ICLR 2026 Papers — Page 27
International Conference on Learning Representations · 5356 papers
LearnIR: Learnable Posterior Sampling for Real-World Image Restoration
Yihang Bao (Guilin University of Electronic Technology), Zengmin Xu (Guilin University of Electronic Technology)
RestorationDiffusion modelImage
🎯 What it does: Propose the LearnIR framework, combining diffusion models with posterior sampling correction to address tasks such as real image dehazing and deshadowing;
LearnPruner: Rethinking Attention-based Token Pruning in Vision Language Models
Rinyoichi Takezoe (Li Auto Inc), Kaiwen Long (Li Auto Inc)
OptimizationComputational EfficiencyRepresentation LearningLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: Propose LearnPruner, a two-stage visual token pruning framework for vision-language models;
LEGACY: A Lightweight Dynamic Gradient Compression Strategy for Distributed Deep Learning
Mostapha Essoullami (Mohammed VI Polytechnic University), Aritra Dutta (University of Central Florida)
CompressionOptimizationFederated LearningTransformerImageTextTabular
🎯 What it does: Proposed LEGACY—a lightweight, pluggable dynamic gradient compression framework that automatically adjusts compression parameters based on layer size and training stage in distributed deep learning training, compatible with any compressor.
LEGATO: Large-scale End-to-end Generalizable Approach to Typeset OMR
Guang Yang (University of Washington), Noah A. Smith (University of Washington)
RecognitionData SynthesisTransformerVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed an end-to-end optical music recognition (OMR) model called Legato, capable of recognizing multi-page, realistically formatted musical score images and generating ABC notation format.
LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models
Ruilin Yao (Wuhan University of Technology), Luc Van Gool (INSAIT Sofia University St Kliment Ohridski)
Large Language ModelPrompt EngineeringMixture of ExpertsVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed the Lens multi-level multi-modal reasoning evaluation benchmark (containing 3.4K images, 60K+ human-annotated QA pairs, 8 tasks) and the SMEC self-driven multi-expert collaborative reasoning framework, and conducted systematic evaluations on 15+ MLLMs released in 2025-2026.
LeRobot: An Open-Source Library for End-to-End Robot Learning
Remi Cadene (Hugging Face), Thomas Wolf (Hugging Face)
Robotic IntelligenceReinforcement LearningDiffusion modelFlow-based Model
🎯 What it does: Proposed an open-source library called lerobot, integrating robot low-level control, data collection and storage, unified data format LeRobotDataset, asynchronous inference stack, and various state-of-the-art robot learning algorithms, achieving end-to-end robot learning workflow;
Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting
Yixing Lao (Hong Kong University), Hengshuang Zhao (Hong Kong University)
GenerationSuper ResolutionTransformerGaussian SplattingImage
🎯 What it does: Propose LGTM, a dual-network framework, achieving high-resolution (4K) text-based Gaussian rendering without scene optimization.
Less Is More: Clustered Cross-Covariance Control for Offline RL
Nan Qiao (Central South University), Ju Ren (Tsinghua University)
Reinforcement LearningBenchmark
🎯 What it does: This paper addresses the distribution shift problem in offline reinforcement learning by proposing the C4 method, which controls the cross-covariance of TD errors through partitioned sampling and gradient covariance penalty, enhancing stability and performance in low-sample and OOD scenarios.
LeSTD: LLM Compression via Learning-based Sparse Tensor Decomposition
Yi Li (University Of Texas At Dallas), Bingzhe Li (University Of Texas At Dallas)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a two-stage post-training LLM compression framework named LeSTD, which achieves high compression rates with minimal performance degradation by learning a shared low-rank base and sparsifying the core tensor on this base;
Let Features Decide Their Own Solvers: Hybrid Feature Caching for Diffusion Transformers
Shikang Zheng (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
GenerationComputational EfficiencyTransformerDiffusion modelImageVideoTextOrdinary Differential Equation
🎯 What it does: Propose the HyCa framework, which assigns different ODE solvers through dimension-level clustering to accelerate feature caching in Diffusion Transformers;
Let LLMs Speak Embedding Languages: Generative Text Embeddings via Iterative Contrastive Refinement
Yu-Che Tsai (National Taiwan University), Shou-De Lin (National Taiwan University)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposed the GIRCSE framework, which achieves richer and more semantically meaningful sentence embeddings through multi-step iterative refinement of text via autoregressive generation of soft tokens;
Let OOD Feature Exploring Vast Predefined Classifiers
Kewen Xia (Shanghai University), PeilinXu
ClassificationAnomaly DetectionConvolutional Neural NetworkTransformerImage
🎯 What it does: Studied open distribution (OOD) detection based on auxiliary outlier samples (Outlier Exposure, OE), proposing the VPC framework that achieves explicit separation of in-distribution (ID) and OOD features through a pre-set Orthogonal Equiangular Feature Space (OEFS), and designs a corresponding loss function and VPC Score based on subspace L2 activation for class-agnostic OOD discrimination.
