, achieving autoregressive mesh generation conditioned on shape.
MeshMask: Physics-Based Simulations with Masked Graph Neural Networks
Paul Garnier (Mines Paris - PSL University), Elie Hachem (Mines Paris - PSL University)
Graph Neural NetworkAuto EncoderMeshGraphPhysics Related
🎯 What it does: A mask-based pre-training technique is proposed for the application of graph neural networks in computational fluid dynamics (CFD) simulations, along with the construction of a corresponding autoencoder architecture.
Meta Flow Matching: Integrating Vector Fields on the Wasserstein Manifold
Lazar Atanackovic (University of Toronto), Kirill Neklyudov (Mila - Quebec AI Institute)
Meta LearningDrug DiscoveryGraph Neural NetworkFlow-based ModelBiomedical Data
🎯 What it does: This paper proposes Meta Flow Matching (MFM), a method for integrating vector fields on the Wasserstein manifold to predict group dynamics under different initial distributions.
Meta-Continual Learning of Neural Fields
Seungyoon Woo (Seoul National University), Gunhee Kim (Seoul National University)
GenerationMeta LearningMixture of ExpertsNeural Radiance FieldImageVideoAudio
🎯 What it does: The Meta-Continual Learning of Neural Fields (MCL-NF) framework is proposed, which combines a modular architecture with optimization-based meta-learning to achieve rapid continual learning of neural fields, and introduces Fisher Information Maximization Loss.
Meta-Dynamical State Space Models for Integrative Neural Data Analysis
Ayesha Vermani (Champalimaud Foundation), Il Memming Park (Champalimaud Foundation)
Meta LearningAuto EncoderTime SeriesSequential
🎯 What it does: A meta-learning framework is proposed to jointly infer and learn the latent dynamics across multiple neural recordings, utilizing low-dimensional dynamic embeddings to capture differences between datasets.
metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models
Alex Kipnis (Human Centered AI Helmholtz Munich), Eric Schulz (Human Centered AI Helmholtz Munich)
Large Language ModelTextBenchmark
🎯 What it does: A sparse benchmark Metabench was constructed to compress six mainstream LLM evaluation benchmarks and achieve reproducible score reconstruction.
MetaDesigner: Advancing Artistic Typography through AI-Driven, User-Centric, and Multilingual WordArt Synthesis
Jun-Yan He (Alibaba Group), Alexander G Hauptmann
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageText
🎯 What it does: MetaDesigner is proposed, a user-centered, cross-language WordArt synthesis system based on large language models (LLM) and a multi-agent framework, capable of generating diverse text and textures according to user needs;
Metalic: Meta-Learning In-Context with Protein Language Models
Jacob Beck (InstaDeep), Paul Duckworth (InstaDeep)
Meta LearningDrug DiscoveryProtein Structure PredictionTransformerBiomedical Data
🎯 What it does: Utilizing a method that combines meta-learning and contextual learning to enhance the performance of low-sample protein fitness prediction.
MetaMetrics: Calibrating Metrics for Generation Tasks Using Human Preferences
Genta Indra Winata (Capital One), Derry Tanti Wijaya (Boston University)
GenerationReinforcement Learning from Human FeedbackTextMultimodality
🎯 What it does: METAMETRICS is proposed, a supervised calibrated meta-metric that linearly or non-linearly combines various existing evaluation metrics to better align with human preferences, supporting text generation and visual language tasks.
Metamizer: A Versatile Neural Optimizer for Fast and Accurate Physics Simulations
Nils Wandel (University of Bonn), Reinhard Klein (University of Bonn)
OptimizationComputational EfficiencyMeta LearningPhysics Related
🎯 What it does: Metamizer is a generalizable neural optimizer that iteratively solves various linear and nonlinear PDEs by minimizing physical constraint loss without the need for explicit datasets.
