arXivSub Start free trial

ICLR 2026 Papers with Code — Page 11

International Conference on Learning Representations · 2207 papers

Language Models are Injective and Hence Invertible

Giorgos Nikolaou (EPFL), Emanuele Rodolà (Sapienza University of Rome)

CodeSafty and PrivacyRepresentation LearningTransformerLarge Language ModelText

🎯 What it does: Prove and verify that the standard decoder-type Transformer is almost surely injective (injective) after initialization and training, and based on this, propose and implement the SIPIT algorithm, which can precisely recover the original input text from the representations of any hidden layer in linear time.

Language-guided Open-world Video Anomaly Detection under Weak Supervision

Zihao Liu (Communication University of China), Linlin Yang (Communication University of China)

CodeData SynthesisAnomaly DetectionTransformerVision Language ModelContrastive LearningVideoTextMultimodality

🎯 What it does: Propose a language-guided open-world video anomaly detection framework LaGoVAD to address the concept drift problem

LapFlow: Laplacian Multi-scale Flow Matching for Generative Modeling

Zelin Zhao (Georgia Institute of Technology), Yongxin Chen (Georgia Institute of Technology)

CodeGenerationTransformerFlow-based ModelAuto EncoderImageOrdinary Differential Equation

🎯 What it does: Proposes LapFlow, a framework that decomposes images into Laplacian pyramid residuals and generates high-quality images through parallel multi-scale flow matching;

LaSeR: Reinforcement Learning with Last-Token Self-Rewarding

Wenkai Yang (Renmin University of China), Yankai Lin (Tencent)

CodeTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: This paper proposes the LaSeR algorithm, which jointly optimizes the reasoning and self-evaluation capabilities of large language models;

Latent Denoising Makes Good Tokenizers

Jiawei Yang (University Of Southern California), Yue Wang

CodeGenerationTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkImageText

🎯 What it does: Propose Latent Denoising Tokenizer (-DeTok), which trains the tokenizer to reconstruct original images under severe noise or occlusion by introducing interpolation noise and random occlusion in the latent space;

Latent Diffusion Model without Variational Autoencoder

Minglei Shi (Tsinghua University), Jiwen Lu (Tsinghua University)

CodeGenerationTransformerDiffusion modelFlow-based ModelImage

🎯 What it does: Propose a VAE-free latent diffusion model SVG, which constructs a discriminative feature space by combining frozen DINO features with a residual encoder, and directly trains the diffusion model on this space.

Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling

Tal Daniel (Carnegie Mellon University), David Held (Carnegie Mellon University)

CodeAuto EncoderWorld ModelVideo

🎯 What it does: Proposed a self-supervised object-centric world model (LPWM) that can autonomously discover key points, bounding boxes, and object masks from video data, learning rich scene decompositions applicable to decision-making.

Latent Planning Emerges with Scale

Michael Hanna (University of Amsterdam), Emmanuel Ameisen (Anthropic)

CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: Investigated the implicit planning capabilities of large language models and analyzed the planning performance across different scales using causal feature circuits.

Latent Speech-Text Transformer

Yen-Ju Lu (Johns Hopkins University), Duc Le (Meta)

CodeComputational EfficiencyTransformerLarge Language ModelContrastive LearningTextAudio

🎯 What it does: This paper proposes Latent Speech-Text Transformer (LST), which significantly improves computational efficiency by aggregating continuous speech tokens into latent speech patches, balancing the modeling granularity between speech and text.

Latent Veracity Inference for Identifying Errors in Stepwise Reasoning

Minsu Kim (Mila Quebec AI Institute), Yoshua Bengio (Mila Quebec AI Institute)

CodeAnomaly DetectionExplainability and InterpretabilitySupervised Fine-TuningTextChain-of-Thought

🎯 What it does: This paper proposes a step-level error detection method based on latent variable models, which introduces veracity variables for each step in chain-of-thought (CoT) reasoning and automatically identifies erroneous reasoning steps through posterior inference using joint model likelihood.

Latent Wasserstein Adversarial Imitation Learning

Siqi Yang (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)

CodeReinforcement LearningGenerative Adversarial NetworkSequential

🎯 What it does: Developed a Wasserstein distance-based adversarial imitation learning framework called LWAIL, which utilizes dynamic perceptual embedding spaces generated by a pre-trained ICVF (intention-conditioned value function). In this space, Euclidean distance replaces the traditional KR Dual's Euclidean metric, enabling expert-level imitation with only a minimal number of state-only expert trajectories.

