ICML 2025 Papers — Page 17
International Conference on Machine Learning · 3257 papers
Learning curves theory for hierarchically compositional data with power-law distributed features
Francesco Cagnetta (International School for Advanced Studies), Matthieu Wyart (Johns Hopkins University)
ClassificationConvolutional Neural NetworkTransformerSequential
🎯 What it does: By analyzing the hierarchical structure of the random hierarchical model (RHM) that includes power-law distribution generation rules, the learning curves for classification and next token prediction tasks were studied;
Learning Distances from Data with Normalizing Flows and Score Matching
Peter Sorrenson (Heidelberg University), Ullrich Koethe
OptimizationScore-based ModelFlow-based ModelAuto EncoderTabularSequential
🎯 What it does: This paper proposes the use of normalized flows and sliced score matching to learn density-based Fermat distances, and solves geodesics through a relaxation algorithm, addressing the issues of slow convergence and high-dimensional failure in traditional nearest neighbor graph methods.
Learning Distribution-wise Control in Representation Space for Language Models
Chunyuan Deng (Rice University), Hanjie Chen (Rice University)
OptimizationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes to use a distributed (random) intervention method in the representation space of language models instead of traditional point interventions to achieve finer-grained behavior control.
Learning Dynamics in Continual Pre-Training for Large Language Models
Xingjin Wang (University of Chinese Academy of Sciences), Daniel Dajun Zeng (University of Chinese Academy of Sciences)
Large Language ModelText
🎯 What it does: A unified learning dynamics framework has been constructed to describe and predict the validation loss curve of large language models during the continuous pre-training (CPT) process; by decomposing the effects of learning rate annealing and distribution shift, a closed-form scaling law for CPT has been proposed, and based on this, the roles of key hyperparameters such as loss potential, replay ratio, peak learning rate, and training steps have been analyzed.
Learning dynamics in linear recurrent neural networks
Alexandra Maria Proca, Pedro A. M. Mediano (University College London)
Recurrent Neural NetworkTime SeriesSequential
🎯 What it does: Analyzed the learning dynamics of linear recurrent neural networks (LRNN) and constructed a theoretical framework that incorporates task dynamics into the learning process, analyzing properties such as learning rate, stability, low-rank connections, and feature learning.
Learning Dynamics under Environmental Constraints via Measurement-Induced Bundle Structures
Dongzhe Zheng (Shanghai Jiao Tong University), Wenjie Mei (Southeast University)
OptimizationSafty and PrivacyRobotic IntelligenceTime SeriesOrdinary Differential Equation
🎯 What it does: A secure dynamic learning method based on fiber bundle geometric framework is proposed under incomplete measurement information.
Learning Efficient Robotic Garment Manipulation with Standardization
zhou changshi, Bin He (Tongji University)
Robotic IntelligenceReinforcement LearningImage
🎯 What it does: This paper proposes APS-Net, a multi-primal action selection network that integrates two primitives: dual-arm dynamic fling and pick-and-place (p&p). It enables garment unfolding and standardization in a single grasping action. Additionally, it presents a factorized reward function based on 2D coverage, IoU, and keypoint distance, along with a spatial action mask and action optimization module, significantly improving unfolding efficiency and standardization quality.
Learning Event Completeness for Weakly Supervised Video Anomaly Detection
Yu Wang (Tongji University), Shiwei Chen (Microsoft Asia-Pacific Technology Co Ltd)
Anomaly DetectionGraph Neural NetworkTransformerContrastive LearningVideo
🎯 What it does: A weakly supervised video anomaly detection method LEC-VAD is proposed, which can learn complete event boundaries with only video-level labels.
Learning Extrapolative Sequence Transformations from Markov Chains
Sophia Hager (Johns Hopkins University), Nicholas Andrews (Johns Hopkins University)
GenerationData SynthesisDrug DiscoveryRecurrent Neural NetworkSupervised Fine-TuningTextSequentialBiomedical Data
🎯 What it does: A self-regressive model is trained using the state sequences sampled from the Markov chain generated by MCMC, achieving extrapolation generation of sequence properties with a small number of steps;
Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning
Lianbo Ma (Northeastern University), Zhichao Lu (City University of Hong Kong)
ClassificationObject DetectionOptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: An adaptive gradient alignment method based on loss curvature sharpness (ASGA) is proposed, which searches for transferable mixed-precision quantization strategies on a small-scale proxy dataset and directly uses them on a large-scale target dataset without the need for extensive fine-tuning.
Learning from others' mistakes: Finetuning machine translation models with span-level error annotations
Lily H Zhang, Markus Freitag (Google)
GenerationTransformerSupervised Fine-TuningText
🎯 What it does: In the machine translation task, fine-tuning the model using fine-grained span-level error annotations from offline data.
Learning from Sample Stability for Deep Clustering
Zhixin Li (Southeast University), Junhui Hou (City University of Hong Kong)
Representation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes to use sample stability as a supervisory signal to improve deep clustering models.
