ICLR 2026 Papers — Page 26
International Conference on Learning Representations · 5356 papers
Learnable Sparsity for Vision Generative Models
Yang Zhang (National University of Singapore), Kenji Kawaguchi (National University of Singapore)
GenerationComputational EfficiencyTransformerDiffusion modelImageText
🎯 What it does: Propose a learnable sparsity framework called EcoDiff to achieve structured pruning for visual generation models;
LearNAT: Learning NL2SQL with AST-guided Task Decomposition for Large Language Models
Weibin Liao (Peking University), Yasha Wang (Peking University)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextTabularBenchmark
🎯 What it does: This paper proposes the LearNAT framework, which enhances the performance of LLMs in NL2SQL tasks through AST-guided task splitting and margin reinforcement learning.
Learned Meta-Tokens for Language Modeling
Alok Shah, Pratik Chaudhari (University of Pennsylvania)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Injecting learnable meta-tokens and incorporating a dedicated meta-attention mechanism during the pre-training phase enables language models to compress and retrieve distant context, thereby achieving better length generalization.
Learning a distance measure from the information-estimation geometry of data
Guy Ohayon (Flatiron Institute), Eero P Simoncelli
RetrievalTransformerDiffusion modelImage
🎯 What it does: This paper proposes and implements a new distance metric based on the geometry of probability density—the Information-Estimation Metric (IEM), and verifies it in the task of image similarity assessment.
Learning a Game by Paying the Agents
Brian Hu Zhang (Massachusetts Institute of Technology), Tuomas Sandholm (Carnegie Mellon University)
OptimizationReinforcement Learning
🎯 What it does: By paying learners money and sending signals through multi-round interactions, the learner's utility function in unknown repeated standard-form games is learned, and the learned utility function is used to guide learners toward optimal cooperative equilibria.
Learning Adaptive Distribution Alignment with Neural Characteristic Function for Graph Domain Adaptation
Wei Chen (Beihang University), deqing wang
Domain AdaptationOptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: Proposed the ADAlign framework for graph domain adaptation through adaptive distribution alignment to achieve cross-domain knowledge transfer.
Learning Admissible Heuristics for A*: Theory and Practice
Ehsan Futuhi (University of Alberta), Nathan R. Sturtevant (University of Alberta)
OptimizationConvolutional Neural NetworkBenchmark
🎯 What it does: This paper proposes a framework for acceptable heuristic learning based on restricted optimization, designs a new cross-entropy acceptable loss function (CEA), and implements an acceptable and efficient heuristic function on multiple pattern databases (PDBs) of the 3×3 Rubik's cube.
Learning an Image Editing Model without Image Editing Pairs
Nupur Kumari (Carnegie Mellon University), Xun Huang (Adobe Research)
Image TranslationGenerationTransformerVision Language ModelDiffusion modelImageText
🎯 What it does: This paper proposes an unpaired training image editing model, which uses gradient feedback from a vision-language model (VLM) and distribution matching loss to perform few-step (4 steps) end-to-end optimization on a pre-trained diffusion model.
Learning AND–OR Templates for Compositional Representation in Art and Design
Liaoruxing Zhang (Beijing Electronic Science and Technology Institute), Song-Chun Zhu (Beijing Institute for General Artificial Intelligence)
Object DetectionSegmentationExplainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: Propose an interpretable AND-OR template based on the maximum entropy log-linear model for automatically learning structured representations in art and design, and provide a unified compatibility score;
Learning Boltzmann Generators via Constrained Mass Transport
Christopher von Klitzing (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)
GenerationFlow-based ModelBiomedical Data
🎯 What it does: Propose a training framework for Boltzmann generators based on constrained mass transport to address the problem of sampling from high-dimensional, multi-modal distributions.
Learning Brain Representation with Hierarchical Visual Embeddings
Jiawen Zheng (Hong Kong University of Science and Technology), Chen Liang (Hong Kong University of Science and Technology)
RestorationRetrievalRepresentation LearningTransformerVision Language ModelDiffusion modelAuto EncoderContrastive LearningImageBiomedical Data
🎯 What it does: Propose a brain-visual interface that aligns brain signals with multi-level visual embeddings, leveraging pre-trained fusion priors to achieve high-quality image reconstruction and retrieval
Learning Collective Variables from BioEmu with Time-Lagged Generation
Seonghyun Park (KAIST), Sungsoo Ahn (KAIST)
Drug DiscoveryProtein Structure PredictionDiffusion modelBiomedical Data
🎯 What it does: By adding a lightweight encoder to the frozen BioEmu generative model and training low-dimensional collective variables (CVs) using time-lagged conditional learning, the method is applied to OPES free energy estimation and SMD transition path sampling.
Learning Concept Bottleneck Models from Mechanistic Explanations
Antonio De Santis (Politecnico di Milano), Lalana Kagal (MIT CSAIL)
ClassificationExplainability and InterpretabilityLarge Language ModelAuto EncoderImageMultimodality
🎯 What it does: Designed and implemented the Mechanistic Concept Bottleneck Model (M-CBM), which first extracts interpretable concepts from a black-box model using a sparse autoencoder, then automatically names and partially annotates images with concepts using a multimodal large language model, and finally trains the CBM for image classification.
