NeurIPS 2025 Papers — Page 25
Conference on Neural Information Processing Systems · 5275 papers
L2RSI: Cross-view LiDAR-based Place Recognition for Large-scale Urban Scenes via Remote Sensing Imagery
Ziwei Shi (Xiamen University), Cheng Wang (Xiamen University)
RecognitionRetrievalContrastive LearningImagePoint Cloud
🎯 What it does: A cross-view laser radar location recognition framework L2RSI is proposed using high-resolution remote sensing images;
LabelAny3D: Label Any Object 3D in the Wild
Jin Yao (University of Virginia), Zezhou Cheng (University of Virginia)
Object DetectionSegmentationImageBenchmark
🎯 What it does: The LabelAny3D scheme is proposed, which automatically generates 3D bounding box annotations using an analysis-synthesis method based on monocular images, and constructs the COCO3D open vocabulary 3D detection benchmark based on this.
LABridge: Text–Image Latent Alignment Framework via Mean-Conditioned OU Process
Huiyang Shao (Tsinghua University), Xuefeng Xiao (Tsinghua University)
GenerationData SynthesisDiffusion modelAuto EncoderImageTextMultimodalityStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes the LABridge framework, which first encodes text into a structured prior aligned with the latent space of images, and then aligns and samples between this prior and the image latent space through a mean-conditioned Ornstein-Uhlenbeck (OU) diffusion process.
LaM-SLidE: Latent Space Modeling of Spatial Dynamical Systems via Linked Entities
Florian Sestak (LIT AI Lab), Johannes Brandstetter (LIT AI Lab)
GenerationData SynthesisComputational EfficiencyGraph Neural NetworkTransformerFlow-based ModelGenerative Adversarial NetworkTime SeriesSequential
🎯 What it does: A modeling method for spatial dynamic systems based on latent space encoding is proposed—LAM-SLIDE, which tracks individual entities in a fixed-size latent system representation using assignable identifiers (IDs) and predicts future trajectories in the latent space through flow models.
LangHOPS: Language Grounded Hierarchical Open-Vocabulary Part Segmentation
Yang Miao (INSAIT Sofia University), Luc Van Gool (INSAIT Sofia University)
Object DetectionSegmentationTransformerLarge Language ModelImageMultimodality
🎯 What it does: This paper proposes LangHOPS, an open vocabulary hierarchical object-part instance segmentation framework based on a multimodal large language model (MLLM), which can simultaneously detect and segment objects and their parts from images, and supports user-defined open vocabulary categories.
LangSplatV2: High-dimensional 3D Language Gaussian Splatting with 450+ FPS
Wanhua Li (Tsinghua University), Hanspeter Pfister (Harvard University)
SegmentationRetrievalComputational EfficiencyContrastive LearningGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes LangSplatV2, which replaces the high-dimensional MLP decoder with a sparse coefficient dictionary to achieve real-time high-dimensional feature splatting for 3D open vocabulary queries, and experiments are conducted on datasets such as LERF, 3D-OVS, and Mip-NeRF360.
Language Model Behavioral Phases are Consistent Across Architecture, Training Data, and Scale
James A. Michaelov (Massachusetts Institute of Technology), Ben Bergen
TransformerLarge Language ModelText
🎯 What it does: A systematic analysis of the behavior of autoregressive language models across multiple architectures (Transformer, Mamba, RWKV), scales (1.4M–12B parameters), and datasets (OpenWebText, The Pile) during the pre-training process is conducted, proposing and validating three simple heuristics—word frequency, n-gram probability, and semantic similarity—that can explain their prediction variance.
Language Modeling by Language Models
Junyan Cheng (Dartmouth College), Kyle Richardson (Allen Institute for AI)
Large Language ModelAgentic AIText
🎯 What it does: This study investigates the use of multi-agent LLMs to automate the discovery of new language model architectures.
Language Models (Mostly) Know When to Stop Reading
Roy Xie (Duke University), Bhuwan Dhingra (Duke University)
CompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study proposes a dynamic context truncation method that allows large language models to self-stop reading after obtaining sufficient information, thereby improving reasoning efficiency.
Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activations
Li Ji-An (University of California San Diego), Marcus K. Benna (University of California San Diego)
Large Language ModelText
🎯 What it does: This paper proposes a neural feedback paradigm that utilizes the monitoring and control capabilities of internal activations in large language models (LLMs) during contextual learning assessments.
Language Models Can Predict Their Own Behavior
Dhananjay Ashok (University of Southern California), Jonathan May (University of Southern California)
ClassificationGenerationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Training a linear probe utilizes the hidden states of the input layer of the language model to predict the overall behavior of the model before generating any tokens, and combines Conformal Prediction to build an early warning system;
Language Models can Self-Improve at State-Value Estimation for Better Search
Ethan Mendes (Georgia Institute of Technology), Alan Ritter (Georgia Institute of Technology)
TransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: A self-supervised framework called STL is proposed, which utilizes language models to generate actions, outcomes, and reasoning in a single-step lookahead to improve state value estimation and enhance search performance without the need for rewards or demonstrations.
