NeurIPS 2025 Papers — Page 26
Conference on Neural Information Processing Systems · 5275 papers
Learning normalized image densities via dual score matching
Florentin Guth (New York University Flatiron Institute Simons Foundation), Eero P Simoncelli
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelScore-based ModelImage
🎯 What it does: Train a network that can directly output normalized energy (log-likelihood) by learning the probability distribution of image data through dual score matching.
Learning Orthogonal Multi-Index Models: A Fine-Grained Information Exponent Analysis
Yunwei Ren (Princeton University), Jason D. Lee (Princeton University)
Optimization
🎯 What it does: This study explores the learning of multi-index models, particularly improving the analysis of sample complexity by considering both low-order and high-order terms.
Learning Parameterized Skills from Demonstrations
Vedant Gupta (Brown University), George Konidaris (Brown University)
Robotic IntelligenceReinforcement Learning from Human FeedbackRecurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: The DEPS algorithm is proposed, which learns parameterized skills and their parameters from expert demonstrations in an end-to-end manner, and constructs a three-layer hierarchical policy (discrete skill selection, continuous parameter selection, low-level action sub-policy).
Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift
Yanru Sun (Tianjin University), Min Wu (Astar Research Institute)
Domain AdaptationTransformerMixture of ExpertsTime Series
🎯 What it does: A TFPS (Time-Frequency Pattern Specific) framework is proposed, utilizing dual-domain encoding, subspace clustering pattern recognition, and a mixture of experts network to achieve dynamic expert allocation for time series patches, resulting in more accurate long-term predictions.
Learning Personalized Ad Impact via Contextual Reinforcement Learning under Delayed Rewards
Yuwei Cheng (University of Chicago), Haifeng Xu (University of Chicago)
Recommendation SystemOptimizationReinforcement LearningSequential
🎯 What it does: This paper studies the adaptive bidding problem in online advertising, modeling it as a contextual Markov decision process (CMDP) with delayed Poisson rewards, and proposes a two-stage maximum likelihood estimator and an approximate optimal learning algorithm based on this estimator.
Learning Preferences without Interaction for Cooperative AI: A Hybrid Offline-Online Approach
Haitong Ma (Shanghai Jiao Tong University), Shuai Li (Shanghai Jiao Tong University)
Recommendation SystemRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: A three-stage hybrid online-offline reinforcement learning framework has been designed and validated, enabling cooperative agents to learn human preferences solely based on offline user preference data while maintaining generalization capabilities with unknown partners.
Learning Provably Improves the Convergence of Gradient Descent
Qingyu Song (Xiamen University), Hong Xu (Chinese University of Hong Kong)
OptimizationTabular
🎯 What it does: This paper proposes a Learning to Optimize (L2O) framework that accelerates the optimization process by learning the hyperparameters of gradient descent (GD) and proves the training convergence of this framework.
Learning quadratic neural networks in high dimensions: SGD dynamics and scaling laws
Gerard Ben Arous, Denny Wu (New York University)
Stochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The study investigates the gradient descent learning dynamics and sample complexity of two-layer quadratic activation neural networks in high dimensions, where the width expands with the dimension, and proposes a rigorous risk decay and scaling law.
Learning Reconfigurable Representations for Multimodal Federated Learning with Missing Data
Manh Duong Nguyen, Phi Le Nguyen (Hanoi University of Science and Technology)
Federated LearningContrastive LearningMultimodalityBiomedical DataElectrocardiogram
🎯 What it does: In the multimodal federated learning scenario, the PEPSY framework is proposed, which constructs data missing profiles by learning the missing pattern features of each client, to reconfigure local models so that they can produce complete and aligned representations even in environments where missing modalities and missing features coexist.
Learning Relative Gene Expression Trends from Pathology Images in Spatial Transcriptomics
Kazuya Nishimura (National Cancer Center Japan), Yasuhiro Kojima (National Cancer Center Japan)
Convolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: The paper proposes a gene expression relative trend learning method based on pathological images called STRank, which improves the robustness of traditional absolute value estimation.
Learning Repetition-Invariant Representations for Polymer Informatics
Yihan Zhu (University of Notre Dame), Meng Jiang (University of Notre Dame)
Representation LearningGraph Neural NetworkTabular
🎯 What it does: This study investigates how to learn invariant representations of polymer repeating units through graph neural networks, proposing the GRIN method.
Learning Robust Spectral Dynamics for Temporal Domain Generalization
En Yu (Australian Artificial Intelligence Institute University of Technology Sydney), Zhen Fang (Australian Artificial Intelligence Institute University of Technology Sydney)
Domain AdaptationAuto EncoderTime Series
🎯 What it does: This paper studies the concept drift problem in Time Domain Generalization (TDG) and proposes the FreKoo framework, which achieves robust generalization to future unseen domains by decomposing model parameter trajectories in the frequency domain, using Koopman operators to predict low-frequency trends, and regularizing high-frequency noise.