Learning Correlated Reward Models: Statistical Barriers and Opportunities
Yeshwanth Cherapanamjeri (Massachusetts Institute of Technology), Sobhan Mohammadpour (Massachusetts Institute of Technology)
Recommendation SystemTabular
🎯 What it does: This paper studies how to learn a Probit model without the IIA assumption from user preference data, and proves that pairwise comparisons alone cannot identify the covariance structure; it proposes using best-of-three preference data to achieve model identifiability and statistical estimation.
Learning Data-Efficient and Generalizable Neural Operators via Fundamental Physics Knowledge
Siying Ma (Simon Fraser University), Vijay Ganesh (Georgia Institute of Technology)
Physics Related
🎯 What it does: Propose a multi-physics training framework that enhances the data efficiency and generalization capability of neural operators in physical simulations by simultaneously learning the original PDE and its simplified fundamental form.
Learning Distributions over Permutations and Rankings with Factorized Representations
Daniel Severo (FAIR at Meta), Niklas Nolte (FAIR at Meta)
Recommendation SystemRepresentation LearningTransformerImageTabular
🎯 What it does: Propose to learn permutations over any probability distribution by utilizing standard MLM/AR training methods through three reversible representations: Lehmer codes, Fisher-Yates sampling, and insertion vectors.
Learning Domain-Aware Task Prompt Representations for Multi-Domain All-in-One Image Restoration
Guanglu Dong (Sichuan University), Lichao Mou (Medai Technology Wuxi Co Ltd)
RestorationDomain AdaptationKnowledge DistillationLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageBiomedical DataMagnetic Resonance ImagingComputed TomographyPositron Emission Tomography
🎯 What it does: Propose the first multi-domain universal image restoration method DATPRL-IR, which utilizes dual Prompt pools for task and domain to achieve domain-aware task prompting representation, unifying the processing of multi-task image restoration across natural, medical, remote sensing, and other domains.
Learning Dynamic Causal Graphs Under Parametric Uncertainty via Polynomial Chaos Expansions
Liang Cao (University of British Columbia)
OptimizationExplainability and InterpretabilityTabular
🎯 What it does: Proposed and implemented the PCT-CD framework, which can learn dynamic causal graphs that change with system parameters, modeling the weight of each edge as a parameter function and achieving interpretable causal mechanisms through polynomial chaos expansion.
Learning Dynamics Feature Representation via Policy Attention for Dynamic Path Planning in Urban Road Networks
Kai Zhang, Qiuhong Wang (Nanjing University of Aeronautics and Astronautics)
Autonomous DrivingOptimizationComputational EfficiencyGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper proposes a Dynamic Feature Representation (DFR) framework for dynamic path planning in urban road networks under reinforcement learning;
Learning Dynamics of Logits Debiasing for Long-Tailed Semi-Supervised Learning
Yue Cheng (Beijing Jiaotong University), Zhanxing Zhu (Beijing Jiaotong University)
ClassificationData-Centric LearningTransformerImageBenchmark
🎯 What it does: This paper provides a theoretical analysis of logit debiasing methods in long-tailed semi-supervised learning from the perspective of learning dynamics, and proposes a dynamic pruning framework called DyTrim to alleviate bias caused by class imbalance.
Learning Efficient and Interpretable Multi-Agent Communication
Wei Du (Shandong University), Lizhen Cui (Shandong University)
Explainability and InterpretabilityLarge Language ModelAuto EncoderContrastive LearningBenchmark
🎯 What it does: Proposes the GLC framework to learn efficient and interpretable multi-agent communication protocols.
Learning Escorted Protocols For Multistate Free-Energy Estimation
Lars Holdijk (University of Oxford), Max Welling (University of Amsterdam)
Graph Neural NetworkFlow-based ModelGraphPhysics RelatedStochastic Differential Equation
🎯 What it does: Proposed a learning framework based on conditional flow matching and conditional density matching, capable of adaptively learning switching protocols for escaped non-equilibrium free energy estimation;
Learning Explicit Single-Cell Dynamics Using ODE Representations
Jan-Philipp von Bassewitz (Institute of Science and Technology Austria), Francesco Locatello (Institute of Science and Technology Austria)
Explainability and InterpretabilityRepresentation LearningBiomedical DataOrdinary Differential Equation
🎯 What it does: Proposes Cell-MNN, an end-to-end encoder-decoder architecture that employs a local linear ODE to represent cell differentiation dynamics and directly learns and interprets gene regulatory interactions through this ODE.
Learning Exposure Mapping Functions for Inferring Heterogeneous Peer Effects
Shishir Adhikari (University of Illinois Chicago), Elena Zheleva (University of Illinois Chicago)
Representation LearningGraph Neural NetworkAuto EncoderGraph
🎯 What it does: Under network interventions, a framework is proposed to automatically learn the exposure mapping function, utilizing the graph neural network EGONETGNN to infer heterogeneous peer effects;
Learning Facts at Scale with Active Reading
Jessy Lin (FAIR at Meta), Barlas Oguz (FAIR at Meta)
Data SynthesisData-Centric LearningTransformerLarge Language ModelTextFinance Related
🎯 What it does: This paper proposes the Active Reading framework, which enhances the learning and recall of factual knowledge by enabling LLMs to self-generate diverse learning strategies (such as rewriting, knowledge linking, active recall, analogical reasoning, etc.) to create self-generated training data from given documents.