Language Ranker: A Lightweight Ranking framework for LLM Decoding
Chenheng Zhang (Peking University), Zhouchen Lin (Peking University)
Recommendation SystemTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The study proposes Language Ranker, a lightweight ranking framework that reorders candidate responses based on features extracted from a base model, thereby enhancing the decoding performance of LLMs.
Language‑Bias‑Resilient Visual Question Answering via Adaptive Multi‑Margin Collaborative Debiasing
Huanjia Zhu (Beijing Institute of Technology), Bingzhi Chen (Beijing Institute of Technology)
ClassificationRecognitionTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: This study investigates the formation mechanism of language bias in Visual Question Answering (VQA) and proposes an Adaptive Multi-Angle Boundary (MMCD) framework that combines frequency, confidence, and difficulty as three adaptive angle boundaries with difficulty-aware contrastive learning to dynamically reshape decision boundaries to suppress language bias.
LaRes: Evolutionary Reinforcement Learning with LLM-based Adaptive Reward Search
Pengyi Li (Tianjin University), Jianye HAO
Robotic IntelligenceLarge Language ModelReinforcement LearningSequential
🎯 What it does: This paper proposes the LaRes framework, which generates and evolves reward functions through LLM and combines evolutionary reinforcement learning to achieve efficient policy learning.
Large Language Bayes
Justin Domke (University of Massachusetts Amherst)
Large Language Model
🎯 What it does: By combining large language models with probabilistic programming languages, non-experts can describe problems in natural language, generate a series of formal Bayesian models, and perform Bayesian model averaging to obtain the final posterior distribution.
Large Language Diffusion Models
Shen Nie (Renmin University of China), Chongxuan Li (Renmin University of China)
TransformerLarge Language ModelSupervised Fine-TuningDiffusion modelText
🎯 What it does: This paper proposes and trains a language diffusion model LLaDA with 8 billion parameters from scratch, demonstrating that discrete diffusion models can achieve core capabilities of LLMs comparable to autoregressive models.
Large Language Models as End-to-end Combinatorial Optimization Solvers
Xia Jiang (Eindhoven University of Technology), Yingqian Zhang (Eindhoven University of Technology)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: An end-to-end combinatorial optimization solving framework based on LLM is proposed, enabling LLMs to directly generate feasible high-quality solutions from natural language descriptions through supervised fine-tuning and constraint- and optimality-friendly reinforcement learning.
Large Language Models as Model Organisms for Human Associative Learning
Camila Kolling (Max Planck Institute for Software Systems), Mariya Toneva (Max Planck Institute for Software Systems)
TransformerLarge Language ModelText
🎯 What it does: Simulating human associative learning in large language models (LLMs) by repeatedly presenting token pairs in context and observing the changes in their internal representations as learning occurs.
Large language models can learn and generalize steganographic chain-of-thought under process supervision
Robert McCarthy (University College London), David Lindner
Large Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: The study demonstrates that under the supervision of reinforcement learning, LLM can hide malicious or reward-defying reasoning through chain-of-thought, showcasing the learnability and generalization of hidden reasoning.
Large Language Models for Lossless Image Compression: Next-Pixel Prediction in Language Space is All You Need
Kecheng Chen (City University of Hong Kong), Haoliang Li (City University of Hong Kong)
CompressionTransformerLarge Language ModelSupervised Fine-TuningImage
🎯 What it does: This paper proposes a lossless image compression framework based on large language models (P²-LLM), which achieves more accurate next pixel prediction through pixel-level priors, context learning, and two-step lossless pixel tokenization, thereby improving the compression rate.
Large Language Models Think Too Fast To Explore Effectively
Lan Pan (Georgia Institute of Technology), Robert Wilson
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This study investigates the performance of LLMs in open-ended exploration tasks (Little Alchemy 2) and compares it with human performance.
Large Stepsizes Accelerate Gradient Descent for Regularized Logistic Regression
Jingfeng Wu (University of California Berkeley), Peter Bartlett
OptimizationTabular
🎯 What it does: This study investigates the use of constant step size gradient descent (GD) with ℓ₂ regularization in logistic regression on linearly separable data. It shows that using a larger step size can achieve non-monotonic convergence, and under a large step size, the convergence speed of GD improves to a level comparable to Nesterov's accelerated method, achieving O(√κ) rate.
LARGO: Latent Adversarial Reflection through Gradient Optimization for Jailbreaking LLMs
Ran Li (Columbia University), Chengzhi Mao (Rutgers University)
OptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The LARGO framework is proposed, utilizing gradient optimization in the continuous latent space of LLMs, and then generating smooth jailbreak suffixes through self-reflective decoding to achieve more efficient and natural attacks.