LaTo: Landmark-tokenized Diffusion Transformer for Fine-grained Human Face Editing

Zhenghao Zhang (Alibaba Cloud Computing), Weizhi Wang (Alibaba Cloud Computing)

CodeImage TranslationTransformerVision Language ModelDiffusion modelAuto EncoderImageTextChain-of-Thought

🎯 What it does: Proposed a landmark-marker-based diffusion transformer called LaTo, achieving fine-grained facial editing while preserving identity.

Lavida-O: Elastic Large Masked Diffusion Models for Unified Multimodal Understanding and Generation

Shufan Li (University of California Los Angeles), Jason Kuen (Adobe)

CodeGenerationTransformerMixture of ExpertsDiffusion modelMultimodality

🎯 What it does: Develop Lavida-O, a unified large-scale masked diffusion model supporting image understanding, object localization, image editing, and high-resolution text-to-image generation.

LayerSync: Self-aligning Intermediate Layers

Yasaman Haghighi (Ecole Polytechnique Federale De Lausanne), Alexandre Alahi (Ecole Polytechnique Federale De Lausanne)

CodeGenerationDiffusion modelImageVideoAudio

🎯 What it does: Proposed a self-aligned intermediate layer regularization method called LayerSync, aimed at improving the generation quality and training efficiency of diffusion models.

LD-MoLE: Learnable Dynamic Routing for Mixture of LoRA Experts

Yuan Zhuang (University of Connecticut), Fei Miao (University of Connecticut)

CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextBenchmark

🎯 What it does: Propose LD-MoLE, a learnable dynamic routing mechanism that combines Mixture of LoRA Experts to achieve token-level and layer-level adaptive expert allocation.

LDT: Layer-Decomposition Training Makes Networks More Generalizable

Zaizuo Tang (Nanjing University), Yu-Bin Yang (Nanjing University)

CodeClassificationSegmentationSuper ResolutionDomain AdaptationImage

🎯 What it does: To address the domain generalization problem, two methods, Layer-Decomposition Training (LDT) and Dynamic Parameter Update (DPU), are proposed. These methods stabilize gradients and improve the network's generalization performance on unseen domains through fine-grained hierarchical gradient variance separation and adaptive EMA updates.

Lean4Physics: Comprehensive Reasoning Framework for College-level Physics in Lean4

Yuxin Li (Hong Kong University of Science and Technology), Yi R. Fung (Hong Kong University of Science and Technology)

CodeLarge Language ModelTextBenchmarkPhysics Related

🎯 What it does: Built a physics reasoning framework called Lean4PHYS based on Lean4, which includes an extensible physics library PhysLib and a benchmark set of 200 formalized physics theorems called LeanPhysBench, and evaluated the performance of multiple LLMs and expert provers on this benchmark.

LEAP: Local ECT-Based Learnable Positional Encodings for Graphs

Juan P Garcia Amboage, Bastian Rieck (University Of Fribourg)

CodeClassificationRepresentation LearningGraph Neural NetworkGraph

🎯 What it does: Proposes LEAP, a differentiable and learnable graph position encoding based on the local Euler Characteristic Transform (ECT), which can be end-to-end integrated with GNNs.

Learn More with Less: Uncertainty Consistency Guided Query Selection for RLVR

Hao Yi (Renmin University of China), Yong Liu (Independent Researcher)

CodeData-Centric LearningReinforcement Learning from Human FeedbackReinforcement LearningText

🎯 What it does: Introduce an uncertainty consistency guided query selection method in RLVR to reduce annotation costs.

Learn to Reason Efficiently with Adaptive Length-based Reward Shaping

Wei Liu (Hong Kong University of Science and Technology), Junxian He (Hong Kong University of Science and Technology)

CodeComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought

🎯 What it does: Proposes a length-based reward shaping method called LASER and its dynamic difficulty-aware version LASER-D, enabling large-scale reasoning models to maintain accuracy while significantly compressing response length during chain-of-thought reasoning generation.

Learn-to-Distance: Distance Learning for Detecting LLM-Generated Text

Hongyi Zhou (Tsinghua University), Chengchun Shi (London School of Economics and Political Science)

CodeAnomaly DetectionLarge Language ModelSupervised Fine-TuningTextBenchmark

🎯 What it does: Proposes a LLM text detection method based on rewrite distance learning, capable of identifying LLM-generated text under unseen prompts.

Learnable Fractional Superlets with a Spectro-Temporal Emotion Encoder for Speech Emotion Recognition

Alaa Nfissi (Université TÉLUQ), Brian L Mishara (University of Québec at Montréal)

CodeClassificationRecognitionTransformerAudio

🎯 What it does: Proposed a learnable fractional-order superwavelet transform (LFST) and a compact spectral-temporal emotional encoder (STEE) for end-to-end speech emotion recognition from raw waveforms.