Let's (not) just put things in Context: Test-time Training for Long-context LLMs
Rachit Bansal (Harvard University), Samy Jelassi (Harvard University)
RetrievalComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Improve the retrieval and reasoning performance of large language models (LLMs) in large-scale contexts by performing only a few gradient updates on query projections during inference (query-only test-time training, qTTT).
Let's Explore Step by Step: Generating Provable Formal Statements with Deductive Exploration
Qi Liu (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextSequential
🎯 What it does: Designed and implemented the DExploration framework, leveraging Lean 4 for step-by-step verifiable mathematical exploration, and generating provable formal propositions and proofs through three atomic actions: introduction, derivation, and submission.
Let's Split Up: Zero-Shot Classifier Edits for Fine-Grained Video Understanding
Kaiting Liu (Leiden University), Hazel Doughty (Leiden University)
ClassificationRecognitionSupervised Fine-TuningVision Language ModelVideoTextBenchmark
🎯 What it does: Proposed the 'category splitting' task of unsupervised splitting of coarse-grained categories into fine-grained subclasses within video classifiers;
Let's Think in Two Steps: Mitigating Agreement Bias in MLLMs with Self-Grounded Verification
Moises Andrade (Georgia Institute of Technology), Zsolt Kira (Georgia Institute of Technology)
TransformerLarge Language ModelPrompt EngineeringMultimodalityChain-of-Thought
🎯 What it does: This paper studies the application of multimodal large language models (MLLM) in evaluating agent behaviors, and systematically evaluates their performance across multiple environments (VisualWebArena, OSWorld, robomimic).
Leveraging Data to Say No: Memory Augmented Plug-and-Play Selective Prediction
Aditya Sarkar (University of Maryland, College Park), Nuno Vasconcelos (University of California, San Diego)
ClassificationTransformerVision Language ModelContrastive LearningImageTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Proposed a memory-enhanced pluggable selective prediction method (MA-PAPSP) without training for open-set tasks in vision-language models, such as image captioning, image-text matching, and classification.
Leveraging Discrete Function Decomposability for Scientific Design
James C Bowden, Jennifer Listgarten (University of California Berkeley)
OptimizationGraph Neural NetworkBiomedical Data
🎯 What it does: Proposes a distributed optimization method called DADO that leverages the decomposability of the objective function for discrete design.
Leveraging Explanation to Improve Generalization of Meta Reinforcement Learning
Shicheng Liu (Pennsylvania State University), Minghui Zhu (Pennsylvania State University)
Explainability and InterpretabilityRobotic IntelligenceMeta LearningReinforcement LearningTabularTime SeriesSequentialFinance Related
🎯 What it does: This paper proposes a two-stage method, first identifying the most critical training tasks for difficult-to-adapt tasks through example explanations, and then enhancing the meta-policy's focus on these critical tasks by maximizing conditional mutual information through task augmentation distribution learning, thereby improving the generalization performance of Meta RL.
Leveraging Pretrained Knowledge at Inference Time: LoRA-Gated Contrastive Decoding for Multilingual Factual Language Generation in Adapted LLMs
Gwangseon Jang (Korea Advanced Institute of Science and Technology), Mun Yong Yi (Korea Advanced Institute of Science and Technology)
GenerationTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed a decoding framework called LoRA-Gated Contrastive Decoding (LGCD) that enhances the factuality of language-specific fine-tuned large language models (LLMs) by dynamically gating and contrastive decoding knowledge extracted from the FFN layers of pre-trained models, without requiring additional training or original pre-training data;
LEXam: Benchmarking Legal Reasoning on 340 Law Exams
Yu Fan (ETH Zurich), Joel Niklaus (Niklaus.ai)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and released the LEXAM legal reasoning benchmark, comprising 340 exam questions from the University of Zurich, covering 7,537 multilingual (English and German) questions, including 2,841 open-ended long-answer questions and 4,696 multiple-choice questions.
LFQA-E: Carefully Benchmarking Long-form QA Evaluation
Yuchen Fan (Shanghai Jiao Tong University), Bowen Zhou (Shanghai AI Lab)
Large Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and constructed LFQA-E, a multilingual, reference-based, long-text question-answering evaluation benchmark covering 15 topics, containing 1,618 questions and 7,323 high-difficulty answer pairs, along with a systematic evaluation of 17 existing evaluation metrics.
LH-DECEPTION: Simulating and Understanding LLM Deceptive Behaviors in Long-Horizon Interactions
Yang Xu (Zhejiang University), Sharon Li (University of Wisconsin-Madison)
Explainability and InterpretabilityLarge Language ModelAgentic AITextSequential
🎯 What it does: Built the LH-DECEPTION framework for evaluating deception behaviors of LLMs in long-term interactions
Libra: Effective yet Efficient Load Balancing for Large-scale MoE Inference
Jaehoon Yang (Seoul National University), Jae W. Lee (Seoul National University)
Computational EfficiencyMixture of ExpertsText
🎯 What it does: Propose the Libra system to address expert load imbalance in large-scale MoE inference
Lifelong Embodied Navigation Learning
Xudong Wang (State Key Laboratory of Robotics and Intelligent Systems, Shenyang Institute of Automation, Chinese Academy of Sciences), Zhi Han (State Key Laboratory of Robotics and Intelligent Systems, Shenyang Institute of Automation, Chinese Academy of Sciences)
Autonomous DrivingMeta LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsVision Language ModelImageTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This work proposes the Lifelong Embodied Navigation Learning (LENL) task and designs the Uni-Walker framework, which can continuously learn new navigation scenarios and instruction styles without forgetting previously learned tasks;
Lifelong Learning with Behavior Consolidation for Vehicle Routing
Jiyuan Pei (Victoria University of Wellington), Xin Yao (Lingnan University)
OptimizationKnowledge DistillationMeta LearningReinforcement LearningBenchmark
🎯 What it does: Proposed the LLR-BC framework to enable lifelong learning for multi-distribution and multi-scale tasks in neural vehicle routing problem solvers, addressing catastrophic forgetting.
