ICML 2024 Papers — Page 15
International Conference on Machine Learning · 2610 papers
Matroid Semi-Bandits in Sublinear Time
Ruo-Chun Tzeng (KTH Royal Institute of Technology), Kaito Ariu (CyberAgent)
Optimization
🎯 What it does: This paper proposes FasterCUCB, an algorithm that can solve the matroid semi-bandit problem in sublinear time;
MaxMin-RLHF: Alignment with Diverse Human Preferences
Souradip Chakraborty (University of Maryland), Mengdi Wang (Princeton University)
OptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningText
🎯 What it does: Proposes the MaxMin-RLHF framework, which utilizes multiple reward models and a max-min (egalitarian) objective to achieve alignment with diverse human preferences;
MC-GTA: Metric-Constrained Model-Based Clustering using Goodness-of-fit Tests with Autocorrelations
Zhangyu Wang (University of California), Ni Lao (Google)
Gaussian SplattingTime Series
🎯 What it does: A new metric-constrained clustering method, MC-GTA, is proposed, which utilizes the Gaussian Markov Random Field model of each observation point and the Wasserstein-2 distance to measure feature similarity, and constructs the clustering objective through multivariate semi-variogram and goodness-of-fit tests.
MD tree: a model-diagnostic tree grown on loss landscape
Yefan Zhou (Dartmouth College), Yaoqing Yang (Zhejiang University)
OptimizationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a model diagnostic tree (MD tree) based on loss landscape metrics, which can diagnose the sources of failure in pre-trained neural networks (such as optimizer hyperparameters, model size, or insufficient data) without retraining and without accessing the original training configuration.
Mean Estimation in the Add-Remove Model of Differential Privacy
Alex Kulesza (Google Research), Yuyan Wang (Google Research)
OptimizationSafty and PrivacyTabular
🎯 What it does: This paper addresses the one-dimensional mean estimation problem under the add/remove neighborhood model, proposing a new algorithm and proving that it achieves the minimum mean square error lower bound for all ε.
Mean Field Langevin Actor-Critic: Faster Convergence and Global Optimality beyond Lazy Learning
Kakei Yamamoto (Massachusetts Institute of Technology), Taiji Suzuki (University of Tokyo)
OptimizationReinforcement LearningSequential
🎯 What it does: This paper proposes the Mean Field Langevin Dynamics-based Actor-Critic algorithm (MFLAC), which achieves feature learning for policy and value functions through two layers of over-parameterized neural networks, and provides theoretical proofs for global optimality and linear convergence.
Mean-field Analysis on Two-layer Neural Networks from a Kernel Perspective
Shokichi Takakura (LY Corporation), Taiji Suzuki (University of Tokyo)
OptimizationRepresentation LearningTabularStochastic Differential Equation
🎯 What it does: This paper studies the feature learning capability of two-layer neural networks in the mean field state, adopting a kernel method perspective and utilizing a double time scale limit to analyze the dynamics of the kernel induced by the first layer.
Mean-field Chaos Diffusion Models
Sungwoo Park (University of California Berkeley), Ahmed Alaa
GenerationData SynthesisDiffusion modelPoint CloudStochastic Differential Equation
🎯 What it does: Mean-field Chaos Diffusion Models (MF-CDMs) are proposed, which are fractional generative models for high cardinality data (such as large-scale point clouds);
Mean-field Underdamped Langevin Dynamics and its Spacetime Discretization
Qiang Fu (Yale University), Ashia Camage Wilson
OptimizationStochastic Differential Equation
🎯 What it does: A N-particle algorithm (N-ULA) based on average field undamped Langevin dynamics is proposed for high-dimensional entropy-regularized mean field optimization problems.
Measures of diversity and space-filling designs for categorical data
Cedric Malherbe (AstraZeneca), Tom Diethe
OptimizationHyperparameter SearchGraph Neural NetworkReinforcement LearningAgentic AIGraphTabular
🎯 What it does: This paper focuses on constructing space-filling designs for discrete (categorical) data spaces and proposes a subset selection method based on combinatorial optimization.
Measuring Stochastic Data Complexity with Boltzmann Influence Functions
Nathan Hoyen Ng (Massachusetts Institute of Technology), Marzyeh Ghassemi (Massachusetts Institute of Technology)
Anomaly DetectionConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a method called IF-COMP, which uses temperature-scaled Boltzmann Influence Functions (BIF) to approximate the predictive normalized maximum likelihood (pNML) distribution, thereby efficiently estimating the random data complexity of deep neural networks. This complexity is utilized for adaptive uncertainty calibration, label noise detection, and out-of-distribution (OOD) detection.
Mechanistic Design and Scaling of Hybrid Architectures
Michael Poli (Stanford University), Stefano Massaroli (RIKEN)
OptimizationNeural Architecture SearchTransformerLarge Language ModelText
🎯 What it does: By constructing a series of small-scale synthetic token operation tasks and evaluating models based on this, a process called 'Mechanized Architecture Design (MAD)' is proposed for rapid prototyping and validation of new deep learning architectures.