Learning Flexible Forward Trajectories for Masked Molecular Diffusion
Hyunjin Seo (Korea Advanced Institute of Science and Technology), Sungsoo Ahn (Korea Advanced Institute of Science and Technology)
Drug DiscoveryTransformerDiffusion modelGraph
🎯 What it does: Proposed a masked diffusion model called MELD for unconditional and conditional generation of molecular graphs
Learning for Highly Faithful Explainability
Yuhan Guo (Beijing Institute of Technology), Guoren Wang (Beijing Institute of Technology)
Explainability and InterpretabilityTransformerImageTextTabular
🎯 What it does: Designed and implemented DeepFaith, an amortized interpreter based on the faithfulness of explanations;
Learning from Algorithm Feedback: One-Shot SAT Solver Guidance with GNNs
Jan Tönshoff (RWTH Aachen University), Martin Grohe (RWTH Aachen University)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: Propose the RLAF paradigm, which utilizes a single forward pass of a GNN to predict variable weights and polarity for a SAT solver, thereby guiding branching decisions;
Learning From Dictionary: Enhancing Robustness of Machine-Generated Text Detection in Zero-Shot Language via Adversarial Training
Yuanfan Li (Xi'an Jiaotong University), Zexuan Xie (Xi'an Jiaotong University)
Anomaly DetectionAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Propose a dictionary-based adversarial training framework called TASTE to enhance the robustness and attack resistance of multilingual machine-generated text detectors on zero-shot languages.
Learning from Historical Activations in Graph Neural Networks
Yaniv Galron (Technion Israel Institute of Technology), Moshe Eliasof (Ben Gurion University of Negev)
Representation LearningDrug DiscoveryGraph Neural NetworkGraphBiomedical Data
🎯 What it does: This paper proposes a graph pooling layer called HISTOGRAPH, which aggregates historical activations from different layers of graph neural networks through a two-stage attention mechanism to generate more informative graph-level representations.
Learning from Label Proportions via Proportional Value Classification
Tianhao Ma (University of Tokyo), Masashi Sugiyama (RIKEN)
ClassificationImage
🎯 What it does: By introducing a Proportional Value Classification (PVC) task, the method learns instance-level classifiers by aggregating predicted results of instances in a bag into proportional values, thereby achieving the goal of classification from label proportion data.
Learning from Noisy Preferences: A Semi-Supervised Learning Approach to Direct Preference Optimization
Xinxin Liu (University of Central Florida), Chen Chen (University of Central Florida)
GenerationOptimizationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImageBenchmark
🎯 What it does: Address multi-dimensional visual preference label noise through a semi-supervised learning framework (Semi-DPO), improving the alignment of Diffusion-DPO
Learning from Synthetic Data Improves Multi-hop Reasoning
Anmol Kabra (Cornell University), Kilian Q Weinberger
Data SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Fine-tune large language models using reinforcement learning (RLVR) on a fully verifiable, rule-generated synthetic dataset to enhance their multi-hop reasoning capabilities.
Learning from the Electronic Structure of Molecules across the Periodic Table
Manasa Kaniselvan (ETH Zurich), Daniel S. Levine (Meta)
Graph Neural NetworkSupervised Fine-TuningGraphPhysics Related
🎯 What it does: This paper proposes the HELM model based on equivariant graph neural networks, which can directly predict the complete Hamiltonian matrix and further pretrain to enhance energy prediction performance, even for molecules with more than 100 atoms, up to 58 elements, and using the def2-TZVPD basis set.
Learning From the Past with Cascading Eligibility Traces
Tokiniaina Raharison Ralambomihanta (Mila Quebec Artificial Intelligence Institute), Blake Aaron Richards (Mila Quebec Artificial Intelligence Institute)
ClassificationRecurrent Neural NetworkReinforcement LearningImage
🎯 What it does: Propose and verify a new delay-tolerant memory mechanism—Cascading Eligibility Traces (CET)—for precise credit assignment in biological learning with fixed delays.
Learning Global Hypothesis Space for Enhancing Synergistic Reasoning Chain
Jiaquan Zhang (University of Electronic Science and Technology of China), Yang Yang (University of Electronic Science and Technology of China)
Explainability and InterpretabilityGraph Neural NetworkLarge Language ModelAgentic AITextChain-of-Thought
🎯 What it does: Proposes a two-stage Global Hypothesis Space and Topological Data Analysis (GHS-TDA) framework to integrate multi-path reasoning results and extract robust reasoning skeletons and self-consistent loops, thereby enhancing the reasoning accuracy and explainability of large language models.
Learning Heterogeneous Degradation Representation for Real-World Super-Resolution
Haowei Li (Sun Yat-sen University), Liang Lin (Sun Yat-sen University)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: Propose a spatially amortized variational learning framework (SAVL) that learns spatially heterogeneous degradation representations using per-pixel Gaussian posteriors and embeds them into an SR network for degradation-aware reconstruction.