LiFR-Seg: Anytime High-Frame-Rate Segmentation via Event-Guided Propagation
Xiaoshan Wu (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
SegmentationTransformerScore-based ModelGaussian SplattingOptical FlowVideoMultimodality
🎯 What it does: Proposed the Anytime Interframe Semantic Segmentation task, achieving high frame rate semantic segmentation on single-frame RGB and asynchronous event streams;
Light Differentiable Logic Gate Networks
Lukas Rüttgers (ETH Zürich), Roger Wattenhofer (ETH Zürich)
ClassificationImage TranslationComputational EfficiencyImageText
🎯 What it does: This paper proposes an input-level parameterization (IWP) scheme, replacing the traditional logic gate network (DLGN) parameterization, significantly reducing the number of parameters while improving gradient stability and discretization accuracy;
Light of Normals: Unified Feature Representation for Universal Photometric Stereo
Houyuan Chen (Hong Kong University Of Science And Technology), Hao Zhao (Tsinghua University)
RestorationTransformerImageBenchmark
🎯 What it does: Proposes a unified photometric stereo framework named LINO UniPS for recovering surface normals under arbitrary unknown illumination.
Light-X: Generative 4D Video Rendering with Camera and Illumination Control
Tianqi Liu (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderVideoPoint Cloud
🎯 What it does: Propose Light-X, which can simultaneously control camera trajectory and illumination from monocular video, generating editable and temporally consistent videos.
LightCtrl: Training-free Controllable Video Relighting
Yizuo Peng (Tsinghua University), Xiaodong Cun (Great Bay University)
GenerationDiffusion modelVideo
🎯 What it does: Designed LightCtrl, a training-free, controllable video relighting method that leverages pre-trained image and video diffusion models to achieve video illumination control under user-specified light trajectories.
LightMem: Lightweight and Efficient Memory-Augmented Generation
Jizhan Fang (Zhejiang University), Ningyu Zhang (Zhejiang University)
GenerationRetrievalComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Designed and implemented LightMem, a lightweight memory framework that helps large language models (LLMs) efficiently filter, organize, retrieve, and maintain historical information in long conversations and cross-turn interactions.
LightRetriever: A LLM-based Text Retrieval Architecture with Extremely Faster Query Inference
Guangyuan Ma (Chinese Academy Of Sciences), Songlin Hu (Langboat Technology)
RetrievalTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose LightRetriever, an asynchronous dual-encoder architecture that simplifies query encoding in LLM retrieval to merely word embedding lookups.
Lightweight Spatio-Temporal Modeling via Temporally Shifted Distillation for Real-Time Accident Anticipation
Patrik Patera (National Taiwan University of Science and Technology), Wen-Hsien Fang (National Taiwan University of Science and Technology)
Autonomous DrivingComputational EfficiencyKnowledge DistillationRepresentation LearningRecurrent Neural NetworkTransformerContrastive LearningVideoMultimodality
🎯 What it does: Propose a lightweight spatiotemporal framework for real-time traffic accident prediction.
Lightweight Transformer for EEG Classification via Balanced Signed Graph Algorithm Unrolling
Junyi Yao (Peking University), Wei Hu (Peking University)
ClassificationTransformerContrastive LearningBiomedical Data
🎯 What it does: A lightweight and interpretable Transformer based on balanced signed graph spectrum denoising algorithm for epilepsy and healthy EEG binary classification.
LikePhys: Evaluating Intuitive Physics Understanding in Video Diffusion Models via Likelihood Preference
Jianhao Yuan (University of Oxford), Daniele De Martini (University of Oxford)
Diffusion modelVideoBenchmarkPhysics Related
🎯 What it does: Proposed a no-training, probabilistic preference-based evaluation method for video diffusion models to assess intuitive physical understanding, named LikePhys;
Linear Mechanisms for Spatiotemporal Reasoning in Vision Language Models
Raphi Kang (California Institute of Technology), Pietro Perona (California Institute of Technology)
Explainability and InterpretabilityImageVideo
🎯 What it does: Studied and quantified the mechanism by which visual spatial and temporal information in vision-language models (VLMs) is linearly bound to object word activation, and verified its causality through interventions.