Mechanistic Neural Networks for Scientific Machine Learning
Adeel Pervez (University of Amsterdam), Stratis Gavves
Time SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes Mechanistic Neural Networks (MNN), which can explicitly learn and solve differential equations by embedding Mechanistic Blocks in neural networks, thereby discovering and simulating dynamic systems directly from data.
Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads
Tianle Cai (Princeton University), Tri Dao
GenerationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: To improve the inference speed of large language models, the paper proposes the MEDUSA framework, which adds multiple decoding heads on the last hidden layer of the original model, allowing for the simultaneous prediction of multiple subsequent tokens and significantly reducing the number of serialized inference steps through tree-structured attention for parallel verification of candidate continuations.
Membership Inference Attacks on Diffusion Models via Quantile Regression
Shuai Tang (Amazon AWS AI/ML), Aaron Roth (University of Pennsylvania)
Anomaly DetectionAdversarial AttackConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: A member inference attack based on quantile regression is proposed to detect whether a certain sample was used for training in diffusion models.
Memoria: Resolving Fateful Forgetting Problem through Human-Inspired Memory Architecture
Sangjun Park (Sungkyunkwan University), JinYeong Bak (Sungkyunkwan University)
ClassificationData-Centric LearningTransformerLarge Language ModelTextSequential
🎯 What it does: A new external memory architecture named Memoria is proposed, mimicking human memory mechanisms to address the problem of long-term forgetting (Fateful Forgetting) in neural networks.
Memorization Through the Lens of Curvature of Loss Function Around Samples
Isha Garg (Purdue University), Kaushik Roy (Purdue University)
Anomaly DetectionComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This paper proposes using the average value of the curvature of the loss function (the square trace of the Hessian) near training samples as an indicator of whether the samples are memorized by the network.
Memory Consolidation Enables Long-Context Video Understanding
Ivana Balazevic (Google DeepMind), Olivier J Henaff (Google DeepMind)
RecognitionComputational EfficiencyTransformerContrastive LearningVideo
🎯 What it does: Using a pre-trained video Transformer, we enhance its modeling capability for long temporal videos through a non-parametric memory integration method.
Memory Efficient Neural Processes via Constant Memory Attention Block
Leo Feng (Mila - Université de Montréal), Mohamed Osama Ahmed (Borealis AI)
GenerationData SynthesisComputational EfficiencyMeta LearningImageTime Series
🎯 What it does: This study proposes a Constant Memory Attention Block (CMAB) and integrates it into Neural Processes (NP), forming Constant Memory Attentive Neural Processes (CMANP), achieving an NP model that uses constant memory and constant computation during the conditioning, querying, and updating phases.
Memory-Space Visual Prompting for Efficient Vision-Language Fine-Tuning
Shibo Jie (Peking University), Yunhe Wang (Huawei)
GenerationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: A method called MemVP is proposed, which directly inserts visual prompts into the FFN weights (memory space) of the language model in visual language models for parameter-efficient fine-tuning.
MEMORYLLM: Towards Self-Updatable Large Language Models
Yu Wang (University of California San Diego), Julian McAuley (University of California San Diego)
TransformerLarge Language ModelText
🎯 What it does: Designed and implemented MEMORYLLM, a large language model (LLM) that embeds a fixed-size memory pool within the latent space of a Transformer, supporting self-updating of new knowledge while retaining old knowledge.
Merging Multi-Task Models via Weight-Ensembling Mixture of Experts
Anke Tang (Wuhan University), Dacheng Tao (Nanyang Technological University)
ClassificationTransformerMixture of ExpertsImage
🎯 What it does: This paper studies the implementation of a unified model by combining a multi-task Transformer model with weight fusion and Mixture-of-Experts (MoE).
Meta Evidential Transformer for Few-Shot Open-Set Recognition
Hitesh Sapkota (Amazon Inc), Qi Yu (Rochester Institute of Technology)
RecognitionMeta LearningTransformerImage
🎯 What it does: A new method for few-shot open set recognition, MET, is proposed, which can identify unknown categories with only a small number of labeled samples.
Meta-Learners for Partially-Identified Treatment Effects Across Multiple Environments
Jonas Schweisthal (Ludwig Maximilian University of Munich), Stefan Feuerriegel (Ludwig Maximilian University of Munich)
Meta LearningTabularTime Series
🎯 What it does: A method for estimating the boundaries of conditional average treatment effects (CATE) for multi-environment observational data is proposed, which derives closed-form upper and lower bounds by treating environmental variables as instrumental variables, and provides a meta-learner implementation that is independent of the two-stage model.
Meta-Reinforcement Learning Robust to Distributional Shift Via Performing Lifelong In-Context Learning
Tengye Xu (Zhejiang University), Qinyuan Ren (Zhejiang University)
Meta LearningTransformerReinforcement LearningTabularSequential
🎯 What it does: A Meta-RL method named PSBL is proposed, which achieves robustness to task distribution drift by utilizing a pre-trained transformer (LILTrans) for gradient-free online inference during the testing phase.