Learning Robust Vision-Language Models from Natural Latent Spaces
Zhangyun Wang (University of Auckland), Aniket Mahanti (University of Auckland)
ClassificationRecognitionDomain AdaptationAdversarial AttackPrompt EngineeringVision Language ModelImage
🎯 What it does: The CoAPT (Collaborative Adversarial Prompt Tuning) method is proposed, which combines pre-trained VLMs with target robust VLMs to enhance adversarial robustness while maintaining natural generalization.
Learning Shared Representations from Unpaired Data
Amitai Yacobi (Bar-Ilan University), Uri Shaham (Bar-Ilan University)
GenerationRetrievalRepresentation LearningContrastive LearningImageTextMultimodalityTabular
🎯 What it does: This paper proposes a framework named SUE, which constructs cross-modal shared representations using almost completely unpaired unimodal data through spectral embedding, CCA, and MMD networks.
Learning Simple Interpolants for Linear Integer Arithmetic
Minchao Wu (University of Tokyo), Naoki Kobayashi (University of Tokyo)
Graph Neural NetworkGraph
🎯 What it does: A lightweight learning framework is proposed for generating concise Craig interpolation clauses for Linear Integer Arithmetic (LIA) problems.
Learning single index models via harmonic decomposition
Nirmit Joshi (Toyota Technological Institute at Chicago), Nathan Srebro (Toyota Technological Institute at Chicago)
🎯 What it does: This paper studies the problem of learning single index models (SIM) under spherical symmetric input distributions, proposing a new perspective based on spherical harmonic decomposition, and constructs two types of estimators that can achieve sample optimal or runtime optimal, along with corresponding lower bounds.
Learning Skill-Attributes for Transferable Assessment in Video
Kumar Ashutosh (University of Texas at Austin), Kristen Grauman (University of Texas at Austin)
ClassificationGenerationRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodality
🎯 What it does: This paper proposes a cross-sport transferable video skill assessment framework—CROSSTRAINER, which can provide actionable feedback and proficiency assessment for new videos by learning general 'skill attributes'.
Learning Source-Free Domain Adaptation for Visible-Infrared Person Re-Identification
Yongxiang Li (Sichuan University), Peng Hu (Sichuan University)
RecognitionDomain AdaptationContrastive LearningImage
🎯 What it does: A source-agnostic domain adaptation framework SVIP is proposed for visible-infrared person recognition, which transfers a pre-trained source model to an unlabeled, unpaired target domain.
Learning Sparse Approximate Inverse Preconditioners for Conjugate Gradient Solvers on GPUs
Zherui Yang (Stanford University), Ligang Liu (University of Science and Technology of China)
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: Study of a sparse approximate inverse preconditioner (SPAI) based on graph neural networks for accelerating conjugate gradient solving on GPUs.
Learning Spatial-Aware Manipulation Ordering
Yuxiang Yan (Fudan University), Jian Pu (Fudan University)
Knowledge DistillationRobotic IntelligenceGraph Neural NetworkTransformerVision Language ModelMultimodalityBenchmark
🎯 What it does: In a crowded environment, a unified OrderMind framework is proposed for learning the operational sequence of spatial perception.
Learning Stochastic Multiscale Models
Andrew Francesco Ilersich, Prasanth B. Nair (Institute for Aerospace Studies University of Toronto)
Mixture of ExpertsPhysics RelatedStochastic Differential Equation
🎯 What it does: By constructing macro and micro layers of hidden states and modeling them in the form of coupled stochastic differential equations (SDEs), a probabilistic model capable of directly predicting multi-scale dynamics from observational data has been developed.
Learning Task-Agnostic Representations through Multi-Teacher Distillation
Philippe Formont (University of Paris-Saclay), Pablo Piantanida (National Center for Scientific Research)
Knowledge DistillationRepresentation LearningTransformerTextMultimodalityGraph
🎯 What it does: This paper proposes a task-agnostic multi-teacher distillation framework that learns general representations by maximizing the mutual information between the student and the embeddings of multiple teachers.
Learning Temporal 3D Semantic Scene Completion via Optical Flow Guidance
Meng Wang, Kenli Li
SegmentationAutonomous DrivingTransformerOptical FlowPoint Cloud
🎯 What it does: This paper proposes a flow-guided temporal 3D semantic scene completion method called FlowScene, which utilizes motion information from historical frames to align and aggregate features for the current frame, and refines them in voxel space with occlusion guidance.
Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks
Guobin Shen (Beijing Institute of AI Safety and Governance), Yi Zeng (Beijing Institute of AI Safety and Governance)
Spiking Neural NetworkReinforcement LearningSequential
🎯 What it does: A Plasticity-Driven Learning Framework (PDLF) is proposed, centered on learnable synaptic plasticity rules, and validated in working memory, multi-task reinforcement learning, and robustness testing.
Learning the Wrong Lessons: Syntactic-Domain Spurious Correlations in Language Models
Chantal Shaib (Northeastern University), Marzyeh Ghassemi (Massachusetts Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This study investigates the erroneous associations between grammatical templates and knowledge domains in the instruction-following process of large language models, revealing that models often rely on shallow grammatical-domain associations rather than true semantic understanding.