Learning from Suboptimal Data in Continuous Control via Auto-Regressive Soft Q-Network
Jijia Liu (Tsinghua University), Yu Wang (Tsinghua University)
Reinforcement LearningSequential
🎯 What it does: This paper proposes the Auto-Regressive Soft Q-Learning (ARSQ) algorithm for efficient training in continuous control tasks using suboptimal data (such as non-expert demonstrations or low-quality trajectories collected online).
Learning from True-False Labels via Multi-modal Prompt Retrieving
Zhongnian Li (China University of Mining and Technology), Xinzheng Xu (China University of Mining and Technology)
ClassificationRetrievalConvolutional Neural NetworkPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: A new weakly supervised labeling method (True-False Labels, TFL) and a corresponding risk-consistent learning framework are proposed, achieving effective transfer to pre-trained visual-language models (VLM) through multimodal convolutional prompt retrieval (MPR).
Learning Fused State Representations for Control from Multi-View Observations
Zeyu Wang (Beijing Institute of Technology), Riashat Islam (Mila - Quebec AI Institute)
Autonomous DrivingRepresentation LearningRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningMultimodality
🎯 What it does: A multi-view fusion state representation framework MFSC is proposed, which utilizes dual affine metric learning and self-attention fusion to generate task-relevant low-dimensional state representations, and enhances robustness to missing or disturbed views through multi-view occlusion and latent reconstruction auxiliary tasks.
Learning Gaussian DAG Models without Condition Number Bounds
Constantinos Costis Daskalakis, Rui Yao (Massachusetts Institute of Technology)
Graph
🎯 What it does: The study addresses the topology learning problem of high-dimensional linear Gaussian DAG models under the condition of homoscedasticity, providing new algorithms and a theoretically optimal sample complexity analysis.
Learning Imbalanced Data with Beneficial Label Noise
Guangzheng Hu (University of Melbourne), Liuhua Peng (University of Melbourne)
ClassificationImageTabular
🎯 What it does: A method called LNR is proposed to rebalance imbalanced data by introducing label noise, and it is validated in binary and multi-class tasks.
Learning Imperfect Information Extensive-form Games with Last-iterate Convergence under Bandit Feedback
Canzhe Zhao (Shanghai Jiao Tong University), Shuai Li (Shanghai Jiao Tong University)
Reinforcement Learning
🎯 What it does: An algorithm is proposed for learning the approximate Nash Equilibrium (NE) of two-player zero-sum incomplete information extensive-form games (IIEFG) under Bandit feedback, and its last-iterate convergence in finite time is proven.
Learning In-context $n$-grams with Transformers: Sub-$n$-grams Are Near-Stationary Points
Aditya Varre (École Polytechnique Fédérale de Lausanne), Nicolas Flammarion (École Polytechnique Fédérale de Lausanne)
TransformerSequential
🎯 What it does: This study investigates the loss landscape of the Transformer in learning in-context n-gram language models, revealing that the sub-n-gram estimator is an approximate steady state point under infinite sequence length and parameter scale, explaining the phase stagnation and mutations during the training process.
Learning Initial Basis Selection for Linear Programming via Duality-Inspired Tripartite Graph Representation and Comprehensive Supervision
Anqi Lu (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A GNN model based on a tripartite graph is proposed for learning the initial basis selection of linear programming, and the prediction quality is enhanced through comprehensive supervision.
Learning Input Encodings for Kernel-Optimal Implicit Neural Representations
Zhemin Li (National University of Defense Technology), Xiaolong Han (National University of Defense Technology)
RetrievalNeural Radiance FieldImage
🎯 What it does: A kernel alignment-based input encoding learning framework (PEAK) is proposed, which enhances the generalization ability of implicit neural representations (INR) by aligning the neural tangent kernel of infinitely wide networks with the theoretically optimal kernel.
Learning Invariant Causal Mechanism from Vision-Language Models
Zeen Song (Institute of Software Chinese Academy of Sciences), Wenwen Qiang (Institute of Software Chinese Academy of Sciences)
Domain AdaptationTransformerVision Language ModelContrastive LearningImageTextBenchmark
🎯 What it does: This paper proposes learning invariant causal mechanisms on the Vision-Language pre-trained model CLIP to enhance OOD generalization.
Learning Joint Interventional Effects from Single-Variable Interventions in Additive Models
Armin Kekić (Max Planck Institute for Intelligent Systems), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
Tabular
🎯 What it does: Using observational data and univariate intervention data, this study investigates how to estimate the causal effects of multivariate joint interventions, providing identifiability proofs and practical estimators.
Learning Latent Graph Structures and their Uncertainty
Alessandro Manenti (Universita della Svizzera italiana), Cesare Alippi (Politecnico di Milano)
Graph Neural NetworkGraph
🎯 What it does: For prediction tasks with unknown graph structures, a method is proposed that combines learning prediction models and latent graph distributions, utilizing distribution matching loss to achieve calibration of the latent graph and optimal point prediction.