MetaOOD: Automatic Selection of OOD Detection Models
Yuehan Qin (University of Southern California), Yue Zhao (University of Southern California)
Anomaly DetectionMeta LearningLarge Language ModelImage
🎯 What it does: MetaOOD is proposed, a meta-learning based unsupervised OOD detection model selection framework that can automatically select the most suitable OOD detector without labels.
MetaUrban: An Embodied AI Simulation Platform for Urban Micromobility
Wayne Wu (University of California), Bolei Zhou (University of California)
Autonomous DrivingRobotic IntelligenceReinforcement LearningVision Language ModelImage
🎯 What it does: This paper proposes MetaUrban, a simulation platform for urban micro-mobility that can generate an infinite variety of complex urban scenes for training and evaluating AI navigation and safety of various mobile machines in urban street environments.
MeteoRA: Multiple-tasks Embedded LoRA for Large Language Models
Jingwei Xu (Nanjing University), Yunpeng Huang (Nanjing University)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes the MeteoRA framework, which utilizes a full-mode MoE to embed multiple task LoRA adapters into a single LLM, achieving unsupervised task awareness and automatic switching.
Methods for Convex $(L_0,L_1)$-Smooth Optimization: Clipping, Acceleration, and Adaptivity
Eduard Gorbunov (Mohammed Bin Zayed University of Artificial Intelligence), Martin Takáč (Mohammed Bin Zayed University of Artificial Intelligence)
Optimization
🎯 What it does: This paper addresses convex (L, L0, 1)-smooth optimization, providing the convergence rates of improved gradient descent and Polyak step size methods, and proposes new acceleration methods and theoretical analysis of adaptive gradient descent.
Methods with Local Steps and Random Reshuffling for Generally Smooth Non-Convex Federated Optimization
Yury Demidovich (King Abdullah University of Science and Technology), Eduard Gorbunov (Mohammed Bin Zayed University of Artificial Intelligence)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes three federated learning methods under the generalized smoothness ((L,L₀,1)-smooth) assumption, namely: Clip-LocalGDJ (a 'jumping' method with local gradient descent), CLERR (a single-round jumping method combining local steps with random reshuffling), and Clipped RR-CLI (a novel method with local steps, random reshuffling, partial participation, and gradient clipping at the end of the meta-period).
MeToken: Uniform Micro-environment Token Boosts Post-Translational Modification Prediction
Cheng Tan (Zhejiang University), Stan Z. Li (Westlake University)
Protein Structure PredictionGraph Neural NetworkSupervised Fine-TuningBiomedical Data
🎯 What it does: A large-scale dataset containing over 1.2M PTM annotations of sequence-structure has been constructed, and the MeToken model has been proposed to predict the types of PTMs in proteins.
Metric-Driven Attributions for Vision Transformers
Chase Walker (University of Florida), Rickard Ewetz (Florida International University)
OptimizationExplainability and InterpretabilityTransformerImage
🎯 What it does: A visual Transformer attribution method MDA is proposed, which is directly driven by evaluation metrics and generates sparse or dense interpretability images through iterative insertion/deletion and magnitude optimization.
MGCFNN: A Neural MultiGrid Solver with Novel Fourier Neural Network for High Wave Number Helmholtz Equations
Yan Xie (Chinese Academy of Sciences), Chen-Song Zhang (Chinese Academy of Sciences)
Convolutional Neural NetworkImageComputed TomographyUltrasoundPhysics Related
🎯 What it does: A multi-grid hierarchical AI solver MGCFNN is proposed for iterative solving of high wave number Helmholtz equations, capable of achieving solutions at arbitrary precision.
MGDA Converges under Generalized Smoothness, Provably
Qi Zhang (Arizona State University), Kaiyi Ji (University at Buffalo)
SegmentationDepth EstimationOptimizationImage
🎯 What it does: This paper studies the convergence of multi-objective gradient descent (MGDA) and its stochastic dual-sampling variant under a more relaxed generalized ℓ-smooth condition, and proposes a warm-start strategy and an efficient MGDA-FA variant.