Learning Theory for Kernel Bilevel Optimization
Fares El Khoury (Universite Grenoble Alpes), Michael Arbel (Universite Grenoble Alpes)
OptimizationTabular
🎯 What it does: This paper studies the learning theory of Kernel Bilevel Optimization (KBO), providing a unified generalization error bound for value functions and gradients under finite samples, and verifying the convergence of gradient descent on synthetic instrumental variable regression tasks;
Learning to Add, Multiply, and Execute Algorithmic Instructions Exactly with Neural Networks
Artur Back de Luca (University of Waterloo), Kimon Fountoulakis (University of Waterloo)
🎯 What it does: This study investigates how a two-layer fully connected network can accurately execute binary encoding algorithm instructions (such as shifting, addition, multiplication, and SBN instructions) through gradient descent at infinite width.
Learning to Better Search with Language Models via Guided Reinforced Self-Training
Seungyong Moon (Seoul National University), Hyun Oh Song (Seoul National University)
Supervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: A Guided-ReST algorithm is proposed, which guides the search by inserting sub-goals of the optimal solution into the self-generated search trajectory, and fine-tunes using PPO based on this.
Learning to Clean: Reinforcement Learning for Noisy Label Correction
Marzi Heidari (Carleton University), Yuhong Guo (Carleton University)
Reinforcement LearningImage
🎯 What it does: Modeling the problem of correcting noisy labels as a reinforcement learning task, using a policy network for recursive label correction.
Learning to cluster neuronal function
Nina S. Nellen (University Göttingen), Alexander S. Ecker (Max Planck Institute for Dynamics and Self-Organization)
Representation LearningBiomedical Data
🎯 What it does: This study proposes a new deep embedding clustering method called DECEMber, aimed at improving the clustering consistency of neuronal functional embeddings by introducing clustering bias, thereby better identifying cell types in the mouse visual cortex.
Learning to Condition: A Neural Heuristic for Scalable MPE Inference
Brij Malhotra (University of Texas at Dallas), Vibhav Giridhar Gogate
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper studies an adaptive variable conditioning method based on neural networks, L2C, to accelerate MPE inference in probabilistic graphical models.
Learning to Control Free-Form Soft Swimmers
Changyu Hu (Tsinghua University), Tao Du (Tsinghua University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: Learning to control arbitrary-shaped soft underwater swimmers autonomously
Learning to Factorize Spatio-Temporal Foundation Models
Siru Zhong (Hong Kong University of Science and Technology), Yuxuan Liang (Hong Kong University of Science and Technology)
Domain AdaptationOptimizationTransformerSupervised Fine-TuningTime SeriesSequential
🎯 What it does: A two-stage spatial-temporal foundational model (FactoST) is proposed, which first learns general temporal knowledge through unsupervised learning on multi-domain data, and then quickly incorporates domain-specific spatial and temporal information through lightweight adapters to complete cross-domain predictions.
Learning to Flow from Generative Pretext Tasks for Neural Architecture Encoding
Sunwoo Kim (Korea Advanced Institute of Science and Technology), Kijung Shin (Korea Advanced Institute of Science and Technology)
Neural Architecture SearchGraph Neural NetworkGenerative Adversarial NetworkGraph
🎯 What it does: A new generative pre-training method FGP is proposed to enable ordinary neural network architecture encoders to learn the 'information flow' features of neural networks, thereby enhancing performance prediction and NAS effectiveness.
Learning to Focus: Causal Attention Distillation via Gradient‐Guided Token Pruning
Yiju Guo (Renmin University of China), Yankai Lin (Renmin University of China)
Knowledge DistillationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A two-stage framework named Learning to Focus (LeaF) is proposed, which automatically identifies and removes distracting words through gradient comparison, achieving causal attention alignment during the distillation process, thereby enhancing the ability of large language models to focus on key information and improve reasoning accuracy in long-context reasoning, code generation, and multi-hop question answering.
Learning to Generalize: An Information Perspective on Neural Processes
Hui Li (Beijing Jiaotong University), Liping Jing (Beijing Jiaotong University)
OptimizationMeta LearningConvolutional Neural NetworkTransformerImageTabular
🎯 What it does: A framework for generalization analysis of neural processes (NP) based on information theory is proposed, introducing dynamic stability regularization and noise injection parameter updates to enhance generalization performance.
Learning to Generate Human-Human-Object Interactions from Textual Descriptions
Jeonghyeon Na (Seoul National University), Hanbyul Joo (Seoul National University)
GenerationData SynthesisPose EstimationDiffusion modelScore-based ModelImageVideoOrdinary Differential Equation
🎯 What it does: A Human-Human-Object Interaction (HHOI) generation framework based on a score diffusion model is proposed, capable of synthesizing collaborative actions of multiple human bodies and objects according to text descriptions.
Learning to Insert for Constructive Neural Vehicle Routing Solver
Fu Luo (Southern University of Science and Technology), Qingfu Zhang (City University of Hong Kong)
OptimizationTransformerSupervised Fine-TuningTabular
🎯 What it does: A neural combinatorial optimization framework L2C-Insert based on the insertion paradigm is proposed to construct high-quality solutions for the vehicle routing problem (TSP and CVRP).