MF-CLR: Multi-Frequency Contrastive Learning Representation for Time Series
Jufang Duan (Lenovo Research), Hongsheng Qi (Lenovo Research)
ClassificationAnomaly DetectionRepresentation LearningRecurrent Neural NetworkContrastive LearningTime SeriesFinance Related
🎯 What it does: Designed and implemented MF-CLR, a self-supervised contrastive learning framework for learning universal representations of multi-frequency time series.
MFTN: A Multi-scale Feature Transfer Network Based on IMatchFormer for Hyperspectral Image Super-Resolution
Shuying Huang (Tiangong University), Yingzhi Wei (Tiangong University)
RestorationSuper ResolutionTransformerImage
🎯 What it does: A multi-scale feature transfer network (MFTN) based on the improved feature matching Transformer (IMatchFormer) is proposed, utilizing cross-modal feature matching and multi-scale dynamic aggregation between low-resolution hyperspectral images (LR-HSI) and high-resolution multispectral images (HR-MSI) to achieve super-resolution reconstruction of high-resolution hyperspectral images (HR-HSI).
MGit: A Model Versioning and Management System
Wei Hao (Columbia University), Junfeng Yang (Columbia University)
CompressionFederated LearningTransformerSupervised Fine-TuningGraph
🎯 What it does: A model versioning and management system named MGit has been designed and implemented, which records model evolution traces and provides storage optimization, testing, updating, and collaboration features.
MH-pFLID: Model Heterogeneous personalized Federated Learning via Injection and Distillation for Medical Data Analysis
Luyuan Xie (Peking University), Zhonghai Wu (Peking University)
ClassificationSegmentationFederated LearningKnowledge DistillationImageTime SeriesBiomedical Data
🎯 What it does: A lightweight information transmission framework MH-pFLID based on injection and distillation is designed to achieve personalized federated learning in multi-model heterogeneous and non-IID data environments.
MILP-FBGen: LP/MILP Instance Generation with Feasibility/Boundedness
Yahong Zhang (Lenovo Research), Junchi Yan (Shanghai Jiao Tong University)
Data SynthesisOptimizationGraph Neural NetworkDiffusion modelTabular
🎯 What it does: A framework for generating instances of linear programming/mixed-integer linear programming based on diffusion models, called MILP-FBGen, is proposed, which ensures the feasibility and boundedness of the generated instances while maintaining structural similarity.
Mimicking Better by Matching the Approximate Action Distribution
Joao Candido Ramos, Alexandros Kalousis (University of Applied Sciences and Arts Western)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: A new observation-based imitation learning algorithm called MAAD is proposed, which infers missing expert actions using an inverse dynamics model and guides policy learning through behavior cloning regularization.
Mind the Boundary: Coreset Selection via Reconstructing the Decision Boundary
Shuo Yang (University of Hong Kong), Liqiang Nie (Harbin Institute of Technology)
ClassificationOptimizationAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: By approximating the distance from training samples to the decision boundary, support vectors are selected to form a core set, which is used to reconstruct the decision boundary of the complete data model.
MindEye2: Shared-Subject Models Enable fMRI-To-Image With 1 Hour of Data
Paul Steven Scotti (Stability AI), Tanishq Mathew Abraham (Stability AI)
GenerationRetrievalTransformerVision Language ModelDiffusion modelImageMultimodalityMagnetic Resonance Imaging
🎯 What it does: The MindEye2 model is proposed to reconstruct visual perception images using only 1 hour of fMRI data, pre-trained and fine-tuned on fMRI data from multiple subjects.
Minimally Modifying a Markov Game to Achieve Any Nash Equilibrium and Value
Young Wu (University of Wisconsin Madison), Qiaomin Xie (University of Wisconsin Madison)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: The study investigates how to minimize reward modification costs in zero-sum Markov games, such that the target strategy becomes the unique Markov perfect Nash equilibrium and meets specified value ranges;
Minimax Optimality of Score-based Diffusion Models: Beyond the Density Lower Bound Assumptions
Kaihong Zhang (University of Illinois Urbana-Champaign), Jingbo Liu (University of Illinois Urbana-Champaign)
GenerationData SynthesisOptimizationDiffusion modelScore-based Model
🎯 What it does: Analyzed the sampling error of score-based diffusion models in the case of large samples, and proved that under the sub-Gaussian assumption, optimal MSE and TV error upper bounds can be achieved, further demonstrating that minimal extremum can be reached on Sobolev class distributions (β≤2).
Minimizing $f$-Divergences by Interpolating Velocity Fields
Song Liu (University of Bristol), Mark Beaumont (University of Bristol)
Domain AdaptationTabular
🎯 What it does: This paper proposes a method to directly estimate the velocity field of Wasserstein Gradient Flow for various f-divergences using interpolation methods (Nadaraya–Watson and local linear regression), proving the consistency of the estimator and applying it to domain adaptation and missing data imputation.
Minimum Norm Interpolation Meets The Local Theory of Banach Spaces
Gil Kur (ETH Zürich), Fanny Yang (ETH Zürich)
🎯 What it does: A general geometric framework is proposed, utilizing the local theory of Banach spaces to analyze the generalization performance of minimum norm interpolators in over-parameterized regression.