LASeR: Learning to Adaptively Select Reward Models with Multi-Arm Bandits
Duy Nguyen (University of North Carolina at Chapel Hill), Mohit Bansal (University of North Carolina at Chapel Hill)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The LASER method is proposed, which utilizes a multi-armed bandit dynamic selection of the reward model (Reward Model) to choose the most suitable RM for each training batch or instance, generating preference pairs for iterative fine-tuning of LLM;
Last Iterate Convergence in Monotone Mean Field Games
Noboru Isobe (RIKEN), Kaito Ariu (CyberAgent)
OptimizationReinforcement LearningTabular
🎯 What it does: In the framework of Mean-Field Game (MFG) without strict monotonicity, a Proximal Point (PP) iteration based on Kullback-Leibler regularization is proposed and proven to achieve Last Iteration Convergence (LIC). Subsequently, an Approximate Proximal-Point (APP) algorithm is designed, which approximates PP through several steps of Regularized Mirror Descent (RMD), thereby converging to the MFG equilibrium under discrete time and non-strict monotonicity conditions.
Last-Iterate Convergence of Smooth Regret Matching$^+$ Variants in Learning Nash Equilibria
Linjian Meng (Nanjing University), Yang Gao (Nanjing University)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the final iteration convergence of the Smooth Regret Matching+ (RM+) variant in learning Nash Equilibrium (NE), proposes a new proof paradigm, and introduces the Smooth Optimistic Gradient-based RM+ (SOGRM+) algorithm.
Latency NMS Attacks: Is It Real Life or Is It Just Fantasy?
Jean-Philippe Monteuuis (Qualcomm Technologies, Inc.), Jonathan Petit (Qualcomm Technologies, Inc.)
Object DetectionAdversarial AttackImage
🎯 What it does: This paper systematically examines the effectiveness of existing non-maximum suppression (NMS) delay attacks in real deployment environments through the design of the evaluation framework EVADE. The results indicate that these attacks show no significant delay gain under most hardware, model versions, quantization, and defense settings, posing essentially no real threat.
Latent Chain-of-Thought for Visual Reasoning
Guohao Sun (Rochester Institute of Technology), Zhiqiang Tao (Rochester Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: This paper proposes a latent chain reasoning method (LaCoT) based on variational inference, which samples diverse visual reasoning chains and generates answers through an autoregressive strategy.
Latent Harmony: Synergistic Unified UHD Image Restoration via Latent Space Regularization and Controllable Refinement
Yidi Liu (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
RestorationSuper ResolutionAuto EncoderImage
🎯 What it does: A two-stage framework called Latent Harmony is proposed for UHD full-scene image restoration.
Latent Mixture of Symmetries for Sample-Efficient Dynamic Learning
Haoran Li (Arizona State University), Yang Weng (Arizona State University)
Mixture of ExpertsTabularTime SeriesSequentialBiomedical DataElectrocardiogramOrdinary Differential Equation
🎯 What it does: A latent dynamics learning framework based on mixed symmetry, called Latent MoS, is proposed for efficiently learning system dynamics in sparse and noisy environments.
Latent Policy Barrier: Learning Robust Visuomotor Policies by Staying In-Distribution
Zhanyi Sun (Stanford University), Shuran Song
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelImage
🎯 What it does: The Latent Policy Barrier (LPB) framework is proposed, which uses diffusion strategies and visual latent dynamics models in the visual latent space to real-time guide behavior cloning strategies back to the expert distribution, thereby enhancing robustness and sample efficiency.
Latent Principle Discovery for Language Model Self-Improvement
Keshav Ramji (IBM Research AI), Ramón Fernandez Astudillo (IBM Research AI)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelText
🎯 What it does: This paper proposes a self-supervised principle learning framework called STaPLe, which utilizes the model's own generation and learning of hidden 'principles' to guide self-correction, thereby achieving self-improvement of the language model.
Latent Refinement via Flow Matching for Training-free Linear Inverse Problem Solving
Hossein Askari (University of Queensland), Fred Roosta (ARC Training Centre for Information Resilience)
RestorationSuper ResolutionFlow-based ModelAuto EncoderImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The LFlow framework is proposed, achieving untrained linear inverse problem solving based on pre-trained latent flow matching.
Latent Retrieval Augmented Generation of Cross-Domain Protein Binders
Zishen Zhang (Tsinghua University), Yang Liu (Tsinghua University)
GenerationRetrievalDrug DiscoveryDiffusion modelAuto EncoderContrastive LearningBiomedical DataRetrieval-Augmented Generation
🎯 What it does: A retrieval-enhanced latent diffusion framework called RADiAnce is proposed for the design of ligands for cross-domain protein complexes.
Latent Space Factorization in LoRA
Shashi Kumar (Idiap Research Institute), Ina Kodrasi (Idiap Research Institute)
ClassificationRepresentation LearningTransformerAuto EncoderImageTextAudio
🎯 What it does: The FVAE-LoRA method is proposed, which combines Variational Autoencoders (VAE) with LoRA, using two separate latent spaces (z1 responsible for task-related information and z2 responsible for remaining information) to achieve low-rank fine-tuning.
Latent Zoning Network: A Unified Principle for Generative Modeling, Representation Learning, and Classification
Zinan Lin (Microsoft Research), Sergey Yekhanin (Microsoft Research)
ClassificationGenerationRepresentation LearningConvolutional Neural NetworkFlow-based ModelRectified FlowAuto EncoderContrastive LearningImage
🎯 What it does: This paper proposes the Latent Zoning Network (LZN), a unified framework that connects generative models, unsupervised representation learning, and classification tasks through a shared Gaussian latent space.