Learning Hierarchical and Geometry-Aware Graph Representations for Text-to-CAD
Shengjie Gong (South China University of Technology), Tianshui Chen (Guangdong University of Technology)
GenerationData SynthesisGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextGraphBenchmark
🎯 What it does: Propose the Graph-CAD framework, implementing a three-stage pipeline for text-to-CAD programs by first generating a geometric decomposition graph, then planning action sequences, and finally generating executable bpy code.
Learning Human Habits with Rule-Guided Active Inference
GONG ZHIREN (Chinese University of Hong Kong), Shuang Li (Chinese University of Hong Kong)
OptimizationExplainability and InterpretabilityReinforcement LearningWorld ModelSequentialBiomedical Data
🎯 What it does: This paper proposes a rule-guided active inference framework for learning and executing human habits.
Learning in Prophet Inequalities with Noisy Observations
Jung-hun Kim (CREST, ENSAE, IP Paris), Vianney Perchet (CREST, ENSAE, IP Paris)
OptimizationTabular
🎯 What it does: Studied oracle inequalities under scenarios with only noisy observations and unknown distributions, and proposed a stopping strategy based on learning and LCB thresholds.
Learning is Forgetting; LLM Training As Lossy Compression
Henry Conklin (Princeton University), Seraphina Goldfarb-Tarrant (Cohere)
CompressionRepresentation LearningLarge Language ModelText
🎯 What it does: This paper views the training of large language models as a lossy compression process and quantifies the evolution of the representation space during pre-training through the Information Bottleneck (IB) theory.
Learning Koopman Representations with Controllability Guarantees
Keyan Miao (University of Oxford), Antonis Papachristodoulou (University of Oxford)
OptimizationRepresentation LearningTabularBiomedical DataBenchmarkOrdinary Differential Equation
🎯 What it does: This paper proposes a neural ODE framework based on Koopman representation that can directly guarantee controllability when learning nonlinear dynamics;
Learning linear state-space models with sparse system matrices
Yasen Wang (China Telecom Research Institute), Gang Lu (China Telecom Research Institute)
OptimizationTabularTime Series
🎯 What it does: Learning sparse linear state space models and recovering their topological structures
Learning Massively Multitask World Models for Continuous Control
Nicklas Hansen (University of California San Diego), Xiaolong Wang (University of California San Diego)
Robotic IntelligenceTransformerReinforcement LearningVision Language ModelVision-Language-Action ModelWorld ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed the MMBench large-scale multi-task reinforcement learning benchmark, constructing 200 continuous control tasks; developed the Newt multi-task world model, which trains a single agent across all tasks through online RL under language instructions and optional visual inputs, combined with four demonstration utilization methods: demonstration pre-training, behavior cloning, and planning constraints.
Learning Mixtures of Linear Dynamical Systems via Hybrid Tensor-EM Method
Lulu Gong (Yale University), Shreya Saxena (Yale University)
Biomedical Data
🎯 What it does: Propose a hybrid tensor-EM framework for learning Mixture of Linear Dynamical Systems (MoLDS), enabling modeling and clustering of multi-trajectory heterogeneity.
Learning Molecular Chirality via Chiral Determinant Kernels
Runhan Shi (Shanghai Jiao Tong University), Yang Yang (Shanghai Jiao Tong University)
Drug DiscoveryGraph Neural NetworkTransformerGraphBiomedical Data
🎯 What it does: A framework named ChiDeK was studied, which learns molecular chirality representations through chiral determinant nuclei and cross-attention mechanisms, capable of simultaneously handling central and axial chirality.
Learning More with Less: A Dynamic Dual-Level Down-Sampling Framework for Efficient Policy Optimization
Chao Wang (Tsinghua University), Wenbo Ding (Tsinghua University)
OptimizationTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes a Dynamic Double-layer Downsampling Framework (DS³), which efficiently trains large language models' policy optimization by maximizing advantage variance at the sample layer and combining entropy with advantage product at the token layer.
Learning multimodal dictionary decompositions with group-sparse autoencoders
Chiraag Kaushik (Georgia Institute of Technology), Andrea Fanelli (Dolby Laboratories)
ClassificationRetrievalRepresentation LearningAuto EncoderImageTextMultimodalityAudio
🎯 What it does: This paper proposes a sparse autoencoder based on group sparse regularization and cross-modal random masking (GSAE/MGSAE) to learn multi-modal concept dictionaries from multi-modal embedding spaces.
Learning Nonlinear Causal Reductions to Explain Reinforcement Learning Policies
Armin Kekić (Max Planck Institute for Intelligent Systems), Michel Besserve (Max Planck Institute for Intelligent Systems)
Explainability and InterpretabilityReinforcement LearningTabularTime Series
🎯 What it does: Investigated a nonlinear target causal reduction (nTCR), which explains the overall behavior of policies and reveals key behavioral patterns and failure modes by injecting random perturbations into trained RL policy actions and learning high-level causal models.