Minimum-Norm Interpolation Under Covariate Shift
Neil Rohit Mallinar, Bin Yu (University of California Berkeley)
Domain AdaptationOptimizationImage
🎯 What it does: The study investigates the generalization behavior of the Minimum Norm Interpolator (MNI) in high-dimensional linear regression under covariate shift, proposing a beneficial and harmful shift classification method based on the degree of over-parameterization, and extending the theoretical results to neural networks;
Mitigating Catastrophic Forgetting in Online Continual Learning by Modeling Previous Task Interrelations via Pareto Optimization
Yichen Wu (City University of Hong Kong), Long-Kai Huang (Nanyang Technological University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a continuous learning algorithm based on Pareto optimization (POCL), which significantly alleviates catastrophic forgetting by jointly weighing the gradients of all learned tasks within a hierarchical gradient aggregation framework.
Mitigating Label Noise on Graphs via Topological Sample Selection
Yuhao Wu (University of Sydney), Tongliang Liu (University of Sydney)
ClassificationOptimizationGraph Neural NetworkGraph
🎯 What it does: A sample selection method based on graph topology information, TSS, is proposed to denoise noisy labels in graph data and enhance the robustness of GNNs.
Mitigating Oversmoothing Through Reverse Process of GNNs for Heterophilic Graphs
MoonJeong Park (Pohang University of Science and Technology), Dongwoo Kim (Pohang University of Science and Technology)
Graph Neural NetworkGraphOrdinary Differential Equation
🎯 What it does: A framework for graph neural networks (GNN) based on reverse propagation is proposed, which alleviates the over-smoothing problem caused by multi-layer aggregation through reverse message passing.
Mitigating Privacy Risk in Membership Inference by Convex-Concave Loss
Zhenlong Liu (Southern University of Science and Technology), Hongxin Wei (Southern University of Science and Technology)
OptimizationSafty and PrivacyConvolutional Neural NetworkImage
🎯 What it does: The Convex-Concave Loss (CCL) is proposed by incorporating a concave function into the cross-entropy loss, significantly improving the variance of training sample loss, thereby reducing the advantage of membership inference attacks (MIA) while maintaining model performance.
Mixtures of Experts Unlock Parameter Scaling for Deep RL
Johan Samir Obando Ceron, Pablo Samuel Castro (Google DeepMind)
Reinforcement LearningMixture of ExpertsSequential
🎯 What it does: Integrate the Mixture-of-Experts (especially Soft MoE) module into deep reinforcement learning value-based networks (DQN, Rainbow) and study its parameter scalability.
MLAgentBench: Evaluating Language Agents on Machine Learning Experimentation
Qian Huang (Stanford University), Jure Leskovec (Stanford University)
TransformerLarge Language ModelAgentic AIPrompt EngineeringImageTextGraphTabularBenchmark
🎯 What it does: A benchmark named MLAgentBench has been designed, which includes 13 diverse machine learning experimental tasks, and an automated experimental agent based on large language models has been implemented, capable of reading/writing files, executing code, and progressively improving model performance in the workspace.
MLI Formula: A Nearly Scale-Invariant Solution with Noise Perturbation
Bowen Tao (Nanjing University), De-Chuan Zhan (Nanjing University)
OptimizationConvolutional Neural NetworkTransformerImageText
🎯 What it does: This paper studies the phenomenon of linear interpolation (MLI) during the training process of neural networks, finding that even when the initialization is changed to random noise or all zeros, the error curve still shows a monotonic decrease, indicating that MLI is more dependent on the properties of the converging model rather than the optimization trajectory.
MLIP: Efficient Multi-Perspective Language-Image Pretraining with Exhaustive Data Utilization
Yu Zhang (Tongji University), Changwei Wang (Institute of Automation Chinese Academy of Sciences)
ClassificationRetrievalOptimizationComputational EfficiencyTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes the MLIP framework, which achieves efficient multi-view and multi-level language-image alignment by incorporating frequency domain transformation, cross-layer alignment, and controllable token merging into CLIP pre-training.
MLLM-as-a-Judge: Assessing Multimodal LLM-as-a-Judge with Vision-Language Benchmark
Dongping Chen (Huazhong University of Science and Technology), Lichao Sun (Lehigh University)
TransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Designed and implemented the MLLM-as-a-Judge benchmark to evaluate the judgment capabilities of multimodal large language models in scoring, comparison, and ranking tasks, and collected human annotations.
MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilities
Weihao Yu (National University of Singapore), Lijuan Wang (Microsoft Azure AI)
RecognitionGenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposes the MM-Vet benchmark, specifically designed to evaluate the performance of large multimodal models (LMM) in integrating various visual-language capabilities.
MMPareto: Boosting Multimodal Learning with Innocent Unimodal Assistance
Yake Wei (Renmin University of China), Di Hu (Renmin University of China)
OptimizationConvolutional Neural NetworkTransformerMultimodality
🎯 What it does: In multimodal learning, identify and mitigate the gradient conflicts between multimodal loss and corresponding unimodal loss, proposing the MMPareto algorithm, which achieves harmless unimodal assistance through direction unification and amplitude enhancement.