LearNAT: Learning NL2SQL with AST-guided Task Decomposition for Large Language Models

Weibin Liao (Peking University), Yasha Wang (Peking University)

CodeAI 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.

Learning Adaptive Distribution Alignment with Neural Characteristic Function for Graph Domain Adaptation

Wei Chen (Beihang University), deqing wang

CodeDomain 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 Concept Bottleneck Models from Mechanistic Explanations

Antonio De Santis (Politecnico di Milano), Lalana Kagal (MIT CSAIL)

CodeClassificationExplainability 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 Domain-Aware Task Prompt Representations for Multi-Domain All-in-One Image Restoration

Guanglu Dong (Sichuan University), Lichao Mou (Medai Technology Wuxi Co Ltd)

CodeRestorationDomain 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 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)

CodeAnomaly 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)

CodeRepresentation 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)

CodeClassificationImage

🎯 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)

CodeGenerationOptimizationReinforcement 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 the Past with Cascading Eligibility Traces

Tokiniaina Raharison Ralambomihanta (Mila Quebec Artificial Intelligence Institute), Blake Aaron Richards (Mila Quebec Artificial Intelligence Institute)

CodeClassificationRecurrent 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 Hierarchical and Geometry-Aware Graph Representations for Text-to-CAD

Shengjie Gong (South China University of Technology), Tianshui Chen (Guangdong University of Technology)

CodeGenerationData 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)

CodeOptimizationExplainability 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)

CodeOptimizationTabular

🎯 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 linear state-space models with sparse system matrices

Yasen Wang (China Telecom Research Institute), Gang Lu (China Telecom Research Institute)

CodeOptimizationTabularTime Series

🎯 What it does: Learning sparse linear state space models and recovering their topological structures

Learning Molecular Chirality via Chiral Determinant Kernels

Runhan Shi (Shanghai Jiao Tong University), Yang Yang (Shanghai Jiao Tong University)

CodeDrug 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 Nonlinear Causal Reductions to Explain Reinforcement Learning Policies

Armin Kekić (Max Planck Institute for Intelligent Systems), Michel Besserve (Max Planck Institute for Intelligent Systems)

CodeExplainability 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)

CodeOptimizationRepresentation 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 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)

CodeReinforcement 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 Posterior Predictive Distributions for Node Classification from Synthetic Graph Priors

Jeongwhan Choi (KAIST), Noseong Park (KAIST)

CodeClassificationData 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 Recursive Multi-Scale Representations for Irregular Multivariate Time Series Forecasting

Boyuan Li (South China University of Technology), Qianli Ma (South China University of Technology)

CodeComputational 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 Semi-Structured Sparsity for LLMs via Shared and Context-Aware Hypernetwork

Lu Sun (RIKEN), Jun Sakuma (RIKEN)

CodeCompressionComputational 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 Survival Distributions with Individually Calibrated Asymmetric Laplace Distribution

Deming Sheng (Duke University), Ricardo Henao (Duke University)

CodeOptimizationExplainability 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 To Draft: Adaptive Speculative Decoding with Reinforcement Learning

Jiebin Zhang (Peking University), Sujian Li (Peking University)

CodeComputational 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 Interpret Weight Differences in Language Models

Avichal Goel (Massachusetts Institute of Technology), Tony T. Wang (Massachusetts Institute of Technology)

CodeData 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 Reason in Structured In-context Environments with Reinforcement Learning

Peng Yu (Shanghai Jiao Tong University), Ying Wen (Shanghai Jiao Tong University)

CodeGraph 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)

CodeExplainability 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 Recall with Transformers Beyond Orthogonal Embeddings

Nuri Mert Vural (University of Toronto), Denny Wu (Flatiron Institute)

CodeRepresentation 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 Segment for Vehicle Routing Problems

Wenbin Ouyang (MIT), Cathy Wu (MIT)

CodeOptimizationComputational 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)

CodeOptimizationTransformerReinforcement 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)

CodeGenerationLarge 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 Unified Representation of 3D Gaussian Splatting

Yuelin Xin (UC Irvine), Xinke Li (City University of Hong Kong)

CodeGenerationRepresentation LearningAuto EncoderGaussian SplattingPoint Cloud

🎯 What it does: Propose a vectorized representation based on an equiprobable ellipsoidal submanifold, replacing traditional 3D Gaussian parameterization;

Learning What Matters Now: Dynamic Preference Inference under Contextual Shifts

Xianwei Cao (Xidian University), Shuang Wang (Xidian University)

CodeReinforcement 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 with Dual-level Noisy Correspondence for Multi-modal Entity Alignment

Haobin Li (Sichuan University), Xi Peng (Sichuan University)

CodeRepresentation 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).

