ICLR 2026 Papers — Page 25
International Conference on Learning Representations · 5356 papers
Just Do It!? Computer-Use Agents Exhibit Blind Goal-Directedness
Erfan Shayegani (Microsoft Research AI Frontiers), Vibhav Vineet (Microsoft Research AI Frontiers)
TransformerLarge Language ModelAgentic AIPrompt EngineeringWorld ModelMultimodalityBenchmark
🎯 What it does: Investigated the Blind Goal-Directedness (BGD) phenomenon in Computer-Use Agents (CUA) and constructed the BLIND-ACT benchmark, comprising 90 tasks, to evaluate the blind goal-pursuing behavior of CUA under different contexts, hypothesis uncertainties, and contradictory/implementable goals.
K-Prism: A Knowledge-Guided and Prompt Integrated Universal Medical Image Segmentation Model
Bangwei Guo (Rutgers University), Dimitris N. Metaxas
SegmentationTransformerPrompt EngineeringMixture of ExpertsMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasound
🎯 What it does: Proposes a unified medical image segmentation framework K-Prism, capable of seamlessly switching between three knowledge paradigms (semantic prior, contextual examples, interactive feedback) and jointly training within the same model.
K-Sort Eval: Efficient Preference Evaluation for Visual Generation via Corrected VLM-as-a-Judge
Zhikai Li (Institute of Automation, Chinese Academy of Sciences), Kurt Keutzer (University of California, Berkeley)
GenerationPrompt EngineeringVision Language ModelImageVideoTextMultimodalityBenchmark
🎯 What it does: Propose a scalable evaluation framework K-Sort Eval based on vision-language models (VLM), which rapidly estimates human preference rankings for generative models using Bayesian calibration and dynamic matching.
K²-Agent: Co-Evolving Know-What and Know-How for Hierarchical Mobile Device Control
Zhe Wu (Tsinghua University), Yuanchun Shi (Huawei Noah's Ark Lab, Institute of Automation, Chinese Academy of Sciences)
Robotic IntelligenceReinforcement LearningAgentic AIVision Language Model
🎯 What it does: Proposes K²-Agent, a hierarchical framework for mobile device control that separates and co-evolves 'knowing what' and 'knowing how'.
Kaleidoscope: In-language Exams for Massively Multilingual Vision Evaluation
Israfel Salazar (University of Copenhagen), Marzieh Fadaee (Cohere Labs)
Vision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed a multilingual, multimodal multiple-choice exam benchmark called KALEIDOSCOPE containing 20,911 questions, and systematically evaluated the performance of various closed-source and open-source vision-language models on this benchmark.
KaLM-Embedding-V2: Superior Training Techniques and Data Inspire A Versatile Embedding Model
Xinping Zhao (Shenzhen Loop Area Institute), Min Zhang (Shenzhen Loop Area Institute)
RetrievalKnowledge DistillationRepresentation LearningTransformerContrastive LearningText
🎯 What it does: Proposed and implemented the 0.5B lightweight text embedding model series KaLM-Embedding-V2
KANO: Kolmogorov-Arnold Neural Operator
Jin Lee (University of California, Santa Barbara), Zheng Zhang (University of California, Santa Barbara)
Explainability and InterpretabilityComputational EfficiencyPhysics Related
🎯 What it does: Proposed and implemented the Kolmogorov-Arnold Neural Operator (KANO), a neural operator that simultaneously sparsely represents and is interpretable in both spatial and frequency domains;
KaVa: Latent Reasoning via Compressed KV-Cache Distillation
Anna Kuzina (Qualcomm AI Research), Babak Ehteshami Bejnordi (Qualcomm AI Research)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Propose the KAVA framework, which uses compressed KV-cache for self-distillation to train the implicit reasoning student to generate only continuous latent thoughts instead of lengthy CoT.
KBVQ-MoE: KLT-guided SVD with Bias-Corrected Vector Quantization for MoE Large Language Models
Zukang Xu (Houmo AI), Dawei Yang
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose KBVQ-MoE, addressing the low-bit vector quantization issue in MoE large models by achieving a balance between compression and performance through input-driven redundancy elimination and bias correction-based output stabilization.