MMT-Bench: A Comprehensive Multimodal Benchmark for Evaluating Large Vision-Language Models Towards Multitask AGI
Kaining Ying (Shanghai Artificial Intelligence Laboratory), Wenqi Shao (Shanghai Jiao Tong University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper presents MMT-Bench, a comprehensive multimodal evaluation benchmark consisting of 31,325 multiple-choice visual questions, 32 core meta-tasks, and 162 sub-tasks, aimed at assessing the performance of large visual-language models (LVLM) in the direction of multi-task AGI.
Mobile Attention: Mobile-Friendly Linear-Attention for Vision Transformers
Zhiyu Yao (Baidu), Mingsheng Long (Tsinghua University)
Object DetectionSegmentationComputational EfficiencyTransformerImage
🎯 What it does: A Mobile-Attention mechanism is proposed to replace the standard attention in ViT with a more efficient linear attention;
MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases
Zechun Liu (Meta), Vikas Chandra (Meta)
OptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Designed and implemented a mobile-oriented sub-billion parameter LLM MobileLLM, providing an efficient language model;
Model Alignment as Prospect Theoretic Optimization
Kawin Ethayarajh (Stanford University), Douwe Kiela (Contextual AI)
OptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This study frames the alignment objectives of LLM as prospect theory optimization, proposing Human-Aware Losses (HALOs) and designing KTO loss to directly maximize the utility of text generation;
Model Assessment and Selection under Temporal Distribution Shift
Elise Han (Columbia University), Kaizheng Wang (Columbia University)
TabularTime Series
🎯 What it does: The research focuses on model evaluation and selection in environments with time distribution drift, proposing an adaptive rolling window estimation of errors and constructing a single-elimination tournament based on this to achieve near-optimal model selection.
Model Tailor: Mitigating Catastrophic Forgetting in Multi-modal Large Language Models
Didi Zhu (Zhejiang University), Kun Kuang (Tencent Youtu Lab)
Domain AdaptationOptimizationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: This paper proposes Model Tailor, a post-training parameter-efficient method that selects key fine-tuned parameters (model patches) through sparse masking and utilizes Hessian inverse estimation for compensation (decorative patches). By modifying only ≤10% of the parameters, it can significantly alleviate catastrophic forgetting in multimodal large models.
Model-Based Minimum Bayes Risk Decoding for Text Generation
Yuu Jinnai (CyberAgent), Kenshi Abe (CyberAgent)
GenerationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a model probability-based Minimum Bayesian Risk (MBMBR) decoding method, which directly uses model probabilities from a set of candidate sentences obtained through sampling to calculate expected utility, rather than relying on Monte Carlo estimation, thereby achieving more accurate text generation.
Model-based Reinforcement Learning for Confounded POMDPs
Mao Hong (Johns Hopkins University), Yanxun Xu (Johns Hopkins University)
OptimizationReinforcement Learning
🎯 What it does: A model-based offline reinforcement learning algorithm is proposed, which implements policy value identification and learning for POMDP problems with hidden state confusion in continuous state and observation spaces.
Model-based Reinforcement Learning for Parameterized Action Spaces
Renhao Zhang (Brown University), George Konidaris (Brown University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: In the parameterized action space Markov decision process, a model-based reinforcement learning framework called DLPA is proposed, which learns action-conditioned dynamics and plans through an improved MPPI;
Model-Based RL for Mean-Field Games is not Statistically Harder than Single-Agent RL
Jiawei Huang (ETH Zurich), Andreas Krause (ETH Zurich)
Reinforcement Learning
🎯 What it does: This paper studies the sample complexity in model-based mean field game reinforcement learning (MFG), proposes a new Partial Model-Based Eluder Dimension (P-MBED) metric, and proves that under reasonable assumptions, learning MFG is statistically as difficult as single-agent RL.
Model-Free Robust $\phi$-Divergence Reinforcement Learning Using Both Offline and Online Data
Kishan Panaganti (California Institute of Technology), Eric Mazumdar (California Institute of Technology)
Reinforcement LearningTabular
🎯 What it does: Two robust reinforcement learning algorithms are proposed: RPQ (suitable for offline data, supporting arbitrary φ-divergence) and HyTQ (hybrid offline + online data, targeting TV divergence), along with a unified theoretical suboptimality and sample complexity analysis.
Modeling Caption Diversity in Contrastive Vision-Language Pretraining
Samuel Lavoie (Meta), Nicolas Ballas (Meta)
GenerationRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A contrastive visual-language pre-training framework named Llip is proposed, which can generate diverse image representations based on different text descriptions.
Modeling Language Tokens as Functionals of Semantic Fields
Zhengqi Pei (Institute of Computing Technology Chinese Academy of Sciences), Qingming Huang (School of Computer Science and Technology University of Chinese Academy of Sciences)
TransformerLarge Language ModelText
🎯 What it does: The LasF module is proposed to simulate the nonlinear mapping of linguistic tokens using the functional representation of semantic fields, thereby achieving more efficient language modeling.