Learning Likelihood-Free Reference Priors
Nicholas George Bishop (Oxford University), Michael J. Wooldridge (Oxford University)
Flow-based ModelTabular
🎯 What it does: This paper proposes several likelihood-free methods for learning reference priors in complex simulation models, achieved through variational approximation and information-theoretic estimation.
Learning Mean Field Control on Sparse Graphs
Christian Fabian (Technische Universität Darmstadt), Heinz Koeppl (Technische Universität Darmstadt)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A sparse graph equilibrium control (LWMFC) model based on local weak convergence is proposed, along with a scalable two-system approximation and reinforcement learning algorithm.
Learning Minimum-Size BDDs: Towards Efficient Exact Algorithms
Christian Komusiewicz (University of Jena), Luca Pascal Staus (University of Jena)
OptimizationTabularBenchmark
🎯 What it does: This paper proposes a branch-and-bound algorithm based on the Witness paradigm, called WitBDD, to find the exact solution for the minimum size Binary Decision Diagram (BDD).
Learning Mixtures of Experts with EM: A Mirror Descent Perspective
Quentin Fruytier (University of Texas), Sujay Sanghavi (University of Texas)
OptimizationMixture of ExpertsImage
🎯 What it does: This paper studies the Expectation-Maximization (EM) algorithm for the Mixture of Experts (MoE) model and proves its equivalence to the Mirror Descent iteration with KL divergence regularization. It further provides convergence theory for EM in the general MoE case, as well as in two special cases: linear experts and logistic experts.
Learning Monotonic Probabilities with a Generative Cost Model
Yongxiang Tang (Kuaishou), Peng Jiang
OptimizationAuto EncoderTabular
🎯 What it does: This paper proposes a generative framework that transforms monotonic probability modeling into latent cost variable modeling, constructing two models: the strictly monotonic Generative Cost Model (GCM) and the non-strictly monotonic Implicit GCM (IGCM);
Learning Multi-Level Features with Matryoshka Sparse Autoencoders
Bart Bussmann, Neel Nanda
Representation LearningAuto EncoderText
🎯 What it does: A new variant of sparse autoencoders is proposed—Matryoshka SAEs, which utilizes multi-layer nested dictionaries to simultaneously learn sparse features of different scales, thereby achieving multi-level, hierarchical feature representation.
Learning multivariate Gaussians with imperfect advice
Arnab Bhattacharyya (University of Warwick), Themis Gouleakis (Nanyang Technological University)
OptimizationTabular
🎯 What it does: The study investigates how to reduce the sample complexity of learning high-dimensional Gaussian distributions using imperfectly accurate predictive information.
Learning Optimal Multimodal Information Bottleneck Representations
Qilong Wu (Zhongnan University of Economics and Law), Xiaobo Sun (Emory University)
Anomaly DetectionRepresentation LearningAuto EncoderMultimodality
🎯 What it does: Proposes the OMIB framework, which learns optimal multimodal representations through the information bottleneck principle to address the issue of task-related information imbalance.
Learning Parametric Distributions from Samples and Preferences
Marc Jourdan (École Polytechnique Fédérale de Lausanne), Nicolas Flammarion (École Polytechnique Fédérale de Lausanne)
🎯 What it does: The study incorporates preference feedback into continuous parametric distribution estimation and compares the statistical performance of sample-only estimation with preference-assisted estimation.
Learning Policy Committees for Effective Personalization in MDPs with Diverse Tasks
Luise Ge (Washington University in St. Louis), Yevgeniy Vorobeychik (Washington University in St. Louis)
Meta LearningLarge Language ModelReinforcement Learning
🎯 What it does: The PACMAN framework is proposed to address the issues of negative transfer, insufficient generalization, and few-shot adaptation caused by task diversity in multi-task reinforcement learning through a Policy Committee.
Learning Progress Driven Multi-Agent Curriculum
Wenshuai Zhao (Aalto University), Joni Pajarinen (Aalto University)
Reinforcement LearningBenchmark
🎯 What it does: A multi-agent curriculum learning method based on learning progress is proposed, which automatically adjusts the number of agents to overcome the sparse reward problem.
Learning Representations of Instruments for Partial Identification of Treatment Effects
Jonas Schweisthal (Ludwig Maximilian University of Munich), Stefan Feuerriegel (Ludwig Maximilian University of Munich)
Representation LearningReinforcement Learning
🎯 What it does: A method is proposed for partially identifying causal effects using complex (high-dimensional, continuous) instrumental variables, providing upper and lower bounds for CATE under observable data.