Lattice Boltzmann Model for Learning Real-World Pixel Dynamicity
Guangze Zheng (Hong Kong University), Jia Pan (Hong Kong University)
Object TrackingTransformerOptical FlowVideo
🎯 What it does: A real-time pixel-level tracking framework based on the lattice Boltzmann model (LBM) is proposed to learn and predict pixel dynamics in the real world, applied to point tracking and open-source object tracking.
LaViDa: A Large Diffusion Model for Vision-Language Understanding
Shufan Li (UCLA), Aditya Grover (Adobe Research)
RecognitionGenerationTransformerVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: LaViDa is proposed, a visual language model based on discrete diffusion models, capable of parallel decoding, bidirectional reasoning, and controllable text generation.
LaX: Boosting Low-Rank Training of Foundation Models via Latent Crossing
Ruijie ZHANG, Zheng Zhang (University of California at Santa Barbara)
CompressionOptimizationTransformerLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: A lightweight 'Latent Crossing (LaX)' module is proposed, which utilizes residual paths to transmit information between low-rank subspaces, thereby enhancing the expressive power of low-rank compression models.
Layer as Puzzle Pieces: Compressing Large Language Models through Layer Concatenation
Fei Wang (South China University of Technology), Changxing Ding (South China University of Technology)
CompressionKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Proposes the CoMe framework for hierarchical pruning and post-training recovery of large language models, reducing the number of parameters while maintaining performance;
Layer-Wise Modality Decomposition for Interpretable Multimodal Sensor Fusion
Park Jae Hyun, Jun Won Choi (Seoul National University)
Autonomous DrivingExplainability and InterpretabilityMultimodality
🎯 What it does: A Layer-Wise Modality Decomposition (LMD) method is proposed to decompose the contributions of each modality layer by layer in a pre-trained multimodal sensor fusion model, achieving interpretability of the model's predictions.
Layer-wise Update Aggregation with Recycling for Communication-Efficient Federated Learning
Jisoo Kim (Inha University), Sunwoo Lee (Inha University)
Federated LearningText
🎯 What it does: A hierarchical update recycling mechanism called FedLUAR is proposed, which uses the gradient-weight ratio to select less important layers and reuses old gradients to significantly reduce communication costs.
LayerCraft: Enhancing Text-to-Image Generation with CoT Reasoning and Layered Object Integration
Yuyao Zhang, Yu-Wing Tai
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageTextChain-of-Thought
🎯 What it does: Proposes the LayerCraft framework, using large language models (LLM) as autonomous agents to achieve structured generation from text to image and layered object integration, supporting batch synthesis and multi-round editing.
LayerIF: Estimating Layer Quality for Large Language Models using Influence Functions
Hadi Askari (University of California), Muhao Chen (University of South Florida)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: The LAYERIF framework is proposed, which evaluates the training quality of each Transformer layer in large language models (LLMs) through influence functions, and applies this evaluation to two types of downstream tasks: LoRA-MoE expert allocation and LLM pruning with hierarchical sparse allocation.
LayerNavigator: Finding Promising Intervention Layers for Efficient Activation Steering in Large Language Models
Hao Sun (Institute of Information Engineering), Yanan Cao (Institute of Information Engineering)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes LayerNavigator, a method for selecting activation-guided layers based on layer guidability scores to enhance the behavioral alignment of large language models.
LBMKGC: Large Model-Driven Balanced Multimodal Knowledge Graph Completion
Yuan Guo (Dalian Maritime University), Shikai Guo (Dalian Maritime University)
Data SynthesisRetrievalKnowledge DistillationGraph Neural NetworkTransformerVision Language ModelContrastive LearningMultimodalityGraph
🎯 What it does: A large model-driven balanced multimodal knowledge graph completion framework LBMKGC is proposed.
LeapFactual: Reliable Visual Counterfactual Explanation Using Conditional Flow Matching
Zhuo Cao (Forschungszentrum Julich), Ira Assent (Aarhus University)
GenerationExplainability and InterpretabilityFlow-based ModelGenerative Adversarial NetworkImage
🎯 What it does: LEAPFACTUAL is proposed, an explainable counterfactual generation algorithm based on Conditional Flow Matching;
Learn and Ensemble Bridge Adapters for Multi-domain Task Incremental Learning
Ziqi Gu (Nanjing University of Science and Technology), Zhen Cui (Beijing Normal University)
Domain AdaptationTransformerVision Language ModelDiffusion modelImageMultimodalityBenchmark
🎯 What it does: The LEBA framework is proposed for multi-domain task incremental learning, aimed at alleviating catastrophic forgetting and enhancing cross-domain generalization.
Learn2Mix: Training Neural Networks Using Adaptive Data Integration
Shyam Venkatasubramanian (Duke University), Vahid Tarokh (Duke University)
ClassificationOptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: An adaptive training strategy called learn2mix is proposed in resource-constrained environments, which accelerates the convergence of neural networks by dynamically adjusting the class ratio in the batch.