KDP: Simplifying Representation Dynamics in Kernel Space
Zeyu Ma (Shanghai Normal University), Jianhong Wu (Shanghai Key Laboratory of Computer Software Testing and Evaluating)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Perform hierarchical pruning on large language models by proposing the Kernelized Dynamics Pruning (KDP) method, which replaces Transformer layer blocks by linearizing the representation of dynamics in kernel space, while maintaining model performance.
Keep the Best, Forget the Rest: Reliable Alignment with Order-Aware Preference Optimization
Jiahui Zhu (Washington State University), Honghao Wei (Washington State University)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: Propose an improved DPO algorithm called RAPPO, which reduces reference strategy bias by filtering low-alignment samples, thereby enhancing the alignment of language models.
KeepLoRA: Continual Learning with Residual Gradient Adaptation
Mao-Lin Luo (Southeast University), Tong Wei (Southeast University)
Representation LearningMeta LearningTransformerVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose a LoRA-based continual learning method called KeepLoRA, which balances pre-trained knowledge, previously learned tasks, and new tasks by projecting gradients into the residual subspace for updates;
KernelFusion: Zero-Shot Blind Super-Resolution via Patch Diffusion
Oliver Heinimann (Weizmann Institute of Science), michal Irani
Super ResolutionConvolutional Neural NetworkDiffusion modelScore-based ModelImage
🎯 What it does: Utilizes a zero-shot diffusion approach to directly infer image-specific down-sampling convolution kernels from a single low-resolution image, while simultaneously restoring the corresponding high-resolution image.
Kevin: Multi-Turn RL for Generating CUDA Kernels
Carlo Baronio (Stanford University Cognition AI), Silas Alberti (Cognition AI)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Proposes a multi-round reinforcement learning training scheme to enhance the ability of large language models to generate CUDA kernels.
KGOT: Unified Knowledge Graph and Optimal Transport Pseudo-Labeling for Molecule-Protein Interaction Prediction
Jiayu Qin (University at Buffalo), zhiqiang xu
Drug DiscoveryGraph Neural NetworkTransformerGraphBiomedical Data
🎯 What it does: Propose a unified framework that integrates knowledge graphs with optimal transport pseudo-labels for molecular-protein interaction retrieval;
Kimi-Dev: Agentless Training as Skill Prior for SWE-agents
Zonghan Yang (Moonshot AI), Tianyu Liu (Shanghai Qi Zhi Institute)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: Propose a code LLM called Kimi-Dev based on Agentless training. First, they develop bug repair and test writing capabilities through mid-training, cold start, reinforcement learning, and self-play. Then, they fine-tune with a small amount of SWE-Agent trajectories to achieve high-performance SWE-Agent.
KinemaDiff: Towards Diffusion for Coherent and Physically Plausible Human Motion Prediction
Ye Lu (Nanyang Technological University), Kim-Hui Yap (Nanyang Technological University)
GenerationPose EstimationGraph Neural NetworkTransformerDiffusion modelGraphSequential
🎯 What it does: Proposes KinemaDiff, a human motion prediction model that integrates a joint-adaptive noise generator and a structural alignment regularizer into the diffusion process.
KL-Regularized Reinforcement Learning for Generative Modelling is Designed to Mode Collapse
Anthony GX-Chen (New York University), Rajesh Ranganath (New York University)
GenerationDrug DiscoveryReinforcement LearningTextBiomedical Data
🎯 What it does: Study the mode collapse problem caused by KL regularization in reinforcement learning, and propose a reward-anchored multimodal target distribution method (MARA) to restore diversity.
KLAS: Using Similarity to Stitch Neural Networks for Improved Accuracy-Efficiency Tradeoffs
Debopam Sanyal (Georgia Institute of Technology), Alexey Tumanov (Cisco Research)
ClassificationSegmentationComputational EfficiencyConvolutional Neural NetworkTransformerImageText
🎯 What it does: Propose a model concatenation framework called KLAS based on KL divergence, which automatically selects concatenation points between pre-trained models to improve accuracy or reduce FLOPs under the same computational budget.