Learning Robust Neural Processes with Risk-Averse Stochastic Optimization
Huafeng Liu (Beijing Jiaotong University), Jian Yu
OptimizationImageTabularStochastic Differential Equation
🎯 What it does: A Robust Neural Process is proposed, which achieves robustness in rapid adaptation performance to tasks by controlling tail risk (CVaR);
Learning Safe Control via On-the-Fly Bandit Exploration
Alexandre Capone (Carnegie Mellon University), Sandra Hirche (Technical University of Munich)
OptimizationRobotic IntelligenceReinforcement LearningTime Series
🎯 What it does: A safe control method utilizing Gaussian process UCB exploration is proposed when safe filtering is infeasible, ensuring system safety under conditions of high model uncertainty.
Learning Safe Strategies for Value Maximizing Buyers in Uniform Price Auctions
Negin Golrezaei (Massachusetts Institute of Technology), Sourav Sahoo (Massachusetts Institute of Technology)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies how value-maximizing buyers in repeated uniform price multi-unit auctions can achieve cumulative value maximization through a safe bidding strategy while satisfying the RoI constraints in each round, and presents a learnable algorithm.
Learning Safety Constraints for Large Language Models
Xin Chen (ETH Zurich), Andreas Krause (ETH Zurich)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes SaP—a Safety Polytope framework that learns safety constraints in the representation space of large language models and geometrically adjusts model activations during inference to achieve safe outputs.
Learning Single Index Models with Diffusion Priors
Anqi Tang (University of Electronic Science and Technology of China), Zhaoqiang Liu (University of Electronic Science and Technology of China)
RestorationDiffusion modelImage
🎯 What it does: Using diffusion models as a prior, efficient signal recovery of the single-index model (SIM) is achieved through reverse sampling starting from intermediate time points.
Learning Smooth and Expressive Interatomic Potentials for Physical Property Prediction
Xiang Fu (Meta), C. Lawrence Zitnick (Meta)
Graph Neural NetworkTabularPhysics Related
🎯 What it does: This study investigates the smoothness and conservativeness of machine learning atomic potential energy functions and proposes the energy-conserving eSEN model to enhance the predictive performance of physical properties.
Learning Soft Sparse Shapes for Efficient Time-Series Classification
Zhen Liu (South China University of Technology), Qianli Ma (South China University of Technology)
ClassificationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkMixture of ExpertsTime Series
🎯 What it does: Proposes the SoftShape model, which uses soft sparse shapes and Mixture-of-Experts to learn time series shapes, achieving efficient and interpretable time series classification.
Learning State-Based Node Representations from a Class Hierarchy for Fine-Grained Open-Set Detection
Spandan Pyakurel (Rochester Institute of Technology), Qi Yu (Rochester Institute of Technology)
ClassificationObject DetectionAnomaly DetectionRepresentation LearningTransformerPrompt EngineeringImage
🎯 What it does: A class-hierarchy-based state representation method is proposed for fine-grained open set detection (FGOD), which avoids the accumulation of local errors and improves the recognition of open set samples by learning global states.
Learning Strategic Language Agents in the Werewolf Game with Iterative Latent Space Policy Optimization
Zelai Xu (Tsinghua University), Yu Wang (Tsinghua University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes an Iterative Latent Space Policy Optimization (LSPO) framework for constructing agents that can perform excellently in strategic language games like 'Werewolf' with free-text dialogue. It combines game-theoretic methods (CFR) and LLM fine-tuning (DPO) to overcome issues of action distribution bias and limited exploration.
Learning Survival Distributions with the Asymmetric Laplace Distribution
Deming Sheng (Duke University), Ricardo Henao (Duke University)
TabularBiomedical DataFinance Related
🎯 What it does: A parameterized survival analysis model based on the Asymmetric Laplace Distribution (ALD) is proposed, which directly learns the distribution parameters for each individual through maximum likelihood, allowing for closed-form calculations of statistics such as mean, median, variance, and quantiles.
Learning the Electronic Hamiltonian of Large Atomic Structures
Chen Hao Xia (ETH Zurich), Mathieu Luisier (ETH Zurich)
Graph Neural NetworkGraphPhysics Related
🎯 What it does: A local reversible graph neural network specifically designed for large-scale, disordered atomic structures is proposed to predict the global electronic Hamiltonian matrix.
Learning the RoPEs: Better 2D and 3D Position Encodings with STRING
Connor Schenck (Google DeepMind), Krzysztof Marcin Choromanski
ClassificationObject DetectionRetrievalRobotic IntelligenceTransformerImage
🎯 What it does: A new separable translation-invariant positional encoding method, STRING, is proposed and its performance is systematically validated across various tasks including image classification, retrieval, open vocabulary object detection, and robotic control.
Learning Time-Aware Causal Representation for Model Generalization in Evolving Domains
Zhuo He (Independent Researcher), Kun Gai (Kuaishou Technology)
Domain AdaptationRepresentation LearningRecurrent Neural NetworkAuto EncoderContrastive LearningTime SeriesSequential
🎯 What it does: This paper proposes a time-aware causal representation learning method called SYNC for the evolving domain generalization (EDG) task. It utilizes a Structural Causal Model (SCM) and a Variational Autoencoder (VAE) framework to learn interpretable static and dynamic causal features, while suppressing spurious associations through information-theoretic objectives, thereby improving the model's generalization performance on future domains.