LEGACY: A Lightweight Dynamic Gradient Compression Strategy for Distributed Deep Learning

Mostapha Essoullami (Mohammed VI Polytechnic University), Aritra Dutta (University of Central Florida)

CodeCompressionOptimizationFederated 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)

CodeRecognitionData 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)

CodeLarge 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.

Less Is More: Clustered Cross-Covariance Control for Offline RL

Nan Qiao (Central South University), Ju Ren (Tsinghua University)

CodeReinforcement 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.

Let LLMs Speak Embedding Languages: Generative Text Embeddings via Iterative Contrastive Refinement

Yu-Che Tsai (National Taiwan University), Shou-De Lin (National Taiwan University)

CodeRepresentation 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

CodeClassificationAnomaly 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 Explore Step by Step: Generating Provable Formal Statements with Deductive Exploration

Qi Liu (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)

CodeOptimizationTransformerLarge 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.

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)

CodeClassificationTransformerVision 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 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)

CodeGenerationTransformerLarge 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;

LFQA-E: Carefully Benchmarking Long-form QA Evaluation

Yuchen Fan (Shanghai Jiao Tong University), Bowen Zhou (Shanghai AI Lab)

CodeLarge 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.

Libra: Effective yet Efficient Load Balancing for Large-scale MoE Inference

Jaehoon Yang (Seoul National University), Jae W. Lee (Seoul National University)

CodeComputational 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)

CodeAutonomous 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)

CodeOptimizationKnowledge 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.

LightCtrl: Training-free Controllable Video Relighting

Yizuo Peng (Tsinghua University), Xiaodong Cun (Great Bay University)

CodeGenerationDiffusion 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)

CodeGenerationRetrievalComputational 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)

CodeRetrievalTransformerLarge 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.

Linear Mechanisms for Spatiotemporal Reasoning in Vision Language Models

Raphi Kang (California Institute of Technology), Pietro Perona (California Institute of Technology)

CodeExplainability 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)

CodeGenerationRetrievalComputational 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;

LINGOLY-TOO: Disentangling Reasoning from Knowledge with Templatised Orthographic Obfuscation

Jude Khouja (University of Oxford), Adam Mahdi (University of Oxford)

CodeData-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.

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

CodeSafty 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)

CodeGenerationRepresentation 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)

CodeLarge 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.

LiveWeb-IE: A Benchmark For Online Web Information Extraction

Seungbin Yang (KAIST AI), Jaegul Choo (KAIST AI)

CodeTransformerLarge 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)

CodeRecognitionTransformerLarge 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.

LLaVAction: evaluating and training multi-modal large language models for action understanding

Haozhe Qi (École Polytechnique Fédérale de Lausanne), Mackenzie W Mathis

CodeRecognitionTransformerLarge 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)

CodeOptimizationDrug 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)

CodeAnomaly 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.

LLM2Fx-Tools: Tool Calling for Music Post-Production

SeungHeon Doh, Yuki Mitsufuji (Sony AI)

CodeTransformerLarge 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 ON TRIAL: Evaluating Judicial Fairness For Large Language Models

Yiran HU, Weixing Shen (Tsinghua University)

CodeExplainability 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.

LMask: Learn to Solve Constrained Routing Problems with Lazy Masking

Tianyou Li (Peking University), Zaiwen Wen (Peking University)

CodeOptimizationTransformerGraph

🎯 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.

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)

CodePose 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)

CodeOptimizationBenchmark

🎯 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)

CodeAnomaly 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.

Locality-Attending Vision Transformer

Sina Hajimiri (École de technologie supérieure), Jose Dolz (École de technologie supérieure)

CodeClassificationSegmentationTransformerSupervised 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

CodeGenerationTransformerImage

🎯 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)

CodeGenerationRepresentation 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)

CodeExplainability 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).

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)

CodeDomain 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)

CodeAnomaly 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-Linear Attention

Han Guo (Massachusetts Institute of Technology), Yoon Kim (Massachusetts Institute of Technology)

CodeComputational 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)

CodeComputational 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)

CodePrompt 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.

Logit‑KL Flow Matching: Non‑Autoregressive Text Generation via Sampling‑Hybrid Inference

Egor Sevriugov (LigandPro), Ivan Oseledets (AXXX)

CodeGenerationTransformerFlow-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)

CodeExplainability 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)

CodeComputational 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;

Long-Context Generalization with Sparse Attention

Pavlo Vasylenko (Instituto Superior Técnico Universidade De Lisboa), Marcos Vinicius Treviso

CodeRetrievalComputational EfficiencyRepresentation LearningTransformerText

🎯 What it does: Replace the softmax attention in Transformers with α-entmax, and propose a learnable temperature and length-adjustable Adaptive-Scalable α-Entmax (ASEntmax) to achieve long-context generalization