Know When to Abstain: Optimal Selective Classification with Likelihood Ratios
Alvin Heng (National University of Singapore), Harold Soh (National University of Singapore)
ClassificationDomain AdaptationImageText
🎯 What it does: This paper proposes to design a selector function using the Neyman-Pearson lemma, unifying and improving selective classification methods, particularly for covariate shift scenarios;
KnowGuard: Knowledge-Driven Abstention for Multi-Round Clinical Reasoning
Xilin Dang (Chinese University of Hong Kong), Pheng-Ann Heng
TransformerLarge Language ModelMultimodalityGraphBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes a knowledge graph-based proactive exploration framework (KnowGuard) for deciding when to stop answering (abstention) and collect evidence in multi-turn clinical dialogues;
Knowing When to Quit: Probabilistic Early Exits for Speech Separation Networks
Kenny Falkær Olsen (Technical University of Denmark), Morten Mørup (Technical University of Denmark)
RestorationRecurrent Neural NetworkTransformerAudio
🎯 What it does: Developed a probability early stopping framework (PRESS) based on uncertainty quantification for single-channel speech separation and enhancement, and designed an architecture capable of outputting at multiple depths of the network;
Knowledge Distillation as Decontamination? Revisiting the “Data Laundering” Concern in Classification Tasks
Hengyu Luo (University of Helsinki), Jörg Tiedemann (University of Helsinki)
ClassificationKnowledge DistillationTransformerTextBenchmark
🎯 What it does: Investigate whether knowledge distillation in classification tasks leads to 'data washing'—i.e., leaking test set information from a polluted teacher model to the student model, and assess the severity of this phenomenon.
Knowledge Distillation for Large Language Models through Residual Learning
Thinh On (Amazon Web Services), George Karypis (Amazon Web Services)
Knowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderTextBenchmark
🎯 What it does: Proposed a two-stage white-box knowledge distillation framework, first compressing the teacher model's hidden representations through self-reconstruction, then guiding the student model training via residual learning.
Knowledge Exchange with Confidence: Cost-Effective LLM Integration for Reliable and Efficient Visual Question Answering
Mahsa Mozaffari (Rochester Institute of Technology), Qi Yu (Rochester Institute of Technology)
Knowledge DistillationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Propose a confidence-based collaborative framework named Uni-VQA that integrates large language models (LLMs) with task-specific VQA models to achieve knowledge exchange and improve visual question answering performance.
Knowledge Externalization: Reversible Unlearning and Modular Retrieval in Multimodal Large Language Models
Jiaqi Li (Southeast University), Guilin Qi (Southeast University)
RetrievalSafty and PrivacyTransformerLarge Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes a reversible, modular framework for knowledge externalization, which can transfer sensitive concepts from the parameters of multimodal large language models to external memory tokens, enabling recovery after forgetting.
Knowledge Fusion of Large Language Models via Modular SkillPacks
Guodong DU (Harbin Institute of Technology), Jing Li (Harbin Institute of Technology)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical DataBenchmarkFinance Related
🎯 What it does: Propose the GraftLLM framework to achieve cross-model capability transfer, encoding source model knowledge into a lightweight SkillPack, supporting the fusion and continuous learning of heterogeneous LLMs.
Knowledge Reasoning Language Model: Unifying Knowledge and Language for Inductive Knowledge Graph Reasoning
Xingrui Zhuo (Hefei University of Technology), Xindong Wu (Hefei University of Technology)
Graph Neural NetworkTransformerLarge Language ModelGraphBenchmark
🎯 What it does: Integrate LLM knowledge with KG context to realize KRLM, addressing the knowledge distortion problem in open-domain induced knowledge graph reasoning
Knowledgeable Language Models as Black-Box Optimizers for Personalized Medicine
Michael S Yao (University of Pennsylvania), Claudia Iriondo (Genentech)
OptimizationTransformerLarge Language ModelPrompt EngineeringBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Explored using large language models (LLMs) as unsupervised black-box optimizers to generate personalized treatment plans through knowledge-driven entropy constraints.