Learning of Population Dynamics: Inverse Optimization Meets JKO Scheme
Mikhail Persiianov (Applied AI Institute), Alexander Korotin (Applied AI Institute)
OptimizationRepresentation LearningTime SeriesBiomedical Data
🎯 What it does: This paper proposes a method called iJKOnet, which combines inverse optimization with the JKO scheme to recover the energy function governing the evolution of dominant particles from discrete-time snapshots of population distribution, thereby learning population dynamics.
Learning on a Razor’s Edge: Identifiability and Singularity of Polynomial Neural Networks
Vahid Shahverdi (KTH Royal Institute of Technology), Kathlén Kohn (KTH Royal Institute of Technology)
Explainability and InterpretabilityConvolutional Neural Network
🎯 What it does: Theoretical study on the geometric properties of neuromanifolds in multi-layer perceptrons (MLPs) and convolutional neural networks (CNNs) using polynomial activation functions, focusing on identifiability, singularities, and exposure properties.
Learning Ordinal Probabilistic Reward from Preferences
Longze Chen (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences), Min Yang (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose an Ordinal Probabilistic Reward Model (OPRM) that models rewards by learning the full probability distribution of response quality; simultaneously design Region Flooding Tuning (RgFT) to calibrate the distribution using a small number of absolute quality labels.
Learning Patient-Specific Disease Dynamics With Latent Flow Matching For Longitudinal Imaging Generation
Hao Chen (University Of Cambridge), Chao Li (University Of Cambridge)
GenerationData SynthesisFlow-based ModelAuto EncoderBiomedical DataMagnetic Resonance ImagingAlzheimer's DiseaseOrdinary Differential Equation
🎯 What it does: Investigated how to achieve continuous, interpretable disease progression simulation by learning patient-specific disease trajectories in latent space and generating future time point MRIs using flow matching.
Learning Physics-Grounded 4D Dynamics with Neural Gaussian Force Fields
Shiqian Li (Peking University), Yixin Zhu (Peking University)
GenerationData SynthesisTransformerGaussian SplattingVideoPoint CloudPhysics RelatedOrdinary Differential Equation
🎯 What it does: Proposes the Neural Gaussian Force Field (NGFF) framework, which integrates multi-view RGB input 3D Gaussian representations with neural force field dynamics to generate physically consistent 4D videos.
Learning Posterior Predictive Distributions for Node Classification from Synthetic Graph Priors
Jeongwhan Choi (KAIST), Noseong Park (KAIST)
ClassificationData SynthesisGraph Neural NetworkTransformerGraph
🎯 What it does: A general node classification model called NodePFN based on a prior fitting network (PFN) is constructed, which can predict unmarked nodes in any graph by only using marked nodes in the context without training separately for each graph;
Learning Pseudorandom Numbers with Transformers: Permuted Congruential Generators, Curricula, and Interpretability
Tao Tao (University of Maryland), Maissam Barkeshli (University of Maryland)
Explainability and InterpretabilityTransformerLarge Language ModelSequential
🎯 What it does: Investigate Transformer's in-context prediction on Permuted Congruential Generators (PCG) sequences, demonstrating that the model can generalize to unseen (a, c) values without requiring parameter-specific knowledge;
Learning Recursive Multi-Scale Representations for Irregular Multivariate Time Series Forecasting
Boyuan Li (South China University of Technology), Qianli Ma (South China University of Technology)
Computational EfficiencyRepresentation LearningTime SeriesElectronic Health Records
🎯 What it does: This study focuses on the prediction problem of irregular multivariate time series (IMTS) and proposes a recursive multi-scale learning framework called ReIMTS, which can recursively split samples, learn multi-scale representations, and capture global-to-local dependencies through irregular-aware fusion while preserving the original sampling timestamps.
Learning residue level protein dynamics with multiscale Gaussians
Mihir Bafna (Massachusetts Institute of Technology), Bonnie Berger (Massachusetts Institute of Technology)
Protein Structure PredictionTransformerBiomedical Data
🎯 What it does: Designed and implemented a lightweight, SE(3) invariant framework called DYNAPROT for directly predicting high-order Gaussian distributions of protein dynamics from static protein structures, including per-residue 3×3 covariance matrices and global N×N scalar coupling matrices.
Learning Retrieval Models with Sparse Autoencoders
Thibault Formal (Naver Labs Europe), Stéphane Clinchant (Naver Labs Europe)
RetrievalKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderText
🎯 What it does: Propose SPLARE, which integrates a pre-trained sparse autoencoder (SAE) into LLM to generate a latent vocabulary space, achieving efficient sparse retrieval.
Learning Robust Intervention Representations with Delta Embeddings
Panagiotis Alimisis (Harokopio University of Athens), Christos Diou (Harokopio University of Athens)
Explainability and InterpretabilityRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes the Causal Delta Embedding (CDE) framework, which directly represents intervention actions by learning the potential space difference between image pairs, achieving interpretable, sparse, and scene-agnostic representations of actions.
Learning Self-Critiquing Mechanisms for Region-Guided Chest X-Ray Report Generation
Sixing Yan (Guangxi Rural Commercial United Bank Co., Ltd), Simon See (NVIDIA Corporation)
GenerationRetrievalAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningImageTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Propose the RadSCR framework, which employs a multi-faceted self-critical mechanism to achieve reliable anomaly localization and report generation in chest X-ray report generation.