Learning to Instruct for Visual Instruction Tuning
Zhihan Zhou (Cooperative Medianet Innovation Center, Shanghai Jiao Tong University), Yanfeng Wang (School of Artificial Intelligence, Shanghai Jiao Tong University)
RecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: By allowing the multimodal LLM to simultaneously learn to generate instructions corresponding to images and responses to regularize the Visual Instruction Tuning (VIT) process.
Learning to Integrate Diffusion ODEs by Averaging the Derivatives
Wenze Liu (MMLab, Chinese University of Hong Kong), Xiangyu Yue (MMLab, Chinese University of Hong Kong)
GenerationData SynthesisTransformerDiffusion modelImageOrdinary Differential Equation
🎯 What it does: A diffusion ODE integration method based on secant losses is proposed, utilizing the secant function to learn the trajectory from noise to the target image for fast inference.
Learning to Learn with Contrastive Meta-Objective
Shiguang Wu (Tsinghua University), Quanming Yao (Tsinghua University)
Meta LearningContrastive LearningImage
🎯 What it does: This paper proposes a method called Contrastive Meta Learning (ConML) that incorporates contrastive meta objectives into meta-learning, enhancing the alignment and discrimination capabilities of meta-learners through task-level contrastive learning, and can be seamlessly integrated into various meta-learning algorithms and contextual learning frameworks.
Learning to Plan Like the Human Brain via Visuospatial Perception and Semantic-Episodic Synergistic Decision-Making
Tianyuan Jia (Beijing Normal University), Xia Wu (Beijing Institute of Technology)
OptimizationRobotic IntelligenceGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper proposes NeuroMP, a two-stage brain-inspired motion planning framework that combines a perceptual segmentation selector and a global alignment heuristic, significantly improving planning efficiency and path quality in high-dimensional continuous spaces.
Learning to price with resource constraints: from full information to machine-learned prices
Ruicheng Ao (Massachusetts Institute of Technology), David Simchi-Levi (Massachusetts Institute of Technology)
OptimizationTabularFinance Related
🎯 What it does: Three algorithms for dynamic pricing and resource constraints are proposed, suitable for scenarios with complete information, no information, and partial information.
Learning to Rank for In-Context Example Retrieval
Yuwen Ji (Zhejiang University), Yue Zhang (Westlake University)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A retrieval method based on learning to rank, SeDPO, is proposed, which uses LLM to generate probabilistic comparisons to train the retriever, thereby improving the example retrieval performance in In-Context Learning (ICL).
Learning to Reason under Off-Policy Guidance
Jianhao Yan (Zhejiang University), Yue Zhang (Westlake University)
Supervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposes the LUFFY framework, which incorporates offline (off-policy) reasoning trajectories into RLVR, achieving a dynamic balance between offline guidance and self-exploration.
Learning to Route: Per-Sample Adaptive Routing for Multimodal Multitask Prediction
Marzieh Ajirak (Weill Cornell Medicine), Logan Grosenick (Weill Cornell Medicine)
OptimizationExplainability and InterpretabilityMixture of ExpertsTextMultimodalityTabularElectronic Health Records
🎯 What it does: This paper proposes a unified adaptive routing framework for multimodal multitask prediction, which can dynamically select the modality processing path and task sharing strategy based on the input features of each sample;
Learning to Solve Complex Problems via Dataset Decomposition
Wanru Zhao (University of Cambridge), Alessandro Sordoni (Microsoft Research)
Large Language ModelSupervised Fine-TuningText
🎯 What it does: A framework called DECOMPX is proposed, which is based on recursively decomposing datasets using a teacher model. It utilizes large language models to automatically break down complex problems into a series of verifiable sub-problems and constructs a curriculum based on difficulty.
Learning to Specialize: Joint Gating-Expert Training for Adaptive MoEs in Decentralized Settings
Yehya Farhat (Rice University), Anastasios Kyrillidis (Rice University)
Federated LearningMixture of ExpertsImageText
🎯 What it does: In a decentralized federated learning environment, a joint training of gating networks and expert models is conducted to achieve adaptive Mixture-of-Experts (MoE) and propose the DDOME system.
Learning to Steer: Input-dependent Steering for Multimodal LLMs
Jayneel Parekh (Sorbonne Université), Matthieu Cord (Valeo.ai)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringImageTextMultimodality
🎯 What it does: This paper proposes a learnable input-dependent model-driven (L2S) approach to guide the behavior of multimodal large language models, enabling the model to automatically adjust its output under different inputs, primarily for safety constraints and reducing hallucinations.
Learning to Think: Information-Theoretic Reinforcement Fine-Tuning for LLMs
Jingyao Wang (Institute of Software Chinese Academy of Sciences), Hui Xiong (Hong Kong University of Science and Technology)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Fine-tuning LLM with reinforcement learning, designing information-theoretic dense process rewards for efficient reasoning.