Learnable Burst-Encodable Time-of-Flight Imaging for High-Fidelity Long-Distance Depth Sensing
Manchao Bao (Nanjing University), Xuemei Hu (Nanjing University)
Depth EstimationConvolutional Neural NetworkSupervised Fine-TuningPoint Cloud
🎯 What it does: A Burst-Encodable Time-of-Flight (BE-ToF) system is proposed and implemented, achieving long-distance high-precision depth perception with a single-frequency pulse bundle mode.
Learnable Sampler Distillation for Discrete Diffusion Models
Feiyang Fu (University of Electronic Science and Technology of China), Zhaoqiang Liu (University of Electronic Science and Technology of China)
GenerationKnowledge DistillationDiffusion modelImageText
🎯 What it does: A fast sampler learning method based on knowledge distillation (LSD/LSD+) is proposed for discrete diffusion models.
Learned Prefix Caching for Efficient LLM Inference
Dongsheng Yang (Princeton University), Wyatt Lloyd (Princeton University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper studies a learning-based prefix caching strategy (LPC) that enhances the inference efficiency of LLMs by predicting whether a session will continue.
Learning “Partner-Aware” Collaborators in Multi-Party Collaboration
Abhijnan Nath (Colorado State University), Nikhil Krishnaswamy (Colorado State University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: The Interruptible Collaborative Roleplayer (ICR) model is proposed, utilizing counterfactual KL regularization with adversarial prefixes to enable large language models to distinguish between useful and misleading interventions in multi-party collaboration, thereby achieving 'partner-aware' collaboration.
Learning (Approximately) Equivariant Networks via Constrained Optimization
Andrei Manolache (University of Stuttgart), Mathias Niepert (University of Stuttgart)
OptimizationGraph Neural NetworkPoint Cloud
🎯 What it does: This paper proposes an adaptive constrained training framework named ACE (Adaptive Constrained Equivariance), which balances the model's symmetry and expressiveness by gradually tightening or relaxing equivariance constraints at the network layer level, thereby enhancing training stability, sample efficiency, and robustness.
Learning 3D Anisotropic Noise Distributions Improves Molecular Force Fields
Xixian Liu (Fudan University), Yixin Cao (Institute of Trustworthy Embodied AI, Fudan University)
Drug DiscoveryTransformerAuto EncoderPoint CloudGraph
🎯 What it does: Proposes the AniDS framework, which utilizes a structure-aware full covariance noise generator for 3D molecular coordinate denoising, thereby learning molecular force fields.
Learning 3D Persistent Embodied World Models
Siyuan Zhou (Hong Kong University of Science and Technology), Chuang Gan (University of Massachusetts Amherst)
GenerationRobotic IntelligenceReinforcement Learning from Human FeedbackDiffusion modelWorld ModelVideo
🎯 What it does: This paper proposes a persistent embedded world model that combines 3D memory, utilizing a video diffusion model to generate RGB-D videos and achieving long-term consistent visual simulation through voxelized 3D maps, which can support robot planning and strategy learning.
Learning a Cross-Modal Schrödinger Bridge for Visual Domain Generalization
Hao Zheng (Tencent Jarvis Lab), Yefeng Zheng (Westlake University)
ClassificationSegmentationDomain AdaptationVision Language ModelImageTextStochastic Differential Equation
🎯 What it does: Proposes a visual domain generalization method based on cross-modal Schrödinger bridge.
Learning Across the Gap: Hybrid Multi-armed Bandits with Heterogeneous Offline and Online Data
Qijia He (Southern University of Science and Technology), Fang Kong (Southern University of Science and Technology)
Recommendation SystemOptimizationTabular
🎯 What it does: This paper proposes a hybrid multi-armed bandit framework that utilizes offline heterogeneous data (absolute feedback or preference feedback) to accelerate online learning, covering both scenarios from absolute to relative feedback and from relative to absolute feedback.
Learning and Planning Multi-Agent Tasks via an MoE-based World Model
Zijie Zhao (University of Chinese Academy of Sciences), Dongbin Zhao (Institute of Automation, Chinese Academy of Sciences)
Robotic IntelligenceReinforcement LearningMixture of ExpertsWorld ModelBenchmark
🎯 What it does: A world model M3W based on Mixture-of-Experts (MoE) is proposed for learning and planning in multi-task multi-agent reinforcement learning.
Learning CAD Modeling Sequences via Projection and Part Awareness
Yang Liu (Beijing Institute of Technology), Fang Deng (Beijing Institute of Technology)
GenerationData SynthesisPoint Cloud
🎯 What it does: Automatically generate editable CAD modeling instruction sequences from 3D point clouds, reconstructing geometry and modeling processes.
Learning Chern Numbers of Multiband Topological Insulators with Gauge Equivariant Neural Networks
Longde Huang (Chalmers University of Technology University of Gothenburg), Jan E Gerken
Physics Related
🎯 What it does: Using a local Gauge Equivariant Neural Network to predict the Chern number of multi-band topological insulators, addressing the challenge of local Gauge symmetry that traditional methods struggle with.