Learning Semi-Structured Sparsity for LLMs via Shared and Context-Aware Hypernetwork
Lu Sun (RIKEN), Jun Sakuma (RIKEN)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Designed and implemented HyperPrune, an n:m semi-structured sparsification framework based on a shared context-aware hypernetwork, for efficiently compressing large language models.
Learning Shrinks the Hard Tail: Training‑Dependent Inference Scaling in a Solvable Linear Model
Noam Itzhak Levi (École Polytechnique Fédérale de Lausanne)
Computational EfficiencyImageText
🎯 What it does: Analyzed the impact of instance difficulty on training and inference scaling in linear models, proposed the Latent Instance Difficulty (LID) model, and derived the training-related inference exponent β_eff.
Learning Structure-Semantic Evolution Trajectories for Graph Domain Adaptation
Wei Chen (Beihang University), HUIMEI HE
Domain AdaptationGraph Neural NetworkDiffusion modelGraphStochastic Differential Equation
🎯 What it does: Propose a graph domain adaptation method called DiffGDA based on a continuous diffusion process, which directly transforms the source graph into the target graph step by step.
Learning Survival Distributions with Individually Calibrated Asymmetric Laplace Distribution
Deming Sheng (Duke University), Ricardo Henao (Duke University)
OptimizationExplainability and InterpretabilityTabularBiomedical DataFinance RelatedPhysics Related
🎯 What it does: Propose a survival analysis framework based on Individually Calibrated Asymmetric Laplace Distribution (ICALD), supporting pre-calibration and post-calibration, unifying parametric and non-parametric ALD methods;
Learning the Inverse Temperature of Ising Models under Hard Constraints using One Sample
Rohan Chauhan (University of California, Irvine), Ioannis Panageas (University of California, Irvine)
GraphPhysics Related
🎯 What it does: Studied the problem of estimating the single-sample inverse temperature parameter in a truncated Ising model under hard constraints (k-SAT satisfiable set)
Learning to Adapt: In-Context Learning Beyond Stationarity
Zhen Qin (University of Michigan), Zhihui Zhu (Ohio State University)
OptimizationMeta LearningTransformerTextTime Series
🎯 What it does: Proposes a theoretical analysis of context learning under non-stationary regression problems, achieving adaptive backtracking bias through gated linear attention (GLA).
Learning to Answer from Correct Demonstrations
Nirmit Joshi (Toyota Technological Institute at Chicago), Nathan Srebro (Toyota Technological Institute at Chicago)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: Studied the process of learning to generate question-answer pairs from correct demonstrations, proposing an imitation learning method aimed at learning how to answer questions through offline demonstrations.
Learning to Be Uncertain: Pre-training World Models with Horizon-Calibrated Uncertainty
Shenghua Wan (Nanjing University), De-Chuan Zhan (Nanjing University of Science and Technology)
Representation LearningRobotic IntelligenceRecurrent Neural NetworkReinforcement LearningAuto EncoderWorld ModelVideo
🎯 What it does: This paper proposes a pre-training framework called HAUWM, based on variable-length prediction and Horizon-Calibrated Uncertainty (HCU) loss, aiming to enable world models to learn temporally increasing prediction uncertainty from action-free video data;
Learning To Draft: Adaptive Speculative Decoding with Reinforcement Learning
Jiebin Zhang (Peking University), Sujian Li (Peking University)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes Learning to Draft (LTD), which uses reinforcement learning to dynamically adjust draft depth and validation scale to maximize throughput per draft-validation cycle, thereby accelerating speculative decoding.
Learning to Generate Stylized Handwritten Text via a Unified Representation of Style, Content, and Noise
Honglie Wang (University of Chinese Academy of Sciences), Cheng-Lin Liu (University of Chinese Academy of Sciences)
Image TranslationGenerationData SynthesisTransformerDiffusion modelAuto EncoderText
🎯 What it does: Propose a unified diffusion Transformer model called InkSpire, which generates and performs fine-grained editing of multi-line handwritten text through a multi-line occlusion filling strategy without relying on independent style/content encoders.
Learning to Generate Unit Test via Adversarial Reinforcement Learning
Dongjun Lee (KAIST), Kimin Lee (Microsoft Research)
AI Code AssistantTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose an adversarial reinforcement learning framework called UTRL to train large language models (LLMs) to automatically generate high-quality unit tests.
Learning to Grasp Anything By Playing with Random Toys
Dantong Niu (University Of California Berkeley), Roei Herzig (University Of California Berkeley)
Robotic IntelligenceTransformerVision-Language-Action ModelImage
🎯 What it does: This paper demonstrates that a robot can successfully grasp real-world objects in a zero-shot setting by training only on small toys composed of four basic geometric shapes randomly combined.