Modelling Microbial Communities with Graph Neural Networks
Albane Ruaud (University of Tübingen), Georg Martius (Max Planck Institute for Intelligent Systems)
Graph Neural NetworkGraph
🎯 What it does: This paper studies the direct prediction of the relative abundance of microbial communities in a steady state using graph neural networks (GNNs) from bacterial genomic features, bypassing the traditional dynamic modeling that requires time series data.
Modular Learning of Deep Causal Generative Models for High-dimensional Causal Inference
Md Musfiqur Rahman (Purdue University), Murat Kocaoglu (Purdue University)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: A modular adversarial training algorithm, Modular-DCM, is proposed for constructing deep causal generative models (DCM), enabling accurate estimation of identifiable causal effects and causal sampling in high-dimensional data (such as images) with potential confounding factors.
MOKD: Cross-domain Finetuning for Few-shot Classification via Maximizing Optimized Kernel Dependence
Hongduan Tian (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
ClassificationDomain AdaptationMeta LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes a two-layer framework MOKD based on optimized kernel dependency for cross-domain few-shot classification fine-tuning.
Mol-AE: Auto-Encoder Based Molecular Representation Learning With 3D Cloze Test Objective
Junwei Yang (Peking University), Hao Zhou (Tsinghua University)
Representation LearningDrug DiscoveryTransformerAuto EncoderGraph
🎯 What it does: This paper proposes a 3D molecular representation learning framework called MOL-AE based on autoencoders, and designs a new 3D Cloze Test training objective to address the inconsistency between pre-training and downstream tasks in traditional coordinate denoising frameworks, as well as the twisted optimization problem caused by coordinate denoising.
MolCRAFT: Structure-Based Drug Design in Continuous Parameter Space
Yanru Qu (Tsinghua University), Wei-Ying Ma (Tsinghua University)
Drug DiscoveryGraph Neural NetworkFlow-based ModelBiomedical Data
🎯 What it does: A generative model called MolCRAFT is proposed for structure-based drug design in continuous parameter space, addressing the issues of mode collapse and discrete-continuous space mismatch in traditional autoregressive and diffusion models for 3D conformation generation.
Mollification Effects of Policy Gradient Methods
Tao Wang (University of California), Sicun Gao
OptimizationReinforcement LearningSequentialPhysics Related
🎯 What it does: This paper establishes a theoretical framework by mapping policy gradient optimization to the backward solution of the heat equation, explaining how policy gradients achieve 'mollification' through Gaussian noise in continuous control problems, and analyzing the benefits and limitations brought by mollification.
MOMENT: A Family of Open Time-series Foundation Models
Mononito Goswami (Carnegie Mellon University), Artur Dubrawski (Carnegie Mellon University)
ClassificationAnomaly DetectionTransformerTime Series
🎯 What it does: MOMENT is proposed and released—a series of Transformer-based foundational models for time series, and a large-scale multi-domain public dataset called Time Series Pile is constructed for unsupervised pre-training;
Momentor: Advancing Video Large Language Model with Fine-Grained Temporal Reasoning
Long Qian (Zhejiang University), Siliang Tang (National University of Singapore)
RecognitionRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodality
🎯 What it does: Momentor has been developed, a video large language model with fine-grained temporal reasoning capabilities, trained on the automatically generated Moment-10M dataset;
Momentum for the Win: Collaborative Federated Reinforcement Learning across Heterogeneous Environments
Han Wang (Columbia University), James Anderson (Columbia University)
Federated LearningReinforcement LearningSequential
🎯 What it does: Two momentum-based joint reinforcement learning algorithms, FEDSVRPG-M and FEDHAPG-M, are proposed to solve shared policies under arbitrary levels of environmental heterogeneity.
Momentum Particle Maximum Likelihood
Jen Ning Lim (University of Warwick), Adam Michael Johansen
OptimizationImageStochastic Differential Equation
🎯 What it does: A Momentum Particle Descent (MPD) algorithm is designed, introducing a momentum mechanism into particle gradient descent for maximum likelihood estimation of latent variable models.
MoMo: Momentum Models for Adaptive Learning Rates
Fabian Schaipp (Technical University of Munich), Robert M. Gower (Flatiron Institute)
OptimizationImage
🎯 What it does: This paper studies the adaptive learning rate MoMo, which can be used for any momentum optimizer, and its Adam version.
Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews
Weixin Liang (Stanford University), James Y. Zou (Stanford University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This study proposes a maximum likelihood-based distribution estimation method that constructs a reference distribution using historical human writing and LLM-generated texts. It can quickly estimate the proportion of texts significantly modified or generated by LLMs in large corpora without the need for document-by-document classification. This method is applied to peer review texts from major AI conferences and the Nature journal in recent years to assess the usage of ChatGPT in the review process.
Monotone Individual Fairness
Yahav Bechavod (University of Pennsylvania)
Optimization
🎯 What it does: A new monotonic individual fairness auditing framework is proposed, along with an online learning algorithm that achieves optimal expected error and fairness violation counts.