MGMapNet: Multi-Granularity Representation Learning for End-to-End Vectorized HD Map Construction
Jing Yang (Tongji University), Hanli Wang (Tongji University)
Autonomous DrivingRepresentation LearningTransformerImagePoint Cloud
🎯 What it does: An end-to-end multi-granularity vectorized high-definition map construction framework MGMapNet is proposed, which utilizes instance-level queries and point-level queries to jointly describe road features.
MIA-Bench: Towards Better Instruction Following Evaluation of Multimodal LLMs
Yusu Qian (Apple), Zhe Gan (Apple)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper presents MIA-Bench, a benchmark for multi-modal large language models (MLLMs) designed to evaluate the models' ability to strictly follow complex hierarchical instructions, and based on this, to improve instruction adherence performance through supervised fine-tuning.
MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models
Ziyu Liu (Shanghai Jiao Tong University), Jiaqi Wang (Shanghai Innovation Institute)
Recommendation SystemOptimizationTransformerReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: Proposes MIA-DPO: a method for multi-image preference alignment of large visual language models by expanding single-image data into multi-image inputs and using attention values to filter generated unlabeled rejection samples.
Microcanonical Langevin Ensembles: Advancing the Sampling of Bayesian Neural Networks
Emanuel Sommer (Ludwig Maximilian University of Munich), David Rügamer (Ludwig Maximilian University of Munich)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerTabular
🎯 What it does: A new Bayesian neural network posterior sampling method is proposed—Microcanonical Langevin Ensembles (MILE). By implementing improvements such as deep ensemble initialization, step size adaptation, energy variance scheduling, and adaptive annealing on the MCLMC sampler, efficient and predictable multimodal sampling is achieved.
MIM-Refiner: A Contrastive Learning Boost from Intermediate Pre-Trained Masked Image Modeling Representations
Benedikt Alkin (Johannes Kepler University Linz), Johannes Brandstetter (Johannes Kepler University Linz)
SegmentationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes MIM-Refiner, which enhances the downstream representation quality of a pre-trained Masked Image Modeling (MIM) model by adding multi-head Instance Discrimination (ID) heads to the intermediate layers and utilizing short-term (a few rounds) refinement.
Min-K%++: Improved Baseline for Pre-Training Data Detection from Large Language Models
Jingyang Zhang (Duke University), Hai Li (Duke University)
TransformerLarge Language ModelText
🎯 What it does: A baseline method Min-K%++ is proposed for detecting pre-training data of large language models by determining whether the input is a mode or high-probability point in the model's conditional distribution to assess training samples.
Mind Control through Causal Inference: Predicting Clean Images from Poisoned Data
Mengxuan Hu (University of Virginia), Sheng Li (University of Virginia)
ClassificationRecognitionAdversarial AttackImage
🎯 What it does: The research directly trains models without backdoors from poisoned data that can recover the original labels, proposing an end-to-end anti-backdoor learning method.
MIND over Body: Adaptive Thinking using Dynamic Computation
Mrinal Mathur (Georgia State University), Sergey M. Plis
ClassificationOptimizationComputational EfficiencyTransformerSupervised Fine-TuningReinforcement LearningImageText
🎯 What it does: The MIND model is proposed, which dynamically adjusts the computation depth through a self-reflection network and fixed-point iteration (FPI), allocating parameters and computational resources according to input complexity.
Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models
Yuda Song (Carnegie Mellon University), Udaya Ghai (Amazon)
OptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper systematically studies the mechanism of self-improvement in large language models (LLMs), proposing the generation-verification gap (GV-Gap) as a metric, and conducts rigorous comparisons and experiments across various model families, tasks, and validation methods.
Mind the GAP: Glimpse-based Active Perception improves generalization and sample efficiency of visual reasoning
Oleh Kolner (IBM Research Europe), Angeliki Pantazi (IBM Research Europe)
RecognitionObject DetectionDomain AdaptationTransformerImage
🎯 What it does: A perspective-focused active perception system called GAP is designed to sequentially focus on the most salient regions in images, extracting dual information of 'what' and 'where', and inputting it into a Transformer or Abstractor to complete visual reasoning tasks.