Learning to Interpret Weight Differences in Language Models
Avichal Goel (Massachusetts Institute of Technology), Tony T. Wang (Massachusetts Institute of Technology)
Data SynthesisExplainability and InterpretabilityTransformerSupervised Fine-TuningText
🎯 What it does: This study proposes a Diff Interpretation Tuning (DIT) method, which utilizes LoRA low-rank adapters to enable fine-tuned language models to describe in natural language the behavioral changes caused by weight differences;
Learning to Lie: Adversarial Attacks on Human-AI Teams and LLMs
Abed K. Musaffar (University of California at Santa Barbara), Francesco Bullo (University of California at Santa Barbara)
Adversarial AttackTransformerLarge Language ModelReinforcement LearningTabular
🎯 What it does: The study constructs a human-AI hybrid team decision-making experiment in an IQ quiz game to explore the impact of malicious AI attacks on team performance.
Learning to Orchestrate Agents in Natural Language with the Conductor
Stefan Nielsen (Sakana AI), Yujin Tang (Sakana AI)
AI Code AssistantLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Proposed and trained the RL Conductor model, using reinforcement learning to automatically design multi-model collaborative workflows, assign subtasks, and establish communication topologies to enhance overall inference performance.
Learning to Parallel: Accelerating Diffusion Large Language Models via Learnable Parallel Decoding
Wenrui Bao (University of Central Florida), Yuzhang Shang (HKUST)
Computational EfficiencyTransformerLarge Language ModelDiffusion modelTextBenchmark
🎯 What it does: To accelerate the inference of diffusion-based large language models, this paper proposes the Learn2PD framework, which uses a lightweight filtering model to predict when to terminate token repetition decoding and introduces the EoTP mechanism to prematurely stop [END] subsequent decoding.
Learning to Play Multi-Follower Bayesian Stackelberg Games
Gerson Personnat (Harvard University), David C. Parkes (Harvard University)
OptimizationReinforcement Learning
🎯 What it does: This paper studies the online learning problem in multi-follower Bayesian Stackelberg games, providing learning algorithms under two feedback models: observing follower types (type feedback) and observing follower actions (action feedback). It also establishes upper and lower bounds on the long-term loss (regret) of the scheduling strategies generated by these algorithms.
Learning to Reason as Action Abstractions with Scalable Mid-Training RL
Shenao Zhang (Northwestern University), Zirui Wang (Apple)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes a new mid-training algorithm RA3 (Reasoning as Action Abstractions), which improves the performance and convergence speed of large language models in the subsequent reinforcement learning (RLVR) phase by learning temporal action abstractions.
Learning to Reason Efficiently with Discounted Reinforcement Learning
Alex Ayoub (Amazon), Karim Bouyarmane (Amazon)
Computational EfficiencyLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes a reinforcement learning method that discounts rewards for reasoning tokens to shorten the reasoning chains of large language models while maintaining answer accuracy.
Learning to Reason for Hallucination Span Detection
Hsuan Su (National Taiwan University), Raviteja Vemulapalli (Apple)
Anomaly DetectionTransformerLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a reinforcement learning-based Hallucination Span Detection framework, RL4HS, which trains LLMs using Chain-of-Thought (CoT) reasoning and Span-F1 reward to precisely locate hallucination spans in generated text.
Learning to Reason in Structured In-context Environments with Reinforcement Learning
Peng Yu (Shanghai Jiao Tong University), Ying Wen (Shanghai Jiao Tong University)
Graph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the Structured In-context Environment (SIE) framework, which automatically constructs scalable and verifiable reinforcement learning (RL) training environments from large-scale structured knowledge graphs, enabling large language models (LLMs) to enhance structured reasoning capabilities and achieve cross-domain generalization in RL.
Learning to Reason over Continuous Tokens with Reinforcement Learning
Yiran Zhao (Salesforce AI Research), Junnan Li (Salesforce AI Research)
Explainability and InterpretabilityComputational EfficiencyTransformerSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose the HyRea framework, which allows LLMs to dynamically switch between explicit (token-based) and implicit (embedding-based) reasoning modes during inference to achieve a balance between reasoning efficiency and accuracy.
Learning to Reason via Mixture-of-Thought for Logical Reasoning
Tong Zheng (University of Maryland), Heng Huang (University of Maryland)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose the Mixture-of-Thought (MoT) framework, enabling large language models to perform logical reasoning in three complementary reasoning modes: natural language, code, and truth tables.
Learning to Reason without External Rewards
Xuandong Zhao (University of California Berkeley), Dawn Song (University of California Berkeley)
Reinforcement LearningText
🎯 What it does: This paper proposes a reinforcement learning framework RLIF without external rewards and implements the INTUITOR method based on model self-certainty (self-certainty).
Learning to Recall with Transformers Beyond Orthogonal Embeddings
Nuri Mert Vural (University of Toronto), Denny Wu (Flatiron Institute)
Representation LearningTransformerSequential
🎯 What it does: Analyzes the learning dynamics and capacity limits of a single-layer Transformer using gradient descent in a simple fact recall task under random embeddings and finite samples;
Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training
Junlin Han (Meta Superintelligence Labs), Filippos Kokkinos (Meta Superintelligence Labs)
Representation LearningHyperparameter SearchData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodalityBenchmark
🎯 What it does: This paper systematically studies the visual priors acquired by large language models (LLMs) through text-only pre-training, decomposing them into perceptual priors and reasoning priors; based on this, a data mixing strategy is proposed to maximize visual priors and validate its effectiveness across different scales; meanwhile, the authors construct a multi-layer existence benchmark (MLE-Bench) to evaluate visual perception capabilities.