Learning to Watermark: A Selective Watermarking Framework for Large Language Models via Multi-Objective Optimization
Chenrui Wang (Harbin Institute of Technology), Jing Li (Harbin Institute of Technology)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The paper proposes the 'Learning to Watermark' (LTW) framework, which achieves selective watermark insertion in large language model texts through training a lightweight selection network, balancing detection rate and text quality.
Learning to Zoom with Anatomical Relations for Medical Structure Detection
Bin Pu (Hunan University), Zhe Jin (Anhui University)
Object DetectionDomain AdaptationTransformerBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasound
🎯 What it does: A Zoom-Aware probabilistic framework ZR-DETR for medical structure detection is proposed, which combines scale-sensitive Zoom embeddings, anatomical relationship consistency constraints, and Gaussian process detection heads to achieve precise localization and uncertainty quantification of anatomical structures at different magnification levels.
Learning Urban Climate Dynamics via Physics-Guided Urban Surface–Atmosphere Interactions
Jiyang Xia (University of Manchester), Zhonghua Zheng (University of Manchester)
TransformerTime SeriesPhysics Related
🎯 What it does: This paper proposes UCformer, a physics-guided Transformer for multi-task prediction of urban 2-meter temperature, specific humidity, and dew point temperature, simulating urban surface-atmosphere interactions.
Learning When to Think: Shaping Adaptive Reasoning in R1-Style Models via Multi-Stage RL
Songjun Tu (Institute of Automation Chinese Academy of Sciences), Dongbin Zhao (Institute of Automation Chinese Academy of Sciences)
Computational EfficiencyReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposes AutoThink, a multi-stage reinforcement learning framework that learns to adaptively decide whether to perform explicit reasoning based on task difficulty in R1-style large models.
Learning with Calibration: Exploring Test-Time Computing of Spatio-Temporal Forecasting
Wei Chen (Hong Kong University of Science and Technology), Yuxuan Liang (Hong Kong University of Science and Technology)
Time SeriesSequentialBenchmark
🎯 What it does: A testing moment computation framework ST-TTC is proposed to compensate for non-stationary errors in spatiotemporal sequence prediction through calibration during the inference phase.
Learning with Restricted Boltzmann Machines: Asymptotics of AMP and GD in High Dimensions
Yizhou Xu (École Polytechnique Fédérale de Lausanne), Lenka Zdeborova
OptimizationTabular
🎯 What it does: In the high-dimensional limit, the likelihood training objective of RBM is mapped to an unsupervised multi-exponential model, and this equivalence is used to derive the asymptotic dynamics of AMP and gradient descent, proving that RBM can achieve the same weak recovery limit as the BBP threshold;
Learning with Statistical Equality Constraints
Aneesh Barthakur (University of Stuttgart), Luiz F. O. Chamon (Ecole Polytechnique)
OptimizationTabular
🎯 What it does: This paper studies statistical learning problems under equality constraints and proposes a set of generalization error theories and practical algorithms based on dual ascent.
Learning Without Augmenting: Unsupervised Time Series Representation Learning via Frame Projections
Berken Utku Demirel (ETH Zurich), Christian Holz (ETH Zurich)
Representation LearningConvolutional Neural NetworkContrastive LearningTime Series
🎯 What it does: A framework for unsupervised time series representation learning is proposed, which generates different views for instance contrastive learning using orthogonal bases and overcomplete frames without manual data augmentation.
Learning World Models for Interactive Video Generation
Taiye Chen (Peking University), Chi Jin (Princeton University)
GenerationData SynthesisTransformerDiffusion modelWorld ModelVideoRetrieval-Augmented Generation
🎯 What it does: An interactive long video generation framework that combines global state conditions and retrieval-augmented generation (VRAG) is proposed to construct an internal world model, significantly enhancing spatiotemporal consistency and reducing cumulative errors.
Learning-Augmented Algorithms for $k$-median via Online Learning
Anish Hebbar (Duke University), Debmalya Panigrahi (Indian Institute of Technology Delhi)
Optimization
🎯 What it does: This paper proposes an online learning-based learning enhancement algorithm that uses past k-median instances to predict the optimal center set for future instances, achieved through online convex optimization and fractional solutions.
Learning-Augmented Facility Location Mechanisms for the Envy Ratio Objective
Haris Aziz (University of New South Wales Sydney), Houyu Zhou (University of New South Wales Sydney)
Optimization
🎯 What it does: A fairness metric for facility location problems, called the envy ratio, is proposed with a learning enhancement mechanism, including the deterministic α-Bounding Interval Mechanism (α-BIM), a non-predictive random (α,p)-LRM constant mechanism, and a predictive Bias-Aware Mechanism (BAM);
Learning-Augmented Online Bidding in Stochastic Settings
Spyros Angelopoulos (CNRS and International Laboratory on Learning Systems), Bertrand Simon (Université Grenoble Alpes)
OptimizationTabular
🎯 What it does: This paper studies the optimal strategies for the online bidding problem under distributed prediction and randomized algorithms, providing Pareto-optimal deterministic and randomized solutions, and extending the methods to dynamic prediction and line search.