Learning Time-Varying Multi-Region Brain Communications via Scalable Markovian Gaussian Processes
Weihan Li (Georgia Institute of Technology), Anqi Wu (Georgia Institute of Technology)
Gaussian SplattingTime SeriesBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This study investigates the time-varying characteristics of multi-regional brain communication and proposes the Adaptive Delay Model (ADM) that combines Markovian Gaussian Process with State-Space Model, utilizing parallel scanning Kalman EM to achieve O(log T) level inference;
Learning to (Learn at Test Time): RNNs with Expressive Hidden States
Yu Sun (Stanford University), Carlos Guestrin (Stanford University)
Recurrent Neural NetworkText
🎯 What it does: A Test-Time Training (TTT) RNN layer is proposed, making the hidden state itself a learnable model, which is updated through self-supervised learning during inference;
Learning to Generate Projections for Reducing Dimensionality of Heterogeneous Linear Programming Problems
Tomoharu Iwata (NTT Corporation), Shinsaku Sakaue (CyberAgent)
OptimizationTabular
🎯 What it does: An adaptive projection method based on neural networks is proposed, which generates specific projection matrices for each linear programming instance to reduce dimensionality and quickly obtain high-quality feasible solutions.
Learning to Incentivize in Repeated Principal-Agent Problems with Adversarial Agent Arrivals
Junyan Liu (University of Washington), Lillian J. Ratliff (University of Washington)
Reinforcement Learning
🎯 What it does: This study explores the repeated principal-agent problem where the principal interacts with various types of agents in an adversarial sequence within a limited time. In each round, the principal selects an incentive to influence the unknown types of agents, who choose actions based on their own utility and the incentives, while the principal receives corresponding rewards.
Learning to Keep a Promise: Scaling Language Model Decoding Parallelism with Learned Asynchronous Decoding
Tian Jin (MIT CSAIL), Michael Carbin (MIT CSAIL)
GenerationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study proposes a system called PASTA that enables large language models to learn to identify semantically independent text blocks during the generation process and achieve asynchronous parallel decoding through a custom annotation language, significantly reducing inference latency.
Learning to Match Unpaired Data with Minimum Entropy Coupling
Mustapha BOUNOUA, Pietro Michiardi (EURECOM)
Image TranslationData SynthesisOptimizationDiffusion modelMultimodalityBiomedical Data
🎯 What it does: A continuous data pairing method based on Minimum Entropy Coupling (MEC) called DDMEC is proposed, which uses two collaborative diffusion models to approximate the joint distribution, achieving coupling and alignment of unpaired multimodal data.
Learning to Plan & Reason for Evaluation with Thinking-LLM-as-a-Judge
Swarnadeep Saha (Meta), Tianlu Wang
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes EvalPlanner, an LLM-as-a-Judge model that evaluates LLM outputs by first generating a problem-solving plan and then executing the plan;
Learning to Quantize for Training Vector-Quantized Networks
Peijia Qin (Southern University of Science and Technology), Jianguo Zhang (Southern University of Science and Technology)
GenerationOptimizationMeta LearningAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a meta-learning-based two-layer optimization framework—Meta-Quantization, which re-parameterizes the codebook using a hyper-net and learns it within vector quantization networks to achieve more efficient and task-related quantization.
Learning to Reuse Policies in State Evolvable Environments
Ziqian Zhang (Nanjing University), Yang Yu (Nanjing University)
Reinforcement LearningGenerative Adversarial NetworkSequential
🎯 What it does: In an environment where the state space evolves over time, this study investigates how to quickly adapt existing RL strategies;
Learning to Route LLMs with Confidence Tokens
Yu-Neng Chuang (Rice University), Helen Zhou (Apple)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Introducing the self-reflection and error feedback fine-tuning strategy Self-REF in LLM models allows the model to use dedicated confidence words to mark whether its answers are trustworthy.
Learning to Steer Learners in Games
Yizhou Zhang (California Institute of Technology), Eric Mazumdar (California Institute of Technology)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: This paper studies how to learn and guide an opponent using a no-regret learning algorithm in a two-player finite action repeated game to reach the Stackelberg equilibrium, exploring the feasibility and necessary conditions when the opponent's payoff matrix is unknown.
Learning to Stop: Deep Learning for Mean Field Optimal Stopping
Lorenzo Magnino (New York University Shanghai), Mathieu Lauriere (New York University Shanghai)
OptimizationReinforcement LearningSequential
🎯 What it does: Proposes and solves the mean field approximation (MFOS) for the multi-agent optimal stopping (MAOS) problem in discrete time and finite state space, providing theoretical guarantees and numerical methods;
Learning to Trust Bellman Updates: Selective State-Adaptive Regularization for Offline RL
Qin-Wen Luo (Nanjing University of Aeronautics and Astronautics), Sheng-Jun Huang (Nanjing University of Aeronautics and Astronautics)
Reinforcement Learning
🎯 What it does: A selective state adaptive regularization method is proposed to enhance the reliability of Bellman updates in offline reinforcement learning.