Learning Cocoercive Conservative Denoisers via Helmholtz Decomposition for Poisson Imaging Inverse Problems
Deliang Wei (East China Normal University), Tieyong Zeng (Beijing Normal-Hong Kong Baptist University)
RestorationSupervised Fine-TuningImageComputed Tomography
🎯 What it does: A Plug-and-Play method for the Poisson inverse problem based on the cocoercive conservative (CoCo) assumption is proposed, constructing a CoCo denoiser and proving it to be a proximal operator of a weakly convex prior. Subsequently, two convergent PnP algorithms, CoCo-ADMM and CoCo-PEGD, are presented.
Learning conformational ensembles of proteins based on backbone geometry
Nicolas Wolf (Max Planck Institute for Polymer Research), Frauke Gräter (Max Planck Institute for Polymer Research)
Protein Structure PredictionGraph Neural NetworkFlow-based ModelTime SeriesBiomedical Data
🎯 What it does: A flow matching generative model BBFlow based on backbone geometry is proposed for rapid sampling of protein conformation sets.
Learning Counterfactual Outcomes Under Rank Preservation
Peng Wu (Beijing Technology and Business University), Kun Zhang (Intuit AI Research)
OptimizationTabular
🎯 What it does: A new adversarial assumption called the 'Rank Preservation Assumption' is proposed, and based on this, an ideal loss function and kernel smoothing estimator are constructed for unbiased estimation of individual counterfactual outcomes under unknown structural causal models.
Learning Crossmodal Interaction Patterns via Attributed Bipartite Graphs for Single-Cell Omics
Xiaotang Wang (Hong Kong University of Science and Technology), Yongqi Zhang (Hong Kong University of Science and Technology)
Graph Neural NetworkTransformerGraphBiomedical Data
🎯 What it does: This paper proposes a method to reformulate the single-cell cross-modal matching problem as a graph classification task on an Attributed Bipartite Graph (ABG), and uses the Bi 2 Former model to learn ATAC-RNA interactions.
Learning Dense Hand Contact Estimation from Imbalanced Data
Daniel Sungho Jung (Seoul National University), Kyoung Mu Lee (Seoul National University)
Pose EstimationTransformerImage
🎯 What it does: A framework named HACO is proposed for learning dense hand contact estimation from hand contact data with category and spatial imbalance.
Learning Differential Pyramid Representation for Tone Mapping
Qirui Yang (Tianjin University), Jingyu Yang (Tianjin University)
Image TranslationRestorationImageVideo
🎯 What it does: Proposes an end-to-end Differential Pyramid Representation Network (DPRNet) to achieve high-fidelity tone mapping from HDR to LDR.
Learning Diffusion Models with Flexible Representation Guidance
Chenyu Wang (Massachusetts Institute of Technology), Tommi Jaakkola (Massachusetts Institute of Technology)
GenerationData SynthesisDrug DiscoveryDiffusion modelImageMultimodalityBiomedical Data
🎯 What it does: A general theoretical framework and implementation method called REED is proposed, which uses pre-trained representations to guide the training of diffusion models, significantly improving the quality of image, protein sequence, and molecular generation while accelerating training.
Learning Dynamics of RNNs in Closed-Loop Environments
Yoav Ger (Technion Israel Institute of Technology), Omri Barak (Technion Israel Institute of Technology)
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: This study investigates the learning dynamics of recurrent neural networks (RNNs) in closed-loop environments and compares them with open-loop training.
Learning Efficient Fuse-and-Refine for Feed-Forward 3D Gaussian Splatting
Yiming Wang (ETH Zurich), Tiancheng Sun (Google)
RestorationCompressionOptimizationComputational EfficiencyTransformerGaussian SplattingOptical FlowVideo
🎯 What it does: Proposes the Fuse-and-Refine module, which projects existing pixel-aligned 3D Gaussian splats into voxel space, utilizing a coarse-to-fine voxel structure and a sparse voxel Transformer to fuse and refine primitives, achieving fast and high-quality reconstruction in both static and dynamic scenes;
Learning Equilibria from Data: Provably Efficient Multi-Agent Imitation Learning
Till Freihaut (University of Zurich), Giorgia Ramponi
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: The study focuses on learning Nash equilibria using expert data in zero-sum Markov games, providing a sample complexity analysis for behavior cloning (BC) and two interactive multi-agent imitation learning methods (MAIL-BRO, MURMAIL).
Learning Expandable and Adaptable Representations for Continual Learning
Ruilong Yu (University of Electronic Science and Technology of China), shijie zhou
Representation LearningTransformerMixture of ExpertsImage
🎯 What it does: A scalable and adaptive representation learning framework called LEAR is designed to achieve multi-domain continual learning using a collaborative backbone and expert networks.
Learning from A Single Markovian Trajectory: Optimality and Variance Reduction
Zhenyu Sun (Northwestern University), Ermin Wei (Northwestern University)
Optimization
🎯 What it does: The study investigates non-convex stochastic optimization problems under a single Markov chain sample, proposing lower and upper bounds.