Learning to Segment for Vehicle Routing Problems
Wenbin Ouyang (MIT), Cathy Wu (MIT)
OptimizationComputational EfficiencyGraph Neural NetworkTransformerGraph
🎯 What it does: Accelerate iterative search solvers for large vehicle routing problems (VRP) by proposing a segment-first then aggregation (FSTA) decomposition technique, and dynamically reduce the search space by learning to identify stable and unstable road segments.
Learning to Solve Orienteering Problem with Time Windows and Variable Profits
Songqun Gao (Università di Trento), Daniele Fontanelli (Università di Trento)
OptimizationTransformerReinforcement LearningBenchmark
🎯 What it does: Propose a two-stage learning framework called DeCoST, which separately solves path planning and service time allocation to address the Orienteering Problem with Time Windows and Variable Profits (OPTWVP).
Learning to Summarize by Learning to Quiz: Adversarial Agentic Collaboration for Long Document Summarization
Weixuan Wang (University of Edinburgh), Alexandra Birch (University of Edinburgh)
GenerationLarge Language ModelAgentic AIGenerative Adversarial NetworkTextBenchmark
🎯 What it does: Proposes SUMMQ, an adversarial multi-agent framework that enhances long document summarization quality through specialized abstract generator/reviewer and quiz generator/reviewer working collaboratively.
Learning to summarize user information for personalized reinforcement learning from human feedback
HyunJi Nam, Natasha Jaques (University of Washington)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Built a joint training framework (PLUS) that learns text summarization through reinforcement learning to capture user preferences, and uses the summary to condition the reward model, achieving multi-user personalized RLHF.
Learning to Weight Parameters for Training Data Attribution
Shuangqi Li (EPFL), Mathieu Salzmann (EPFL)
GenerationExplainability and InterpretabilityTransformerSupervised Fine-TuningDiffusion modelImageText
🎯 What it does: This paper studies gradient-based data attribution methods, proposing to improve attribution quality by learning parameter group importance weights, and supports fine-grained semantic attribution for image generation tasks (e.g., subject, style, background).
Learning under Quantization for High-Dimensional Linear Regression
Dechen Zhang (University of Hong Kong), Difan Zou (University of Hong Kong)
OptimizationTabular
🎯 What it does: This paper studies the use of low-precision quantized stochastic gradient descent (SGD) in high-dimensional linear regression and provides a theoretical analysis of quantization's impact on learning performance;
Learning Unified Representation of 3D Gaussian Splatting
Yuelin Xin (UC Irvine), Xinke Li (City University of Hong Kong)
GenerationRepresentation LearningAuto EncoderGaussian SplattingPoint Cloud
🎯 What it does: Propose a vectorized representation based on an equiprobable ellipsoidal submanifold, replacing traditional 3D Gaussian parameterization;
Learning Video Generation for Robotic Manipulation with Collaborative Trajectory Control
Xiao Fu, Dahua Lin
GenerationRobotic IntelligenceTransformerDiffusion modelVideo
🎯 What it does: Proposed a collaborative trajectory-based robot manipulation video generation framework called RoboMaster, which can generate multi-stage interactive videos from initial frames, prompts, trajectories, and optional object masks.
Learning What Matters Now: Dynamic Preference Inference under Contextual Shifts
Xianwei Cao (Xidian University), Shuang Wang (Xidian University)
Reinforcement Learning
🎯 What it does: In non-stationary multi-objective environments, we propose a dynamic preference inference framework (DPI) that adapts to environmental changes through online inference of preference weights.
Learning What Reinforcement Learning Can't: Interleaved Online Fine-Tuning for Hardest Questions
Lu Ma (Peking University), Bin CUI
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Propose a training framework called ReLIFT that interleaves reinforcement learning with online fine-tuning to enhance the reasoning capabilities of large language models, especially for solving the hardest problems.
Learning with Dual-level Noisy Correspondence for Multi-modal Entity Alignment
Haobin Li (Sichuan University), Xi Peng (Sichuan University)
Representation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposed a robust multi-modal entity alignment method that models and eliminates dual-layer noise correspondence (entity-attribute and cross-graph entity/attribute mismatch).
Learning-Augmented Moment Estimation on Time-Decay Models
Soham Deepak Nagawanshi (Texas A&M University), Samson Zhou (Texas A&M University)
OptimizationComputational EfficiencyTabularTime Series
🎯 What it does: This paper proposes a learning-enhanced algorithm for Fp frequency, rectangular Fp frequency, and cascaded norm estimation in time-decaying flow models (including sliding window, power-law decay, and exponential decay), leveraging a suffix-compatible heavy-hitter oracle capable of predicting tail frequencies, achieving almost optimal space complexity;
Learning-Time Encoding Shapes Unlearning in LLMs
Ruihan Wu (University of California San Diego), Kamalika Chaudhuri (University of California San Diego)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper investigates how different knowledge encoding methods during the training phase of large language models (LLMs), such as multiple paraphrasing and text block structures, affect post-training unlearning performance through controlled experiments.