Learning Utilities from Demonstrations in Markov Decision Processes
Filippo Lazzati (Politecnico di Milano), Alberto Maria Metelli (Politecnico di Milano)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: A risk-sensitive IRL model is proposed, utilizing a utility function to describe the behavior of non-Markov decision-makers and defining the utility learning problem.
Learning Vision and Language Concepts for Controllable Image Generation
Shaoan Xie (Carnegie Mellon University), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImageTextMultimodality
🎯 What it does: This paper proposes a theoretical framework for learning and aligning atomic visual and textual concepts from multimodal distributions, and based on this, designs a controllable text-to-image generation model called ConceptAligner.
Learning with Exact Invariances in Polynomial Time
Ashkan Soleymani (Massachusetts Institute of Technology), Patrick Jaillet (Massachusetts Institute of Technology)
Tabular
🎯 What it does: A polynomial-time kernel regression algorithm has been developed that can obtain precise invariance estimators under a given finite group action.
Learning with Expected Signatures: Theory and Applications
Lorenzo Lucchese (Imperial College London), Almut E. D. Veraart (Imperial College London)
TabularTime SeriesSequentialFinance RelatedStochastic Differential Equation
🎯 What it does: This paper proposes a learning method based on expected signatures, aimed at mapping data streams to low-dimensional representations, and connects the theoretical results of discrete-time estimates of expected signatures with the gap between continuous-time values.
Learning With Multi-Group Guarantees For Clusterable Subpopulations
Jessica Dai (University of California Berkeley), Eric Zhao (University of California Berkeley)
🎯 What it does: A multi-group guarantee learning framework under sub-group clusterable distribution is proposed, along with a corresponding online calibration algorithm.
Learning with Selectively Labeled Data from Multiple Decision-makers
Jian Chen (Tsinghua University), Xiaojie Mao (Tsinghua University)
ClassificationData-Centric LearningTabularFinance Related
🎯 What it does: This study investigates the selective labeling problem in a multi-decision-maker environment, establishes a framework for instrumental variables, and proposes a unified cost-sensitive learning method to train classifiers that are robust to selection bias.
Learning without Isolation: Pathway Protection for Continual Learning
Zhikang Chen (Tsinghua University), Tianling Ren (Tsinghua University)
Knowledge DistillationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: A channel path protection framework LwI based on graph matching is proposed for exemplar-free continual learning to avoid catastrophic forgetting.
Learning-Augmented Algorithms for MTS with Bandit Access to Multiple Predictors
Matei Gabriel Cosa (Bocconi University), Marek Elias (Bocconi University)
Reinforcement Learning
🎯 What it does: This paper studies how to combine the suggestions of multiple predictors to solve the Metric Task System (MTS) problem under the condition of being able to query the predictor only once (m-delay bandit access).
Learning-Augmented Hierarchical Clustering
Vladimir Braverman (Johns Hopkins University), Samson Zhou (Texas A&M University)
Optimization
🎯 What it does: A hierarchical clustering algorithm is proposed within a learning-enhanced framework, utilizing a triplet-based 'splitting oracle' to obtain a structure close to the optimal tree, thereby achieving approximately optimal hierarchical clustering trees under various objective functions.
Learning-Order Autoregressive Models with Application to Molecular Graph Generation
Zhe Wang (Google DeepMind), Michalis Titsias
GenerationData SynthesisDrug DiscoveryGraph Neural NetworkTransformerReinforcement LearningGraph
🎯 What it does: A learnable autoregressive model LO-ARM is proposed for generating sequences, suitable for high-dimensional data without a natural order (such as molecular graphs).
Learnings from Scaling Visual Tokenizers for Reconstruction and Generation
Philippe Hansen-Estruch (University of Texas at Austin), Xinlei Chen (Meta)
RestorationGenerationTransformerAuto EncoderGenerative Adversarial NetworkImageVideo
🎯 What it does: Proposes ViTok, an autoencoder based on Vision Transformer, for image and video reconstruction and generation;
Learnware Specification via Dual Alignment
Wei Chen (Southeast University), Min-Ling Zhang (Southeast University)
Domain AdaptationImage
🎯 What it does: A dual alignment method (Discriminative Alignment and Distribution Alignment) called DALI is proposed to generate learnware specifications, enhancing model reuse effectiveness.
Lego Sketch: A Scalable Memory-augmented Neural Network for Sketching Data Streams
Yuan Feng (University of Science and Technology of China), S Kevin Zhou
Computational EfficiencyMeta LearningTime Series
🎯 What it does: This paper proposes Lego Sketch, a scalable memory-augmented neural network architecture for frequency estimation of infinite data streams within limited space.