LinearRAG: Linear Graph Retrieval Augmented Generation on Large-scale Corpora
Luyao Zhuang (Hong Kong Polytechnic University), Xiao Huang (Hong Kong Polytechnic University)
GenerationRetrievalComputational EfficiencyGraph Neural NetworkTextGraphRetrieval-Augmented Generation
🎯 What it does: Propose LinearRAG, a framework for retrieval-augmented generation that constructs a three-layer graph (Tri-Graph) based on entities without relying on relationships;
LinearSR: Unlocking Linear Attention for Stable and Efficient Image Super-Resolution
Xiaohui Li (Shanghai Jiao Tong University), Yihao Liu (Shanghai Artificial Intelligence Laboratory)
RestorationSuper ResolutionMixture of ExpertsVision Language ModelDiffusion modelFlow-based ModelImage
🎯 What it does: Proposed the LinearSR framework, which first combines linear attention (O(N)) with diffusion generative models in super-resolution tasks, achieving high-fidelity image reconstruction.
LingoLoop Attack: Trapping MLLMs via Linguistic Context and State Entrapment into Endless Loops
Jiyuan Fu (Fudan University), Wenqiang Zhang (Fudan University)
Adversarial AttackLarge Language ModelImage
🎯 What it does: Studied an energy consumption attack targeting multimodal large language models, inducing the model to generate extremely long and repetitive outputs, leading to resource exhaustion.
LINGOLY-TOO: Disentangling Reasoning from Knowledge with Templatised Orthographic Obfuscation
Jude Khouja (University of Oxford), Adam Mahdi (University of Oxford)
Data-Centric LearningLarge Language ModelTextBenchmark
🎯 What it does: This paper develops a novel language reasoning benchmark, LINGOLY-TOO, by obscuring UK Linguistics Olympiad problems through expert-designed orthographic substitutions, maintaining the reasoning logic while eliminating the model's reliance on knowledge and memory shortcuts.
LinguaMap: Which Layers of LLMs Speak Your Language and How to Tune Them?
J. Ben Tamo (Georgia Institute of Technology), Mingxian Wang (Amazon)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed a hierarchical diagnostic framework for evaluating the language control capability of multilingual LLMs;
LINK: Learning Instance-level Knowledge from Vision-Language Models for Human-Object Interaction Detection
WU EASTMAN Z Y (Tsinghua University), Shengjin Wang (Tsinghua University)
Object DetectionKnowledge DistillationVision Language ModelImageTextMultimodality
🎯 What it does: Proposed a pluggable HOI detection framework called LINK, integrating a person-object geometry encoder with a VLM connection decoder, and achieving instance-level dense supervision through progressive learning and teacher-student distillation;
Linking Process to Outcome: Conditional Reward Modeling for LLM Reasoning
Zheng Zhang (ShanghaiTech University), Kan Ren (ShanghaiTech University)
Large Language ModelReinforcement LearningText
🎯 What it does: Proposed a Conditional Reward Modeling (CRM) method to enhance the reasoning capabilities of large language models (LLMs) by treating the reasoning process as a time series and gradually guiding the model toward the correct answer.
LipNeXt: Scaling up Lipschitz-based Certified Robustness to Billion-parameter Models
Kai Hu (Carnegie Mellon University), Matt Fredrikson (Carnegie Mellon University)
Computational EfficiencyAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Proposed a novel 1-Lipschitz network architecture called LipNeXt, achieving scalable robust training on large-scale models with billions of parameters.
Lipschitz Bandits with Stochastic Delayed Feedback
Zhongxuan Liu (University of California, Davis), Thomas Lee
OptimizationReinforcement Learning
🎯 What it does: This paper studies the Lipschitz bandit problem with random delayed feedback and proposes two algorithms: the Delay-Aware Zooming algorithm for bounded delays and the Phased Pruning algorithm DLPP for unbounded delays.
LiteGuard: Efficient Task-Agnostic Model Fingerprinting with Enhanced Generalization
Guang Yang (Virginia Commonwealth University), Changqing Luo (University of Houston)
Safty and PrivacyComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkGraph Neural NetworkImageGraphTabularTime SeriesBiomedical Data
🎯 What it does: Proposed a task-agnostic model fingerprinting framework called LiteGuard, aiming to address the overfitting and high computational costs of existing MetaV when the model set size is limited.
LitmusValues: Will AI Tell Lies to Save Sick Children? Litmus-Testing AI Values Prioritization with AIRiskDilemmas
Yu Ying Chiu (University of Washington), Evan J Hubinger
Safty and PrivacyExplainability and InterpretabilityLarge Language ModelTextBenchmark
🎯 What it does: Proposes the LITMUSVALUES framework, which uses AIRISKDILEMMAS scenario oppositions to measure AI models' prioritization of 16 shared values and associate them with risk behaviors;
LiTo: Surface Light Field Tokenization
Jen-Hao Rick Chang, Oncel Tuzel (Apple)
GenerationRepresentation LearningTransformerDiffusion modelScore-based ModelFlow-based ModelNeural Radiance FieldContrastive LearningGaussian SplattingImageMesh
🎯 What it does: Proposed a new 3D latent representation capable of simultaneously modeling object geometry and view-dependent lighting effects, achieving the ability to generate complete 3D objects from a single image.