Learning-Augmented Online Bipartite Fractional Matching
Davin Choo (Harvard University), Yongho Shin (University of Wrocław)
OptimizationGraph
🎯 What it does: This study investigates the online bipartite fractional matching problem with learning enhancement, proposing two algorithms: LAB (supporting fractional predictions) and PAW (suitable for unweighted integral predictions). It analyzes their robustness-consistency trade-off and extends them under the AdWords small bidding hypothesis.
Learning-Augmented Streaming Algorithms for Correlation Clustering
Yinhao Dong (University of Science and Technology of China), Pan Peng (University of Science and Technology of China)
Graph
🎯 What it does: A learning-enhanced single transmission flow algorithm is proposed to address the related clustering problems of complete graphs and general graphs.
Least squares variational inference
Yvann Le Fay (ENSAE), Simon Barthelmé (GIPSA-Lab)
OptimizationTabular
🎯 What it does: A new gradient-free, least squares-based variational inference method called LSVI is proposed to approximate the target distribution within the exponential family.
Leaving No OOD Instance Behind: Instance-Level OOD Fine-Tuning for Anomaly Segmentation
YUXUAN ZHANG, Wei Yang (University of Science and Technology of China)
SegmentationAnomaly DetectionImage
🎯 What it does: A new instance-level OOD fine-tuning framework called LNOIB is proposed, aimed at improving the detection capability of small anomalies in anomaly segmentation.
LEDiT: Your Length-Extrapolatable Diffusion Transformer without Positional Encoding
Shen Zhang (JIIOV Technology), Yao Tang (JIIOV Technology)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: This paper proposes and implements a diffusion Transformer (LEDiT) that does not use explicit position encoding and can generate high-quality images at training resolutions beyond (scalable up to 4×).
LeMiCa: Lexicographic Minimax Path Caching for Efficient Diffusion-Based Video Generation
Huanlin Gao (China Unicom), Shiguo Lian (China Unicom)
GenerationOptimizationComputational EfficiencyDiffusion modelVideo
🎯 What it does: A training-free, globally optimized caching framework called LeMiCa is proposed to accelerate video generation with diffusion models. This framework models cache scheduling as a directed acyclic graph (DAG) path search problem and controls the worst-case error caused by caching through global error metrics and lexicographic minimax optimization, significantly improving inference speed while maintaining video quality.
Length Generalization via Auxiliary Tasks
Pranjal Awasthi, Ravi Kumar
TransformerSequential
🎯 What it does: A robust training framework based on auxiliary tasks is proposed to improve the performance of Transformers in sequence length generalization.
Less but More: Linear Adaptive Graph Learning Empowering Spatiotemporal Forecasting
Jiaming Ma (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
Graph Neural NetworkTransformerMixture of ExpertsTime Series
🎯 What it does: The MAGE framework is proposed, utilizing sparse balanced multi-expert linear adaptive graph learning to achieve efficient spatiotemporal prediction.
Less Greedy Equivalence Search
Adiba Ejaz (Columbia University), Elias Bareinboim (Columbia University)
Graph
🎯 What it does: Developed Less Greedy Equivalence Search (LGES) and I-ORIENT algorithms to improve the greedy search of traditional GES, enhancing the efficiency and accuracy of causal structure learning under observational and interventional data.
Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior
Yulin Li (Harbin Institute of Technology), Zhuotao Tian (Shenzhen Loop Area Institute)
CompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideo
🎯 What it does: A training-agnostic dynamic video token compression method called DyToK is proposed, which utilizes the internal attention mechanism of VLLM to automatically generate keyframe priors and dynamically allocate the number of tokens based on frame importance, thereby significantly improving inference efficiency while maintaining video understanding capabilities.
Less is More: an Attention-free Sequence Prediction Modeling for Offline Embodied Learning
Wei Huang (Shanghai AI Laboratory), Qinying Gu (Shanghai AI Laboratory)
Robotic IntelligenceTransformerReinforcement LearningSequential
🎯 What it does: This paper proposes a hierarchical sequence prediction model called Decision HiFormer (DHi), which merges the <state, action, reward> triplet of each time step into a single representation through Token Merger, and captures local dependencies between time steps using a non-parametric average pooling Token Mixer, significantly improving the inference speed and policy performance of offline reinforcement learning.
Less is More: Improving LLM Alignment via Preference Data Selection
Xun Deng (University of Science and Technology of China), Xiangnan He
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A Bayesian aggregation strategy based on marginal maximization (BeeS) is proposed to select high-quality preference samples during Direct Preference Optimization (DPO) training, thereby enhancing the alignment effect of large language models and significantly reducing training costs.
Less is More: Local Intrinsic Dimensions of Contextual Language Models
Benjamin Matthias Ruppik (Heinrich Heine University Düsseldorf), Milica Gasic
Representation LearningTransformerLarge Language ModelText
🎯 What it does: This paper studies the geometric structure of internal representations in large-scale language models (LLMs), using the local intrinsic dimension (LID) of context embeddings as an unsupervised diagnostic metric to analyze processes such as model training, fine-tuning, grokking, overfitting, and generalization.