LEMoN: Label Error Detection using Multimodal Neighbors
Haoran Zhang (Massachusetts Institute of Technology), Marzyeh Ghassemi (Massachusetts Institute of Technology)
ClassificationAnomaly DetectionTransformerContrastive LearningImageTextMultimodality
🎯 What it does: A label error detection method based on multimodal neighbors, LEMON, is proposed, which utilizes a multimodal model pre-trained with contrastive learning (such as CLIP) to embed image-text pairs, combining multimodal similarity and the nearest neighbor distances of images/texts to calculate mislabeling scores.
LensLLM: Unveiling Fine-Tuning Dynamics for LLM Selection
Xinyue Zeng (Virginia Tech), Dawei Zhou (Virginia Tech)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the LENSLLM framework, using PAC-Bayesian generalization bounds and NTK to model the dynamics of LLM fine-tuning, providing an interpretable model selection method.
Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective
Lele Fu (Sun Yat-sen University), Chuan Chen (Sun Yat-sen University)
Federated LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes the FedATH method, which uses an edge evaluator to partition each client's graph into causal subgraphs and bias subgraphs, uploading only the GNN parameters of the causal subgraphs to eliminate topological heterogeneity in federated graph learning.
Let LLM Tell What to Prune and How Much to Prune
Mingzhe Yang (University of Science and Technology of China), Xiaojun Chang (University of Science and Technology of China)
Large Language ModelText
🎯 What it does: Proposes a structured pruning framework with a dynamic ratio of multi-structure units.
LETS Forecast: Learning Embedology for Time Series Forecasting
Abrar Majeedi (University of Wisconsin Madison), Yin Li (University of Wisconsin Madison)
Anomaly DetectionTime Series
🎯 What it does: The DeepEDM framework is proposed, which combines time delay embedding with deep networks for predicting nonlinear time series.
Leveraging Diffusion Model as Pseudo-Anomalous Graph Generator for Graph-Level Anomaly Detection
Jinyu Cai (National University of Singapore), See-Kiong Ng (National University of Singapore)
Anomaly DetectionGraph Neural NetworkDiffusion modelAuto EncoderGraphBiomedical Data
🎯 What it does: The AGDiff framework is proposed, which utilizes latent diffusion models to generate pseudo-anomalous graphs that closely resemble normal graphs but contain subtle perturbations, and is jointly trained with GNN-level anomaly detectors.
Leveraging Model Guidance to Extract Training Data from Personalized Diffusion Models
Xiaoyu Wu (Carnegie Mellon University), Steven Wu
GenerationData SynthesisDiffusion modelScore-based ModelImage
🎯 What it does: This work proposes the FineXtract framework, which utilizes the gradual transfer of the distribution of pre-trained and fine-tuned Diffusion models to guide the training data distribution, and extracts fine-tuned training set images through clustering;
Leveraging Offline Data in Linear Latent Contextual Bandits
Chinmaya Kausik (University of Michigan), Ambuj Tewari (University of Michigan)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: This work proposes a complete offline-online end-to-end method that utilizes offline data to estimate the low-dimensional subspace of linear potential contextual bandits (SOLD), and based on this, designs the optimal online algorithm LOCAL-UCB and the feasible algorithm ProBALL-UCB, completing the proof of theoretical upper bounds and matching lower bounds, and providing a proof of the de Finetti theorem to demonstrate the generality of potential bandits.
Leveraging Online Olympiad-Level Math Problems for LLMs Training and Contamination-Resistant Evaluation
Sadegh Mahdavi (University of British Columbia), Renjie Liao (Vector Institute for AI)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Extracted Olympiad-level math Q&A pairs from the AoPS forum to construct a training set of 652K QA pairs called AoPS-Instruct and generated a real-time, contamination-resistant evaluation set called LiveAoPSBench;
Leveraging Per-Instance Privacy for Machine Unlearning
Nazanin Mohammadi Sepahvand (McGill University), Gintare Karolina Dziugaite (Google)
OptimizationSafty and PrivacyConvolutional Neural NetworkTransformerImageStochastic Differential Equation
🎯 What it does: This study investigates the privacy loss calculated per instance in machine unlearning and uses it to measure and predict the unlearning difficulty of each data point.
Leveraging Predictive Equivalence in Decision Trees
Hayden McTavish (Duke University), Cynthia Rudin (Duke University)
OptimizationExplainability and InterpretabilityReinforcement LearningTabular
🎯 What it does: Transform decision trees into Boolean logic expressions (DNF) to eliminate predictive equivalence and apply it in multiple machine learning tasks.
Leveraging Randomness in Model and Data Partitioning for Privacy Amplification
Andy Dong (Stanford University), Ayfer Ozgur (Stanford University)
Federated LearningSafty and PrivacyConvolutional Neural NetworkImage
🎯 What it does: This study explores how to leverage the inherent randomness during the training process to enhance privacy protection, including model partitioning and data partitioning.