LiveClin: A Live Clinical Benchmark without Leakage
Xidong Wang (Chinese University of Hong Kong), Benyou Wang (Chinese University of Hong Kong)
Large Language ModelAgentic AIImageTextMultimodalityTabularBiomedical DataBenchmark
🎯 What it does: Constructed and evaluated a real-time, dynamic, cross-modal clinical benchmark called LiveClin, covering complete clinical pathways and updated every six months.
LiveMoments: Reselected Key Photo Restoration in Live Photos via Reference-guided Diffusion
Clara Xue (vivo BlueImage Lab), Qingnan Fan (vivo BlueImage Lab)
RestorationDiffusion modelOptical FlowImageVideo
🎯 What it does: To address the reselected low-quality keyframes in Live Photo, this paper proposes a dual-branch network based on diffusion models, which restores images by referring to the original high-quality keyframes.
LiveResearchBench: A Live Benchmark for User-Centric Deep Research in the Wild
Jiayu Wang (University of Wisconsin Madison), Shafiq Joty (Salesforce Ai Research)
Large Language ModelAgentic AITextBenchmark
🎯 What it does: Proposed the LiveResearchBench benchmark and DeepEval evaluation framework to measure the performance of agent systems on dynamic, user-centric, unambiguous, and multi-dimensional deep research tasks
LiveWeb-IE: A Benchmark For Online Web Information Extraction
Seungbin Yang (KAIST AI), Jaegul Choo (KAIST AI)
TransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed a new benchmark for real-time web information extraction (WIE), LIVEWEB-IE, and designed a vision-based multi-stage extraction framework, VGS;
LLaVA-4D: Embedding SpatioTemporal Prompt into LMMs for 4D Scene Understanding
Hanyu Zhou (National University of Singapore), Gim Hee Lee (National University of Singapore)
RecognitionTransformerLarge Language ModelVision Language ModelOptical FlowImageVideoTextMultimodality
🎯 What it does: Propose LLaVA-4D, a multimodal model that combines dynamic-aware 4D coordinate encoding with spatiotemporal decoupled visual embeddings to achieve full spatiotemporal understanding of dynamic scenes.
LLaVA-FA: Learning Fourier Approximation for Compressing Large Multimodal Models
Pengcheng Zheng (University of Electronic Science and Technology of China), Yang Yang (University of Electronic Science and Technology of China)
CompressionVision Language ModelMultimodality
🎯 What it does: In large-scale multimodal models, propose a method to compress weights through joint low-rank decomposition and quantization in the frequency domain, achieving more efficient Large Multimodal Models (LMM).
LLaVAction: evaluating and training multi-modal large language models for action understanding
Haozhe Qi (École Polytechnique Fédérale de Lausanne), Mackenzie W Mathis
RecognitionTransformerLarge Language ModelVision-Language-Action ModelVideoMultimodalityBenchmark
🎯 What it does: This paper proposes an action understanding framework based on multimodal large language models (MLLM), constructs a new action recognition benchmark EPIC‑KITCHENS‑100‑MQA, and trains the LLaVAction model on this benchmark;
LLEMA: Evolutionary Search with LLMs for Multi-Objective Materials Discovery
Nikhil Abhyankar (Virginia Tech), Chandan K. Reddy (Virginia Tech)
OptimizationDrug DiscoveryGraph Neural NetworkTransformerLarge Language ModelTextGraphTabularBenchmarkPhysics Related
🎯 What it does: The LLEMA framework combines LLM, chemical rules, and evolutionary search to achieve multi-objective material discovery.
LLM as an Algorithmist: Enhancing Anomaly Detectors via Programmatic Synthesis
Hangting Ye (Jilin University), Hongyuan Zha (CUHK-Shenzhen)
Anomaly DetectionTransformerLarge Language ModelTabularBenchmark
🎯 What it does: Propose the LLM-DAS framework, which generates 'hard anomalies' targeting weaknesses of detectors without accessing data, thereby enhancing the robustness of table anomaly detectors.
LLM DNA: Tracing Model Evolution via Functional Representations
Zhaomin Wu (National University of Singapore), Bingsheng He (National University of Singapore)
Explainability and InterpretabilityRepresentation LearningLarge Language ModelText
🎯 What it does: Proposes the concept of LLM DNA, constructing an interpretable low-dimensional representation to track the evolution and relationships of LLMs.
LLM Fingerprinting via Semantically Conditioned Watermarks
Thibaud Gloaguen (ETH Zurich), Martin Vechev (ETH Zurich)
Anomaly DetectionSafty and PrivacyLarge Language ModelText
🎯 What it does: The paper proposes a semantic conditional watermark-based LLM fingerprinting method that remains detectable even when the model is fine-tuned, quantized, pruned, or maliciously deployed.
LLM Pretraining with Continuous Concepts
Jihoon Tack (Meta), Xian Li (Meta)
Explainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelAuto EncoderContrastive LearningText
🎯 What it does: Propose the CoCoMix framework, which combines discrete next-token prediction with continuous concept prediction in large language model pretraining, and enhances performance by compressing the predicted concepts and inserting them into hidden layers.