Learning from Delayed Feedback in Games via Extra Prediction
Yuma Fujimoto (CyberAgent), Kaito Ariu (CyberAgent)
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: This paper proposes and studies the negative impact of time delays in multi-agent game learning, and overcomes this issue by introducing Weighted Optimized Forecasting (WOFTRL);
Learning from Demonstrations via Capability-Aware Goal Sampling
Yuanlin Duan (Rutgers University), He Zhu (Rutgers University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningWorld ModelSequential
🎯 What it does: This paper proposes a learning method called Cago, which dynamically tracks the capability boundaries of agents on demonstration trajectories to sample intermediate sub-goals that guide exploration, thereby achieving more efficient learning in long-horizon tasks with sparse rewards.
Learning from Disjoint Views: A Contrastive Prototype Matching Network for Fully Incomplete Multi-View Clustering
Yiming Wang (Nanjing University of Posts and Telecommunications), Yao Zhao (Beijing Jiaotong University)
Auto EncoderContrastive LearningImageText
🎯 What it does: A Contrastive Prototype Matching Network (CPMN) is proposed for completely uncorresponded multi-view clustering, achieving cross-view prototype alignment and contrastive learning without corresponding information, ultimately resulting in more consistent clustering outcomes.
Learning from Interval Targets
Rattana Pukdee (Carnegie Mellon University), Chirag Gupta (Bloomberg)
Tabular
🎯 What it does: For regression tasks with only a given target interval, this paper proposes projection loss and worst-case loss, and provides non-asymptotic generalization bounds.
Learning from positive and unlabeled examples -Finite size sample bounds
Farnam Mansouri (University of Waterloo and Vector Institute), Shai Ben-David (University of Waterloo and Vector Institute)
ClassificationDomain Adaptation
🎯 What it does: This paper studies PU learning (only positive samples and unlabeled samples) under the imbalance of positive and negative labels, providing upper and lower bounds on finite sample complexity under various label distribution assumptions (SCAR, SAR, PCS, APDS), and discusses the unlearnability when class priors are unknown.
Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models
Vlad Sobal (New York University), Yann LeCun (New York University)
OptimizationRobotic IntelligenceReinforcement LearningContrastive LearningSequential
🎯 What it does: This paper systematically evaluates various methods for learning from offline data without rewards, particularly comparing reinforcement learning (RL) and model-based planning (control) paradigms. It proposes a new method called 'Planning with a Potential Dynamics Model (PLDM)' and conducts comprehensive experiments in a custom navigation environment.
Learning from Videos for 3D World: Enhancing MLLMs with 3D Vision Geometry Priors
Duo Zheng (Chinese University of Hong Kong), Liwei Wang (Chinese University of Hong Kong)
Object DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodality
🎯 What it does: This paper proposes a framework named VG LLM, which utilizes a 3D visual geometry encoder to extract potential 3D geometric priors from inputs containing only RGB videos and integrates them into a multimodal large language model (MLLM) for understanding 3D scenes and spatial reasoning.
Learning Generalizable Shape Completion with SIM(3) Equivariance
Yuqing Wang (Technical University of Munich), Xiao Xiang Zhu (Technical University of Munich)
RestorationDomain AdaptationAutonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper presents the first fully SIM(3) equivariant 3D shape completion network, which maintains consistency under arbitrary rotations, translations, and scale transformations, and achieves generalization on unaligned real scans.
Learning Gradient Boosted Decision Trees with Algorithmic Recourse
Kentaro Kanamori (Fujitsu Limited), Takuya Takagi (Fujitsu Limited)
OptimizationTabularFinance Related
🎯 What it does: The research proposes RABIT, a new framework for training Gradient Boosting Decision Trees (GBDT) under the premise of ensuring feasible recourse.
Learning Grouped Lattice Vector Quantizers for Low-Bit LLM Compression
Xi Zhang (Nanyang Technological University), Weisi Lin (Nanyang Technological University)
CompressionLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A Group Lattice Vector Quantization (GLVQ) framework is designed to achieve post-training quantization of large language models at extremely low bit widths (≤3bit) while maintaining or improving performance.
Learning Human-Like RL Agents Through Trajectory Optimization With Action Quantization
Jian-Ting Guo (National Yang Ming Chiao Tung University), I-Chen Wu (National Yang Ming Chiao Tung University)
OptimizationRobotic IntelligenceReinforcement LearningAgentic AISequential
🎯 What it does: The paper proposes a method for achieving human-like reinforcement learning agents through trajectory optimization and action quantization, namely Macro Action Quantization (MAQ).
Learning Human-Object Interaction as Groups
Jiajun Hong (Zhejiang University), Wenguan Wang (Zhejiang University)
RecognitionObject DetectionTransformerVision Language ModelImage
🎯 What it does: The GroupHOI framework is proposed, treating human-object interactions as groups, constructing groups using geometric proximity and semantic similarity, and achieving relational reasoning through local context aggregation and global transformers.
Learning in Compact Spaces with Approximately Normalized Transformer
Jörg K.H. Franke (University of Freiburg), Michael Hefenbrock (Perspix.ai)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes an approximate normalization Transformer (anTransformer) architecture, which utilizes the concentration phenomenon of high-dimensional vector norms to replace traditional normalization operations like LayerNorm with feasible scalar multiplication, thereby reducing hyperparameters and computational overhead while maintaining or improving convergence speed.