Leveraging Skills from Unlabeled Prior Data for Efficient Online Exploration
Max Wilcoxson (University of California), Sergey Levine (University of California)
Robotic IntelligenceReinforcement Learning
🎯 What it does: Utilize unlabeled offline trajectory data to pre-train low-level skills and use these trajectories as additional offline data through pseudo-labeling in the online phase to learn high-level policies for efficient exploration of sparse reward tasks.
Leveraging Sparsity for Sample-Efficient Preference Learning: A Theoretical Perspective
Yunzhen Yao (École Polytechnique Fédérale de Lausanne), Michael Gastpar (École Polytechnique Fédérale de Lausanne)
Recommendation SystemOptimizationTabular
🎯 What it does: This study investigates the sample efficiency of sparse random utility models in preference learning, providing lower and upper bounds, and proposes a computable ℓ1 regularization estimator.
LEVIS: Large Exact Verifiable Input Spaces for Neural Networks
Mohamad Fares El Hajj Chehade (University of Texas at Austin), Hao Zhu (University of Texas at Austin)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the LEVIS framework for identifying the largest spherical subset of the verifiable input space of neural networks.
Lexico: Extreme KV Cache Compression via Sparse Coding over Universal Dictionaries
Junhyuck Kim (Krafton), Dimitris Papailiopoulos (University of Wisconsin-Madison)
CompressionTransformerText
🎯 What it does: Utilizing sparse coding and a universal dictionary to compress the KV cache of the Transformer, significantly reducing memory usage while maintaining model generation performance.
LGDM: Latent Guidance in Diffusion Models for Perceptual Evaluations
Shreshth Saini (University of Texas at Austin), Alan Bovik
Diffusion modelImage
🎯 What it does: A framework for no-reference image quality assessment (NR-IQA) using a pre-trained Latent Diffusion Model (LDM) is proposed, called LGDM. It extracts multi-scale, multi-time-step diffusion hyperfeatures through Perceptual Manifold Guidance (PMG) in the latent space and predicts quality scores using a lightweight regression network.
LieRE: Lie Rotational Positional Encodings
Sophie Ostmeier (Stanford University), Curtis Langlotz (Stanford University)
ClassificationRecognitionTransformerImageVideo
🎯 What it does: In the attention mechanism of the Transformer, a new positional encoding method called LieRE (Lie Rotational Positional Encodings) is proposed, which learns dense antisymmetric matrices and maps them to high-dimensional rotation matrices, directly acting on keys and queries to encode relative and absolute positional information in multi-dimensional spaces (2D, 3D).
LIFT the Veil for the Truth: Principal Weights Emerge after Rank Reduction for Reasoning-Focused Supervised Fine-Tuning
Zihang Liu (University of California), Shiwei Liu (University of Oxford)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A sparse fine-tuning method called LIFT is proposed, which updates only the maximum modulus parameters obtained after low-rank approximation, thereby achieving efficient inference fine-tuning on large models.
Liger: Linearizing Large Language Models to Gated Recurrent Structures
Disen Lan (Shanghai AI Laboratory), Yu Cheng (Chinese University of Hong Kong)
Recurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Construct a gating mechanism by projecting keys from a pre-trained Transformer LLM and rewrite it into a linear recursive structure to achieve model linearization.
LightGTS: A Lightweight General Time Series Forecasting Model
Yihang Wang (East China Normal University), Chenjuan Guo (East China Normal University)
TransformerTime SeriesFinance Related
🎯 What it does: A lightweight time series forecasting model, LightGTS, is proposed, which achieves unified modeling and forecasting of multi-source, different scale time series through periodical tokenization and periodical parallel decoding.
LightningDrag: Lightning Fast and Accurate Drag-based Image Editing Emerging from Videos
Yujun Shi (National University of Singapore), Jiashi Feng (ByteDance Inc)
Image TranslationGenerationOptimizationDiffusion modelOptical FlowImageVideo
🎯 What it does: A drag-and-drop image editing method that utilizes a video-trained conditional generative model to achieve approximately 1 second of editing time is proposed.
Lightspeed Geometric Dataset Distance via Sliced Optimal Transport
Khai Nguyen (University of Texas at Austin), Nhat Ho
Computational EfficiencyData-Centric LearningImageText
🎯 What it does: A training-free, embedding-free dataset distance based on sliced optimal transport (s-OTDD) is proposed, which maps the label distribution to a scalar through Moment Transform Projection, and then maps the entire dataset to a one-dimensional distribution to compute the expected Wasserstein distance.
Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty
Yeseul Cho (KAIST), Chulhee Yun (KAIST)
ClassificationData-Centric LearningImage
🎯 What it does: A lightweight dataset pruning method based on training dynamics is proposed, capable of identifying important samples without complete training.
Lightweight Online Adaption for Time Series Foundation Model Forecasts
Thomas L Lee, Martin Asenov (Huawei)
Time SeriesBenchmark
🎯 What it does: A lightweight online adaptive method called ELF is proposed to enhance the predictive performance of time series foundational models (FM) through online feedback during the deployment phase.