LLM Unlearning with LLM Beliefs
Kemou Li (University of Macau), Jiantao Zhou (University of Macau)
Safty and PrivacyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a bootstrapping-based LLM forgetting framework that balances target text with the model's own high-confidence outputs, thereby effectively achieving deep forgetting.
LLM-as-a-Prophet: Understanding Predictive Intelligence with Prophet Arena
Qingchuan Yang (University of Southern California), Haifeng Xu (University of Chicago)
TransformerLarge Language ModelTextTime SeriesSequentialBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Proposed the LLM-as-a-Prophet paradigm and constructed the Prophet Arena real-time prediction market benchmark to evaluate the predictive intelligence of LLMs.
LLM-Guided Evolutionary Program Synthesis for Quasi-Monte Carlo Design
Amir Sadikov (University of California San Francisco)
OptimizationLarge Language ModelPoint CloudTabularFinance Related
🎯 What it does: Employ an LLM-driven evolutionary program synthesis framework to automatically generate and optimize 2D/3D low discrepancy point sets and Sobol direction numbers, achieving lower star discrepancy or lower rQMC mean squared error than existing state-of-the-art methods.
LLM-JEPA: Large Language Models Meet Joint Embedding Predictive Architectures
Hai Huang (Atlassian), Randall Balestriero (Brown University)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a training objective combining generative tasks with JEPA (LLM-JEPA), improving fine-tuning and pre-training performance across multiple models and datasets.
LLM2Fx-Tools: Tool Calling for Music Post-Production
SeungHeon Doh, Yuki Mitsufuji (Sony AI)
TransformerLarge Language ModelSupervised Fine-TuningMultimodalityChain-of-ThoughtAudio
🎯 What it does: Propose LLM2Fx-Tools, a multimodal tool calling framework, which uses large language models to generate executable audio effect chains (Fx-chain) and provides chain-of-thought reasoning and natural language responses, supporting reverse engineering and blind estimation;
LLMs are Greedy Agents: Effects of RL Fine-tuning on Decision-Making Abilities
Thomas Schmied (JKU Linz), Razvan Pascanu (Google DeepMind)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextChain-of-Thought
🎯 What it does: Analyze the shortcomings of large language models in decision-making tasks and enhance their exploration and decision-making capabilities through reinforcement learning fine-tuning (RLFT) on self-generated chain-of-thought (CoT).
LLMs are Single-threaded Reasoners: Demystifying the Working Mechanism of Soft Thinking
Junhong Wu (Baidu Inc), Hua Wu (Baidu Inc)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Analyze the behavior of Soft Thinking in LLMs and propose Stochastic Soft Thinking to break the greedy trap.
LLMs as Rules Oracles: Exploring Real-World Multimodal Reasoning in Tabletop Strategy Game Environments
Joseph J Peper, Lu Wang (University of Michigan)
Large Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Developed and released LUDOBENCH, a benchmark for evaluating the multimodal reasoning capabilities of visual language models in board games.
LLMs Can Hide Text in Other Text of the Same Length
Antonio Norelli (University of Oxford Project CETI), Michael M. Bronstein (University of Oxford Project CETI)
GenerationSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Utilize large language models (LLM) to embed arbitrary messages into another text of the same length and readability, achieving text camouflage and information hiding.
LLMs Get Lost In Multi-Turn Conversation
Philippe Laban (Microsoft Research), Jennifer Neville (Microsoft Research)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Compared the performance of LLMs in single-turn full-information tasks versus multi-turn partial-information tasks through large-scale simulations, revealing the significant impact of multi-turn interactions on LLM reliability and accuracy;
LLMS ON TRIAL: Evaluating Judicial Fairness For Large Language Models
Yiran HU, Weixing Shen (Tsinghua University)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Constructed an evaluation framework for judicial fairness, designed and implemented the JudiFair dataset containing 65 labels and 161 values, proposed a three-dimensional fairness evaluation metric encompassing inconsistency, bias, and imbalanced inaccuracy, and conducted systematic experiments on 16 LLMs from different sources.
LLMs Process Lists With General Filter Heads
Arnab Sen Sharma (Northeastern University), David Bau (Northeastern University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Study the mechanisms for implementing list filtering operations in large language models, identify and locate 'filter heads', and verify their transferability.
LLMs Struggle to Balance Reasoning and World Knowledge in Causal Narrative Understanding
Khurram Yamin (Carnegie Mellon), Bryan Wilder (Carnegie Mellon)
Data SynthesisExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextGraphSequentialChain-of-Thought
🎯 What it does: This paper investigates the causal reasoning ability of large language models (LLMs) in narrative texts, explores the interaction between world knowledge and reasoning capabilities, and discovers through synthetic, semi-synthetic, and real data experiments that models exhibit biases toward event order, prior knowledge, and graph structure complexity.
LMask: Learn to Solve Constrained Routing Problems with Lazy Masking
Tianyou Li (Peking University), Zaiwen Wen (Peking University)
OptimizationTransformerGraph
🎯 What it does: Propose the LMask framework, combining LazyMask and backtracking mechanism, utilizing Transformer's autoregressive model to generate feasible routing solutions that satisfy complex constraints.
lmgame-Bench: How Good are LLMs at Playing Games?