Learning in Stackelberg Mean Field Games: A Non-Asymptotic Analysis
Sihan Zeng (JPMorgan AI Research), Alec Koppel (JPMorgan AI Research)
OptimizationReinforcement Learning
🎯 What it does: Proposes a single-loop actor-critic algorithm AC-SMFG to learn the optimal leader strategy in Stackelberg equilibrium games.
Learning Individual Behavior in Agent-Based Models with Graph Diffusion Networks
Francesco Cozzi (Sapienza University), Corrado Monti (CENTAI)
Graph Neural NetworkDiffusion modelGraph
🎯 What it does: A differentiable surrogate model based on Graph Diffusion Network (GDN) is proposed, which can learn individual-level stochastic behaviors from the data of the original Agent-Based Model (ABM) and generate future states.
Learning Interactive World Model for Object-Centric Reinforcement Learning
Fan Feng (University of California San Diego), Sara Magliacane (University of Amsterdam)
Robotic IntelligenceReinforcement LearningWorld ModelGraph
🎯 What it does: A hierarchical world model is proposed, implemented through objects and their interactions, learning the static and dynamic properties of objects and representing them with interaction graphs.
Learning Interestingness in Automated Mathematical Theory Formation
George Tsoukalas (University of Texas at Austin), Swarat Chaudhuri (University of Texas at Austin)
Large Language ModelReinforcement LearningGraph
🎯 What it does: The FERMT environment and EvoAbstract algorithm are proposed, completing a reinforcement learning framework from symbolic actions to theoretical discovery, and implementing an automatic learning 'interestingness' evaluator to guide the formation of mathematical theories.
Learning Intractable Multimodal Policies with Reparameterization and Diversity Regularization
Ziqi Wang (Hong Kong University of Science and Technology), Ling Pan (Hong Kong University of Science and Technology)
Robotic IntelligenceReinforcement LearningMultimodalityBenchmark
🎯 What it does: The Diversity-regularized Actor Critic (DrAC) algorithm is proposed, which utilizes reparameterization and distance-based multimodal policy diversity regularization to train intractable multimodal decision policies.
Learning Juntas under Markov Random Fields
Gautam Chandrasekaran (University of Texas at Austin), Adam Klivans
🎯 What it does: This paper proposes an efficient algorithm for learning O(log n) Veldand functions (juntas) on smooth Markov Random Fields (MRFs), combining unsupervised structure learning and a greedy supervised learning two-phase process.
Learning Latent Variable Models via Jarzynski-adjusted Langevin Algorithm
James Cuin (Imperial College London), O. Deniz Akyildiz (Imperial College London)
TabularBiomedical DataStochastic Differential Equation
🎯 What it does: Proposes an EM method based on the Jarzynski-adjusted Langevin algorithm (JALA-EM) to estimate the maximization of marginal likelihood for latent variable models.
Learning Linear Attention in Polynomial Time
Morris Yau (Massachusetts Institute of Technology), Jacob Andreas
OptimizationTransformerTabular
🎯 What it does: This study investigates the polynomial time learnability of single-layer multi-head linear attention networks and provides theoretical guarantees for strong, noise-free PAC learning.
Learning long range dependencies through time reversal symmetry breaking
Guillaume Pourcel (University of Groningen), Maxence Ernoult (Google DeepMind)
OptimizationComputational EfficiencyRecurrent Neural NetworkTime SeriesOrdinary Differential Equation
🎯 What it does: A forward learning algorithm based on Hamiltonian systems, RHEL, is proposed to train recursive state space models (SSM), achieving gradient computation without explicit Jacobian matrices and using only forward propagation.
Learning Memory-Enhanced Improvement Heuristics for Flexible Job Shop Scheduling
Jiaqi Wang (Jilin University), You Zhou (Jilin University)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This study investigates an improved heuristic method based on deep reinforcement learning to solve the flexible job shop scheduling problem.
Learning Multi-Source and Robust Representations for Continual Learning
Fei Ye (University of Electronic Science and Technology of China), shijie zhou
Representation LearningTransformerImage
🎯 What it does: A continuous learning framework LMSRR based on multi-source pre-trained ViT is proposed, which integrates multi-source features and enhances the model's adaptability to new tasks through dynamic fusion.
Learning Neural Exposure Fields for View Synthesis
Michael Niemeyer (Google), Federico Tombari (Google)
GenerationData SynthesisNeural Radiance FieldImage
🎯 What it does: This paper proposes Neural Exposure Fields (NExF), which learns the exposure field in 3D space and conditions on lighting, enabling the generation of high-quality, 3D-consistent view synthesis results under different exposure conditions.
Learning non-equilibrium diffusions with Schrödinger bridges: from exactly solvable to simulation-free
Stephen Y. Zhang (University of Melbourne), Michael Stumpf
Score-based ModelFlow-based ModelTime SeriesBiomedical DataPhysics RelatedStochastic Differential Equation
🎯 What it does: This paper studies the Schrödinger bridge problem in non-equilibrium systems, using multivariate Ornstein-Uhlenbeck (mvOU) as the reference process. It derives a closed-form analytical solution for Gaussian endpoints and proposes a simulation-free solver MVOU-OTFM based on score and flow matching to handle non-Gaussian endpoints.