Monotone, Bi-Lipschitz, and Polyak-Łojasiewicz Networks
Ruigang Wang (University of Sydney), Ian Manchester
OptimizationImage
🎯 What it does: This paper proposes BiLipNet (reversible bi-Lipschitz network) and PLNet (scalar output network satisfying the Polyak-Łojasiewicz condition), and presents a method to construct reversible residual layers while ensuring strong monotonicity and Lipschitz continuity.
More Benefits of Being Distributional: Second-Order Bounds for Reinforcement Learning
Kaiwen Wang (Cornell University), Wen Sun (Cornell University)
Reinforcement LearningTabular
🎯 What it does: Proves that distributed reinforcement learning (DistRL) can achieve second-order (variance-related) upper bounds in both online and offline RL, and implements this theory in low-rank MDPs and contextual bandit problems;
More Flexible PAC-Bayesian Meta-Learning by Learning Learning Algorithms
Hossein Zakerinia (Institute of Science and Technology Austria), Christoph H. Lampert (Institute of Science and Technology Austria)
Meta LearningImageTabular
🎯 What it does: A new PAC-Bayesian framework is proposed to directly analyze the generalization of meta-learning algorithms and derive more general risk upper bounds.
Moreau Envelope for Nonconvex Bi-Level Optimization: A Single-Loop and Hessian-Free Solution Strategy
Risheng Liu (Dalian University of Technology), Jin Zhang (Southern University of Science and Technology)
OptimizationTabular
🎯 What it does: A single-loop, Hessian-free Moreau envelope-based algorithm (MEHA) is proposed for solving non-convex-non-convex bilevel optimization (BLO) problems.
MorphGrower: A Synchronized Layer-by-layer Growing Approach for Plausible Neuronal Morphology Generation
Nianzu Yang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
GenerationData SynthesisRecurrent Neural NetworkGraph Neural NetworkAuto EncoderGraphBiomedical Data
🎯 What it does: The MorphGrower method is proposed, which generates neuron morphologies through layer-by-layer synchronous growth, incrementally generating based on the tree structure of real samples.
MS-TIP: Imputation Aware Pedestrian Trajectory Prediction
Pranav singh chib, Pravendra Singh (Indian Institute of Technology)
Graph Neural NetworkTransformerTime SeriesBenchmark
🎯 What it does: An end-to-end MS-TIP framework is proposed, which can simultaneously perform trajectory missing value imputation and future pedestrian trajectory prediction, utilizing self-attention imputation combined with multi-scale hypergraphs and scene attention for interactive modeling.
MS$^3$D: A RG Flow-Based Regularization for GAN Training with Limited Data
Jian Wang (Sichuan University), Jiancheng Lv (Sichuan University)
GenerationData SynthesisFlow-based ModelGenerative Adversarial NetworkImagePhysics Related
🎯 What it does: In the context of GAN training with limited data, a multi-scale structure self-similarity (MS D) regularization method based on RG flow is proposed, which constrains the discriminator's gradient field to remain consistent across different scales, thereby enhancing the robustness of the generator and the quality of the images.
Multi-Agent Reinforcement Learning Meets Leaf Sequencing in Radiotherapy
Riqiang Gao (Siemens Healthineers), Ali Kamen (Siemens Healthineers)
OptimizationReinforcement LearningBiomedical Data
🎯 What it does: A blade serialization model RLS based on multi-agent deep reinforcement learning is proposed to replace traditional optimization methods.
Multi-Agent Reinforcement Learning with Hierarchical Coordination for Emergency Responder Stationing
Amutheezan Sivagnanam (Pennsylvania State University), Aron Laszka (Vanderbilt University)
OptimizationTransformerReinforcement LearningAgentic AITabularTime Series
🎯 What it does: This paper proposes a hierarchical coordination framework based on multi-agent reinforcement learning for the proactive relocation of rescue vehicles in emergency response management systems.
Multi-Factor Adaptive Vision Selection for Egocentric Video Question Answering
Haoyu Zhang (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
RecognitionObject DetectionTransformerVideo
🎯 What it does: A multi-factor adaptive visual selection framework named MFAS is proposed to address the challenges of small object recognition, noise suppression, and spatiotemporal reasoning in self-perspective video question answering.
Multi-Fidelity Residual Neural Processes for Scalable Surrogate Modeling
Ruijia Niu (University of California San Diego), Rose Yu (University of California San Diego)
Time SeriesPhysics Related
🎯 What it does: A multi-fidelity residual neural process framework (MFRNP) is proposed, which achieves scalable surrogate modeling by aggregating low-fidelity model predictions and explicitly modeling the residuals.
Multi-group Learning for Hierarchical Groups
Samuel Deng (Columbia University), Daniel Hsu (Columbia University)
ClassificationExplainability and InterpretabilityTabular
🎯 What it does: This study investigates the theory and algorithms of multi-group learning under hierarchical grouping, proposing a multi-group learning algorithm that can output interpretable decision tree predictors and conducting experimental validation.