Less is More: Unlocking Specialization of Time Series Foundation Models via Structured Pruning
Lifan Zhao (Shanghai Jiao Tong University), Jiaji Deng (Alibaba Group)
TransformerTime Series
🎯 What it does: Proposed the 'prune first, fine-tune later' paradigm, which focuses on the sub-networks important for specific tasks in time series foundational models (TSFMs) through structured pruning, thereby enhancing the overall sample prediction performance.
Lessons Learned: A Multi-Agent Framework for Code LLMs to Learn and Improve
Yuanzhe Liu (Rensselaer Polytechnic Institute), Jie Chen (IBM Research)
OptimizationAI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: The LessonL framework is proposed and implemented, allowing multiple large language models (LLMs) to collaboratively improve code in code optimization and code generation tasks by exchanging 'lessons' with each other.
Let a Neural Network be Your Invariant
Mirco Giacobbe (University of Birmingham), Michael Tautschnig (Amazon Web Services)
OptimizationSafty and PrivacyTabular
🎯 What it does: A neural network-based model checking method is proposed, capable of simultaneously verifying safety and liveness properties, and generating formal proofs through a neural ranking function.
Let Brain Rhythm Shape Machine Intelligence for Connecting Dots on Graphs
Jiaqi Ding (University of North Carolina), Guorong Wu (University of North Carolina)
Graph Neural NetworkGraphBiomedical DataMagnetic Resonance ImagingOrdinary Differential Equation
🎯 What it does: This paper proposes the BRICK framework, which models cognitive states through brain rhythm synchronization and extends it to the graph neural network BIG-NOS.
Let LRMs Break Free from Overthinking via Self-Braking Tuning
Haoran Zhao (Tianjin University), Yueting Zhuang (Zhejiang University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A Self-Braking Tuning (SBT) framework is proposed, which trains large-scale reasoning models to autonomously identify and terminate excessive reasoning steps, significantly reducing token consumption while maintaining accuracy.
Let Me Think! A Long Chain of Thought Can Be Worth Exponentially Many Short Ones
Parsa Mirtaheri (University of California San Diego), Enric Boix-Adserà (University of Pennsylvania)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningGraphChain-of-Thought
🎯 What it does: This paper systematically studies the performance differences in reasoning computation between sequential expansion (long chain thinking) and parallel expansion (multiple short chain thinking) through the design of a graph connectivity task.
Let the LLM Stick to Its Strengths: Learning to Route Economical LLM
Yi-Kai Zhang (Nanjing University), Han-Jia Ye (Nanjing University)
Recommendation SystemLarge Language ModelText
🎯 What it does: The LLMRec framework is constructed to achieve intelligent routing of large language models based on the Pareto front of cost and accuracy.
Let Them Talk: Audio-Driven Multi-Person Conversational Video Generation
Zhe Kong, Wenhan Luo
GenerationData SynthesisTransformerDiffusion modelVideoTextAudio
🎯 What it does: A multi-person dialogue video generation framework called MultiTalk is proposed, which can generate multi-character dialogue videos based on multiple audio inputs and text prompts, maintaining lip-sync and natural interaction.
Let's Revise Step-by-Step: A Unified Local Search Framework for Code Generation with LLMs
Zhiyi Lyu (Nanyang Technological University), Bo An (Nanyang Technological University)
GenerationOptimizationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: The ReLoc framework is proposed, unifying the code generation problem as a local search, achieving iterative code revision through four core components (initial draft, neighborhood generation, candidate evaluation, and master update), and can be instantiated as Hill-Climbing and Genetic-Algorithm;
Leveraging Conditional Dependence for Efficient World Model Denoising
Shaowei Zhang (Nanjing University), De-Chuan Zhan (Nanjing University)
Autonomous DrivingRobotic IntelligenceReinforcement Learning from Human FeedbackRecurrent Neural NetworkReinforcement LearningAuto EncoderWorld ModelVideo
🎯 What it does: A recursive state space model based on collision structure (CsRSSM) and an extended Dreamer framework (CsDreamer) are proposed for denoising observations and extracting task-relevant information in environments with noise interference.
Leveraging Depth and Language for Open-Vocabulary Domain-Generalized Semantic Segmentation
Siyu Chen (Jimei University), Meiliu Wu (University of Glasgow)
SegmentationDomain AdaptationAutonomous DrivingTransformerPrompt EngineeringImageMultimodality
🎯 What it does: A single-stage framework named Vireo is proposed to achieve the joint task of open-vocabulary semantic segmentation (OVSS) and domain generalization semantic segmentation (DGSS) without accessing target domain samples or new category annotations—open-vocabulary domain generalization semantic segmentation (OV-DGSS).
Leveraging Importance Sampling to Detach Alignment Modules from Large Language Models
Yi Liu (People's Daily Online), Zhendong Mao (University of Science and Technology of China)
Large Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: The Residual Alignment Model (RAM) is proposed, which splits the alignment of large models into a pre-training Proposal module and a lightweight Residual Aligner. Alignment is achieved through importance sampling, and the aligner is trained on sentence-level data.