Lanxiang Hu (University Of California San Diego), Hao Zhang (University Of California San Diego)
Large Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the LMGame-Bench benchmark, employing switchable perception, memory, and reasoning modules to evaluate the performance of LLM/VLM across various games.
LoC-Decomp: LLM Autoformalization via Logical Concept Decomposition and Iterative Feedback Correction
Jiangze Shi (Beijing Institute of Technology), Guoren Wang (National Computer System Engineering Research Institute of China)
AI Code AssistantTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose an automatic formalization framework based on logical concept decomposition (LoC-Decomp), combining semantic consistency checks and the Lean 4 compiler to achieve iterative improvement;
Loc$^{2}$: Interpretable Cross-View Localization via Depth-Lifted Local Feature Matching
Zimin Xia (École Polytechnique Fédérale de Lausanne), Alexandre Alahi (École Polytechnique Fédérale de Lausanne)
Pose EstimationDepth EstimationAutonomous DrivingTransformerContrastive LearningImage
🎯 What it does: Propose a cross-view localization method based on direct correspondence of local features between ground and aerial images, achieving 3-DoF pose estimation through monocular depth prediction and scale-aware Procrustes alignment, with end-to-end trainability and interpretability.
Local Entropy Search over Descent Sequences for Bayesian Optimization
David Stenger (RWTH Aachen University), Sebastian Trimpe (RWTH Aachen University)
OptimizationBenchmark
🎯 What it does: Propose a Bayesian optimization method based on Local Entropy Search (LES), specifically designed for iterative descent sequences starting from a given initial point, achieving efficient search for local optima.
Local Geometry Attention for Time Series Forecasting under Realistic Corruptions
Dongbin Kim (Seoul National University), Jaewook Lee (Seoul National University)
Anomaly DetectionTransformerTime SeriesBenchmark
🎯 What it does: This paper proposes a local geometric attention (LGA) mechanism and constructs the first benchmark for robustness evaluation in time series prediction, TSRBench, based on real statistical pulse/horizontal displacement noise.
Local Linear Attention: An Optimal Interpolation of Linear and Softmax Attention For Test-Time Regression
Yifei Zuo (Northwestern University), Zhaoran Wang (Northwestern University)
OptimizationComputational EfficiencyTransformer
🎯 What it does: Proposed the Local Linear Attention (LLA) mechanism, designed and implemented a scalable block-level FlashLLA implementation based on local linear regression;
Local Reinforcement Learning with Action-Conditioned Root Mean Squared Q-Functions
Zequan Wu, Mengye Ren (New York University)
Reinforcement LearningBenchmark
🎯 What it does: Proposed a local reinforcement learning method without backpropagation, ARQ, based on the forward-forward (FF) algorithm, which estimates Q-values using the root mean square (RMS) affinity function and action conditioning, enabling direct scalar value prediction in each layer of units.
Local Success Does Not Compose: Benchmarking Large Language Models for Compositional Formal Verification
Xu Xu (Hong Kong University of Science and Technology), Binhang Yuan (Hong Kong University of Science and Technology)
AI Code AssistantLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes the DAFNYCOMP benchmark to evaluate the compositional formal specification generation and verification capabilities of large language models in multi-function composite programs.
Locality-Attending Vision Transformer
Sina Hajimiri (École de technologie supérieure), Jose Dolz (École de technologie supérieure)
ClassificationSegmentationTransformerSupervised Fine-TuningImage
🎯 What it does: A lightweight plug-and-play module called LocAt is proposed to enhance the performance of Vision Transformer (ViT) on dense prediction tasks such as semantic segmentation, while maintaining ViT's original classification performance.
Locality-aware Parallel Decoding for Efficient Autoregressive Image Generation
Zhuoyang Zhang, Song Han
GenerationTransformerImage
🎯 What it does: Proposed Locality-aware Parallel Decoding (LPD), significantly reducing image generation steps and lowering inference latency through flexible parallel autoregressive modeling and generation order scheduling based on spatial locality.
Localized Concept Erasure in Text-to-Image Diffusion Models via High-Level Representation Misdirection
Uichan Lee (Seoul National University of Science and Technology), Sangheum Hwang (Seoul National University of Science and Technology)
GenerationRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Achieving concept elimination in text-to-image diffusion models by misleading high-level semantic representations on the early layers of the text encoder
Localizing Task Recognition and Task Learning in In-Context Learning via Attention Head Analysis
Haolin Yang (University of Chicago), Naoya Inoue (JAIST)
Explainability and InterpretabilityMeta LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose the TSLA method, which identifies and locates attention heads responsible for task recognition (TR) and task learning (TL) in large language models through task subspace alignment quantization, and verifies their independent roles in in-context learning (ICL).
Locally Subspace-Informed Neural Operators for Efficient Multiscale PDE Solving
Alexander Rudikov (Institute of Numerical Mathematics), Ivan Oseledets (Institute of Numerical Mathematics)
Computational EfficiencyMeshPhysics Related
🎯 What it does: Propose GMsFEM-NO, a hybrid framework that integrates Generalized Multiscale Finite Element Method (GMsFEM) with neural operators (NO), where NO directly predicts the multiscale basis space of GMsFEM to accelerate solving high-contrast heterogeneous PDEs.