Multi-layer Rehearsal Feature Augmentation for Class-Incremental Learning
Bowen Zheng (Nanjing University), De-Chuan Zhan (Nanjing University)
ClassificationKnowledge DistillationContrastive LearningImage
🎯 What it does: This paper studies the overfitting problem of replay samples in continual learning and proposes a Multi-layer Replay Feature Augmentation (MRFA) method, which enhances the features of replay samples through gradient ascent at each layer to increase the inter-layer margin, thereby alleviating catastrophic forgetting.
Multi-Patch Prediction: Adapting Language Models for Time Series Representation Learning
Yuxuan Bian (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
Anomaly DetectionRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTime Series
🎯 What it does: Transforming time series forecasting into a self-supervised multi-patch prediction task, utilizing causal Transformer (GPT2) to learn time series representations under a two-stage pre-training + fine-tuning framework.
Multi-Region Markovian Gaussian Process: An Efficient Method to Discover Directional Communications Across Multiple Brain Regions
Weihan Li (Georgia Institute of Technology), Anqi Wu (Georgia Institute of Technology)
Gaussian SplattingTime SeriesBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A multi-regional Markov Gaussian process (MRM-GP) is proposed, which combines multi-output GP with linear dynamical systems to efficiently capture directional communication related to frequency between brain regions.
Multi-Sender Persuasion: A Computational Perspective
Safwan Hossain (Harvard University), Haifeng Xu (University of Chicago)
OptimizationReinforcement Learning
🎯 What it does: This paper studies the Bayesian persuasion problem with multiple information senders, proving that solving the optimal sending strategy and Nash equilibrium is computationally extremely difficult. It proposes a method that combines a differentiable deep network with an extra gradient algorithm to find ε-local Nash equilibria. Experiments show that this method significantly outperforms traditional ReLU, DeLU networks, and full information revelation equilibria, achieving higher sender utility and social welfare.
Multi-Source Conformal Inference Under Distribution Shift
Yi Liu (North Carolina State University), Larry Han (Northeastern University)
OptimizationFederated LearningTabularElectronic Health Records
🎯 What it does: A distribution-independent compliant inference method is proposed under multi-source data distribution shift to construct prediction intervals for the target population with missing outcomes.
Multi-Track Message Passing: Tackling Oversmoothing and Oversquashing in Graph Learning via Preventing Heterophily Mixing
Hongbin Pei (Xi'an Jiaotong University), Xiaohong Guan
Graph Neural NetworkGraph
🎯 What it does: A multi-track message passing model (MTGCN) is proposed, which prevents heterogeneous mixing by independently propagating messages on different category tracks, addressing the issues of oversmoothing and overcompression in graph neural networks.
Multi-View Clustering by Inter-cluster Connectivity Guided Reward
Hao Dai (Sichuan University), Jiancheng Lv (Sichuan University)
Graph Neural NetworkReinforcement LearningGraph
🎯 What it does: A graph-based multi-view clustering algorithm is designed, which automatically infers the unknown number of clusters k through reinforcement learning during the iterative process and provides the final clustering results.
Multi-View Stochastic Block Models
Vincent Cohen-Addad (Google Research), Rajai Nasser (Google Research)
Graph
🎯 What it does: A Multi-View Stochastic Block Model (Multi-View SBM) is proposed, along with an algorithm for weakly recovering community labels from multiple graphs.
Multicalibration for Confidence Scoring in LLMs
Gianluca Detommaso (Amazon Web Services), Aaron Roth (University of Pennsylvania)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes using multi-calibration techniques to provide interpretable and reliable confidence scores for answers generated by LLMs.
Multigroup Robustness
Lunjia Hu (Stanford University), Judy Hanwen Shen (Stanford University)
ClassificationOptimizationTabular
🎯 What it does: The concept of Multigroup Robustness is proposed, ensuring that when the training data is corrupted, the prediction error of each subgroup is only affected by the degree of corruption within itself;
MultiMax: Sparse and Multi-Modal Attention Learning
Yuxuan Zhou (University of Mannheim), Margret Keuper (Max Planck Institute for Informatics)
ClassificationTransformerImageTextMultimodality
🎯 What it does: A learnable MultiMax function is proposed to replace the traditional SoftMax, achieving sparse and multimodal weight allocation for attention and classification outputs.
Multimodal Prototyping for cancer survival prediction
Andrew H. Song (Mass General Brigham), Faisal Mahmood (Mass General Brigham)
ClassificationExplainability and InterpretabilityComputational EfficiencyTransformerMultimodalityBiomedical Data
🎯 What it does: This paper proposes a multimodal prototyping framework (MMP) to compress whole slide images and transcriptomic data into a small number of prototypes, which are then fused to predict the survival time of cancer patients.
Multiplicative Weights Update, Area Convexity and Random Coordinate Descent for Densest Subgraph Problems
Ta Duy Nguyen (Boston University), Alina Ene (Boston University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: Three iterative algorithms based on Multiplicative Weight Update (MWU), area convexity, and Random Coordinate Descent (RCD) are proposed to solve the Dense Subgraph (DSG) and Dense Subgraph Decomposition problems;