Leveraging robust optimization for llm alignment under distribution shifts
Mingye Zhu (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A distribution-aware robust alignment framework called DoRA is proposed to mitigate distribution shifts caused by synthetic data in LLM preference alignment;
Leveraging semantic similarity for experimentation with AI-generated treatments
Lei Shi (University of California), Avi Feller (University of California)
OptimizationRepresentation LearningTransformerLarge Language ModelTextTabular
🎯 What it does: This paper proposes a dual-core representation learning framework for estimating causal effects and achieving adaptive allocation in online experiments, targeting the variable content generated by large language models.
LeVo: High-Quality Song Generation with Multi-Preference Alignment
Shun Lei (Shenzhen International Graduate School, Tsinghua University), Dong Yu (Shenzhen International Graduate School, Tsinghua University)
GenerationTransformerReinforcement LearningTextAudio
🎯 What it does: Developed the LeVo framework, which implements parallel prediction of mixed tokens and dual-track tokens for language models, generating high-quality songs and supporting text or audio prompt control.
Lie Detector: Unified Backdoor Detection via Cross-Examination Framework
Xuan Wang (National University of Defense Technology), Xitong Gao (Shenzhen Institutes of Advanced Technology)
ClassificationAnomaly DetectionSupervised Fine-TuningImageMultimodality
🎯 What it does: A unified cross-validation framework is proposed to detect backdoors in third-party trained models, utilizing inconsistencies between models for identification.
Lifelong Safety Alignment for Language Models
Haoyu Wang (Tsinghua University), Tianyu Pang (Sea AI Lab)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A lifelong safety alignment framework is proposed, allowing large language models to continuously learn and resist the ever-evolving jailbreak attacks.
Lifelong Test-Time Adaptation via Online Learning in Tracked Low-Dimensional Subspace
Dexin Duan (Shanghai Jiao Tong University), Fei Wen (Shanghai Jiao Tong University)
Domain AdaptationContrastive LearningImage
🎯 What it does: A lifelong testing-time adaptation (LCoTTA) method based on gradient low-dimensional subspaces is proposed to continuously update the model in continuously evolving target domains without the need for source data or labels.
Light-Weight Diffusion Multiplier and Uncertainty Quantification for Fourier Neural Operators
Albert Matveev (PhysicsX), Michalis Michaelides (PhysicsX)
Diffusion modelTabularBenchmark
🎯 What it does: A lightweight Fourier Neural Operator (DINOZAUR) based on thermal kernels is proposed, significantly reducing the number of parameters while maintaining prediction accuracy and supporting Bayesian uncertainty quantification.
LightFair: Towards an Efficient Alternative for Fair T2I Diffusion via Debiasing Pre-trained Text Encoders
Boyu Han (Institute of Computing Technology, Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
GenerationData SynthesisComputational EfficiencyTransformerSupervised Fine-TuningDiffusion modelImageText
🎯 What it does: This paper proposes a lightweight debiasing method called LightFair, which reduces gender and racial bias in text-to-image diffusion models by fine-tuning a pre-trained text encoder and combining two-stage text-guided sampling.
LILO: Learning to Reason at the Frontier of Learnability
Thomas Foster, Jakob Nicolaus Foerster
Reinforcement LearningText
🎯 What it does: This paper proposes an active curriculum learning method based on learnability called LILO, which dynamically selects the easiest problems to improve model performance for reinforcement learning training.
Limitations of Normalization in Attention
Timur Mudarisov (University of Luxembourg), Radu State (University of Luxembourg)
TransformerLarge Language ModelText
🎯 What it does: This paper theoretically derives and experimentally verifies the limitations of normalization (mainly softmax) in attention mechanisms, providing non-asymptotic upper bounds in terms of distance, geometry, and gradient, and tests these bounds on the GPT-2 model.
Limited Preference Data? Learning Better Reward Model with Latent Space Synthesis
Leitian Tao (University of Wisconsin Madison), Sharon Li (University of Wisconsin Madison)
Data SynthesisRecommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelAuto EncoderText
🎯 What it does: This paper studies a framework called LENS that synthesizes preference data in the latent space of LLMs using VAE, training a reward model with the generated embedding pairs.
LIMOPro: Reasoning Refinement for Efficient and Effective Test-time Scaling
Yang Xiao (Hong Kong Polytechnic University), Pengfei Liu (Shanghai Jiao Tong University)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: To address the lengthy chain reasoning in large language models for inference tasks, the authors propose a perplexity-based hierarchical importance filtering framework called PIR, which automatically removes functional steps to generate a more concise and accurate reasoning chain.
Linear Attention for Efficient Bidirectional Sequence Modeling
Arshia Afzal (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
Computational EfficiencyRecurrent Neural NetworkTransformerImageText
🎯 What it does: The LION framework is proposed, extending linear Transformers to bidirectional sequence modeling, and provides three equivalent representations: full attention, bidirectional RNN, and chunked parallelism, balancing training speed and inference efficiency.