MIND: Math Informed syNthetic Dialogues for Pretraining LLMs
Syeda Nahida Akter (Carnegie Mellon University), Bryan Catanzaro (NVIDIA)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the MIND method, which utilizes a pre-trained LLM to convert raw mathematical text into multi-turn dialogues, and continuously pre-trains a 7B language model on these dialogues to enhance mathematical reasoning and general reasoning capabilities.
MindSearch: Mimicking Human Minds Elicits Deep AI Searcher
Zehui Chen (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
RetrievalTransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Proposes the MindSearch framework, utilizing LLM multi-agent to achieve cognitive decomposition and hierarchical retrieval of complex problems, improving network information retrieval and integration.
MindSimulator: Exploring Brain Concept Localization via Synthetic fMRI
Guangyin Bao (Tongji University), Duoqian Miao (Tongji University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImageMagnetic Resonance Imaging
🎯 What it does: Developed MindSimulator, a generative fMRI encoder based on diffusion models, capable of synthesizing brain activation signals from visual stimuli and used for concept localization.
Mini-batch Coresets for Memory-efficient Language Model Training on Data Mixtures
Dang Nguyen (University of California Los Angeles), Baharan Mirzasoleiman (University of California Los Angeles)
Large Language ModelSupervised Fine-TuningText
🎯 What it does: The CoLM method is proposed, which selects a small batch core set (coreset) that approximates the gradient in large batches to achieve the effects of large batch training while reducing the memory consumption of LLMs.
Mini-Monkey: Alleviating the Semantic Sawtooth Effect for Lightweight MLLMs via Complementary Image Pyramid
Mingxin Huang (South China University of Technology), Xiang Bai (Huazhong University of Science and Technology)
RecognitionSegmentationCompressionTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: This paper proposes the Complementary Image Pyramid (CIP) and Scale Compression Mechanism (SCM), which dynamically constructs a multi-scale image pyramid and compresses visual tokens to address the semantic aliasing effect encountered by cropping-based MLLMs when processing high-resolution images, further developing a lightweight MLLM called Mini-Monkey.
miniCTX: Neural Theorem Proving with (Long-)Contexts
Jiewen Hu (Carnegie Mellon University), Sean Welleck (Carnegie Mellon University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: The miniCTX benchmark is proposed to evaluate whether neural network models can complete theorem proving in practical Lean projects that contain a large amount of new definitions, lemmas, annotations, and other contexts.
Minimal Impact ControlNet: Advancing Multi-ControlNet Integration
Shikun Sun (Tsinghua University), Jia Jia (Tsinghua University)
GenerationOptimizationImage
🎯 What it does: This paper proposes Minimal Impact ControlNet (MIControlNet) to address the issues of control signal conflicts and silent signal suppression in texture generation when combining multiple ControlNets.
Minimal Variance Model Aggregation: A principled, non-intrusive, and versatile integration of black box models
Theo Bourdais, Houman Owhadi (California Institute of Technology)
TabularTime Series
🎯 What it does: The Minimal Empirical Variance Aggregation (MEVA) framework is proposed, which non-invasively aggregates the predictions of various models (such as machine learning and numerical solvers) with weighted aggregation, aiming to minimize the variance of the aggregation error.
Minimalistic Predictions for Online Class Constraint Scheduling
Dorian Guyot (University of Fribourg), Alexandra Anna Lassota
OptimizationReinforcement Learning
🎯 What it does: Research on online class-constrained scheduling problems, presenting a learning-enhanced algorithm.