LoFT: Low-Rank Adaptation That Behaves Like Full Fine-Tuning
Nurbek Tastan (Mohamed bin Zayed University of Artificial Intelligence), Samuel Horváth (Mohamed bin Zayed University of Artificial Intelligence)
Domain AdaptationOptimizationComputational EfficiencyTransformerSupervised Fine-TuningImageTextBiomedical DataBenchmark
🎯 What it does: Propose a novel low-rank adaptation method called LoFT, aiming to align the optimization dynamics of low-rank parameter updates with full-parameter fine-tuning, thus significantly narrowing the performance gap with full fine-tuning without increasing inference costs.
Log Probability Tracking of LLM APIs
Timothee Chauvin, Gilles Tredan (LAAS)
Anomaly DetectionComputational EfficiencyLarge Language ModelTextBenchmark
🎯 What it does: Proposed a method that utilizes the logprob of the first token returned by LLM for continuous monitoring, enabling efficient detection of subtle changes in API models.
Log-Augmented Generation: Scaling Test-Time Reasoning with Reusable Computation
Peter Baile Chen (MIT), Mike Cafarella
Computational EfficiencyTransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Proposed the Log-Augmented Generation (LAG) framework, enabling LLMs to directly reuse past task inference logs during testing, enhancing inference efficiency and accuracy.
Log-Linear Attention
Han Guo (Massachusetts Institute of Technology), Yoon Kim (Massachusetts Institute of Technology)
Computational EfficiencyRepresentation LearningText
🎯 What it does: Proposed the Log-Linear Attention mechanism, utilizing Fenwick trees to hierarchically partition historical information, maintaining logarithmic-scale hidden states, and providing efficient parallel training and decoding algorithms; applied it to Mamba-2 and Gated DeltaNet, resulting in corresponding Log-Linear variants.
LogART: Pushing the Limit of Efficient Logarithmic Post-Training Quantization
Jiawei Xu (University of Macau), Zhuo Zou (Fudan University)
Computational EfficiencyConvolutional Neural NetworkTransformerImageText
🎯 What it does: Propose the LogART method, achieving learnable logarithmic rounding and multi-baseline, asynchronous, and robust-to-outliers quantizers for 3~4-bit weight quantization;
LogiConBench: Benchmarking Logical Consistencies of LLMs
Zheng CHEN, Zhouchen Lin (Peking University)
Prompt EngineeringTextGraphBenchmark
🎯 What it does: Constructed an expandable logical consistency benchmark called LogiConBench, generating 280K samples covering logical diagrams, reasoning paths, symbolic rewriting, and natural language formulation for 2–5 propositions.
LogicReward: Incentivizing LLM Reasoning via Step-Wise Logical Supervision
Jundong Xu (National University of Singapore), Wynne Hsu (National University of Singapore)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Propose a method to train large language models through step-by-step logical supervision (LogicReward), ensuring that the reasoning process satisfies logical validity at each step.
LogicXGNN: Grounded Logical Rules for Explaining Graph Neural Networks
Chuqin Geng (McGill University), Xujie Si (University of Toronto)
Explainability and InterpretabilityGraph Neural NetworkGraphBenchmark
🎯 What it does: Provide a post-hoc rule explanation framework called LOGICXGNN for graph neural networks (GNNs), which can generate credible and interpretable logical rules at the final subgraph level.
LogiStory: A Logic-Aware Framework for Multi-Image Story Visualization
Chutian Meng (Zhejiang University), Yueting Zhuang (Zhejiang University)
GenerationLarge Language ModelDiffusion modelMultimodalityBenchmark
🎯 What it does: Proposed and implemented the LogiStory framework for generating multi-image story visualization sequences, focusing on enhancing visual logical consistency.
Logit‑KL Flow Matching: Non‑Autoregressive Text Generation via Sampling‑Hybrid Inference
Egor Sevriugov (LigandPro), Ivan Oseledets (AXXX)
GenerationTransformerFlow-based ModelTextOrdinary Differential Equation
🎯 What it does: Proposed a conditional flow matching (KL-Flow) framework based on KL-geometry trajectories for non-autoregressive text generation, along with theoretical proofs and an iterative sampling and hybrid reasoning algorithm.
Long Chain-of-Thought Reasoning Across Languages
Josh Barua (University of California, Berkeley), Alane Suhr (University of California, Berkeley)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: The system conducted a comprehensive evaluation of Long CoT across nine non-English languages, comparing three reasoning setups—En-CoT, Target-CoT, and En-Only—and performed experiments across four stages: model scale, pre-training, multilingual pre-training, post-training, and inference efficiency.
Long-Context Attention Benchmark: From Kernel Efficiency to Distributed Context Parallelism
Tao Bu (Nanjing University), Jingwei Xu (Nanjing University)
Computational EfficiencyTransformerTextBenchmark
🎯 What it does: Proposes a unified long context attention benchmark (LongCA-bench), covering various implementations from single-device kernels to large-scale distributed context parallel mechanisms, supporting multiple mask patterns and sequence lengths;