Minimax Optimal Reinforcement Learning with Quasi-Optimism
Harin Lee (Seoul National University), Min-hwan Oh (Seoul National University)
OptimizationReinforcement LearningTabular
🎯 What it does: A new reinforcement learning algorithm based on 'quasi-optimism' called EQO is proposed, which implements exploration using a reward term that relies solely on visit counts;
Minimax Optimal Two-Stage Algorithm For Moment Estimation Under Covariate Shift
Zhen Zhang (Shanghai University of Finance and Economics), Jiaye Teng (Shanghai University of Finance and Economics)
Domain AdaptationOptimizationTabular
🎯 What it does: An optimal two-stage algorithm is proposed for estimating the moments of an unknown function under covariate shift, along with corresponding lower and upper bounds.
Mining your own secrets: Diffusion Classifier Scores for Continual Personalization of Text-to-Image Diffusion Models
Saurav Jha (University of New South Wales Sydney), Yuki Mitsufuji (Sony Group Corporation)
GenerationData SynthesisDiffusion modelScore-based ModelImage
🎯 What it does: This paper proposes a method for continuous personalization of text-to-image diffusion models under no replay conditions, utilizing diffusion classifier (DC) scores as regularization to mitigate forgetting during the incremental learning process.
MiniPLM: Knowledge Distillation for Pre-training Language Models
Yuxian Gu (Tsinghua University), Minlie Huang (Tsinghua University)
Knowledge DistillationTransformerLarge Language ModelText
🎯 What it does: This paper proposes the MINIPLM framework, which integrates the knowledge of the teacher LM into the pre-training data distribution through differential sampling to train efficient small language models.
MIRACLE 3D: Memory-efficient Integrated Robust Approach for Continual Learning on 3D Point Clouds via Shape Model Construction
Hossein Resani (K. N. Toosi University of Technology), Behrooz Nasihatkon (K. N. Toosi University of Technology)
ClassificationSafty and PrivacyComputational EfficiencyKnowledge DistillationPoint Cloud
🎯 What it does: A continuous learning framework named MIRACLE3D is proposed, which focuses on 3D point cloud classification by only storing the mean shape of each class and a few key deformation patterns. Samples are generated from these shape models for replay, achieving low memory usage and privacy-friendly learning.
MIRAGE: Evaluating and Explaining Inductive Reasoning Process in Language Models
Jiachun Li (University of Chinese Academy of Sciences), Jun Zhao (University of Chinese Academy of Sciences)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: A synthetic dataset named MIRAGE has been constructed to flexibly generate test samples for a comprehensive evaluation of large language models' performance in inductive reasoning (rule induction and deduction).
Misspecified $Q$-Learning with Sparse Linear Function Approximation: Tight Bounds on Approximation Error
Ally Yalei Du (Carnegie Mellon University), Ruosong Wang (Peking University)
Reinforcement Learning
🎯 What it does: A new reinforcement learning algorithm is designed, providing polynomial sample complexity and a suboptimality of O(Hϵ) with theoretical guarantees in the presence of sparse linear function approximation and error ϵ.
Mitigate the Gap: Improving Cross-Modal Alignment in CLIP
Sedigheh Eslami (Hasso Plattner Institute), Gerard de Melo (Hasso Plattner Institute)
ClassificationRetrievalTransformerContrastive LearningImageTextMultimodality
🎯 What it does: The AlignCLIP framework is proposed, which enhances the cross-modal alignment of CLIP by sharing the parameters of the visual and text encoders and separating the semantic regularization of unimodal embeddings.
Mitigating Information Loss in Tree-Based Reinforcement Learning via Direct Optimization
Sascha Marton (University of Mannheim), Heiner Stuckenschmidt (University of Mannheim)
OptimizationExplainability and InterpretabilityReinforcement LearningTabular
🎯 What it does: This paper presents SYMPOL, a method that can directly optimize axis-aligned decision tree (DT) policies within a policy gradient-based online reinforcement learning framework, achieving interpretable policies without information loss.
Mitigating Memorization in Language Models
Mansi Sakarvadia (University of Chicago), Michael W. Mahoney (Lawrence Berkeley National Laboratory)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper addresses the issue of language models (LM) potentially 'remembering' training data during inference, leading to privacy or copyright information leakage, and proposes and evaluates various memory mitigation methods.
Mitigating Modality Prior-Induced Hallucinations in Multimodal Large Language Models via Deciphering Attention Causality
Guanyu Zhou (Hong Kong University of Science and Technology), Xuming Hu (Hong Kong University of Science and Technology)
GenerationData SynthesisOptimizationTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: The CAUSALMM method is proposed, which utilizes structural causal models and counterfactual reasoning to intervene in the visual and language attention layers of multimodal LLMs, aiming to mitigate hallucinations caused by modality priors and enhance alignment quality.
Mitigating Object Hallucination in MLLMs via Data-augmented Phrase-level Alignment
Pritam Sarkar (Queen's University), Tomas Pfister (Google Cloud AI Research)
Object DetectionGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: A data augmentation-based phrase-level alignment (DPA) method is proposed, which utilizes generated 'hallucination' and 'correct' response pairs to finely constrain the language generated by multimodal large models, significantly reducing object hallucination.
Mitigating Parameter Interference in Model Merging via Sharpness-Aware Fine-Tuning
Yeoreum Lee (Hanyang University), Sungyong Baik (Hanyang University)
ClassificationOptimizationTransformerSupervised Fine-TuningImage
🎯 What it does: A Sharpness-Aware Fine-Tuning (SAFT) method is proposed within a pre-training-fine-tuning framework to reduce parameter interference in multi-task model fusion by finding flat local minima without the need for joint training.
Mitigating Reward Over-Optimization in RLHF via Behavior-Supported Regularization
Juntao Dai (Zhejiang University), Gang Pan (Zhejiang University)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a Behavior-Supported Policy Optimization (BSPO) method, which defines the behavior policy using the next-word distribution of reward training data and incorporates a behavior-supported Bellman operator into the value function to suppress reward over-optimization caused by out-of-distribution (OOD) responses.
Mitigating Spurious Correlations in Zero-Shot Multimodal Models
Shenyu Lu (Purdue University), Xiaoqian Wang (Purdue University)
ClassificationRecognitionImage TranslationTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a text-prompt-based image embedding translation method (TIE) and its unlabeled version (TIE*), which eliminates pseudo-correlation and enhances group robustness in zero-shot Vision-Language models.
Mitigating the Backdoor Effect for Multi-Task Model Merging via Safety-Aware Subspace
Jinluan Yang (Zhejiang University), Fei Wu (Zhejiang University)
ClassificationSafty and PrivacyMeta LearningTransformerImage
🎯 What it does: This paper proposes a security-aware multi-task model merging method (DAM) to address the security issues of backdoor injection during the model merging process.
Mix-CPT: A Domain Adaptation Framework via Decoupling Knowledge Learning and Format Alignment
Jinhao Jiang (Renmin University of China), Ji-Rong Wen (Renmin University of China)
Domain AdaptationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: In the Mix-CPT framework, the original domain text is mixed with general instructions and alignment data for continuous pre-training, followed by format alignment using a small number of seen instructions to achieve domain adaptation for LLMs;
Mix-LN: Unleashing the Power of Deeper Layers by Combining Pre-LN and Post-LN
Pengxiang Li (Dalian University of Technology), Shiwei Liu (University of Oxford)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper studies the phenomenon of deep inactivity in large language models and proposes the Mix-LN normalization technique, which balances gradients across all layers, thereby improving model pre-training and downstream performance.
MixEval-X: Any-to-any Evaluations from Real-world Data Mixture
Jinjie Ni (National University of Singapore), Michael Shieh
ImageVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Proposes the MixEval-X benchmark, covering multimodal tasks with arbitrary input-output and reconstructing a low-bias evaluation system through real-world queries;
MixMax: Distributional Robustness in Function Space via Optimal Data Mixtures
Anvith Thudi (University of Toronto), Chris J. Maddison (University of Toronto)
OptimizationTabularSequential
🎯 What it does: A group distributionally robust optimization method called MixMax is proposed, which achieves group DRO through optimal data mixing.