KnowledgeSmith: Uncovering Knowledge Updating in LLMs with Model Editing and Unlearning
Yinyi Luo (Carnegie Mellon University), Jindong Wang (Carnegie Mellon University)
Large Language ModelGraphBenchmark
🎯 What it does: This paper proposes the KnowledgeSmith framework, which unifies the analysis of knowledge editing and machine forgetting (unlearning) in LLMs, and constructs a large-scale structured benchmark through automated knowledge graph data generation;
KnowProxy: Adapting Large Language Models by Knowledge-guided Proxy
Gukhyeon Lee (Korea University), SangKeun Lee (Korea University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose KNOWPROXY, an agent method for fine-tuning large language models using textual knowledge rather than probability distributions;
Koopman-Assisted Trajectory Synthesis: A Data Augmentation Framework for Offline Imitation Learning
Jin Wang (Nanjing University of Aeronautics and Astronautics), Xiaoyang Tan (Nanjing University of Aeronautics and Astronautics)
Data SynthesisReinforcement LearningAuto EncoderSequentialBenchmark
🎯 What it does: Propose a trajectory-level data augmentation framework called KATS, which leverages the Koopman transformation to enhance policy robustness and generalization in offline imitation learning.
KRAMABENCH: A Benchmark for AI Systems on Data-to-Insight Pipelines over Data Lakes
Eugenie Lai (MIT CSAIL), Tim Kraska (MIT CSAIL)
Data-Centric LearningTransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Evaluate the capability of large language models (LLMs) to design and execute complete data-to-insight pipelines on real-world data lakes, constructing a benchmark called KRAMABENCH with 104 multi-domain tasks.
KV Cache Transform Coding for Compact Storage in LLM Inference
Konrad Staniszewski (NVIDIA), Adrian Łańcucki (NVIDIA)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a lightweight transformation encoding method (KVTc) for compressing KV caches during LLM inference, achieving significant compression ratios while maintaining inference quality.
KVComm: Enabling Efficient LLM Communication through Selective KV Sharing
Xiangyu Shi (KTH Royal Institute of Technology), Dejan Kostic (KTH Royal Institute of Technology)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the KVComm framework, which achieves efficient communication between LLMs by sharing key-value (KV) pairs in critical layers only.
La-Proteina: Atomistic Protein Generation via Partially Latent Flow Matching
Tomas Geffner (NVIDIA), Arash Vahdat (NVIDIA)
GenerationProtein Structure PredictionTransformerFlow-based ModelAuto EncoderBiomedical Data
🎯 What it does: Proposes La-Proteina, a full-atom protein generation model based on partial latent flow matching.
Label Smoothing Improves Machine Unlearning
Zonglin Di (University of California, Santa Cruz), Yang Liu (University of California, Santa Cruz)
Safty and PrivacyComputational EfficiencyImageTextMultimodality
🎯 What it does: Proposed a machine unlearning method called UGradSL, which combines gradient ascent with label smoothing (negative label smoothing), enabling rapid elimination of the memory of specific data while maintaining model accuracy and enhancing local differential privacy;
Label-Free Mitigation of Spurious Correlations in VLMs using Sparse Autoencoders
Bharat Chandra Yalavarthi (University at Buffalo), Venu Govindaraju (University at Buffalo)
Explainability and InterpretabilityRepresentation LearningVision Language ModelAuto EncoderImageMultimodalityBiomedical DataBenchmark
🎯 What it does: Propose a fully unsupervised, zero-shot audio-visual model debiasing method called DIAL, which utilizes sparse autoencoders to separate and remove spurious correlation subspaces in the embedding space.
LadderSym: A Multimodal Interleaved Transformer for Music Practice Error Detection
Benjamin Shiue-Hal Chou (Purdue University), Yung-Hsiang Lu (Purdue University)
ClassificationTransformerPrompt EngineeringMultimodalityAudio
🎯 What it does: Propose a multi-modal interleaved Transformer called LadderSym for detecting errors in music practice
LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning
Haoqiang Kang (University of California San Diego), Lianhui Qin (University of California San Diego)
Explainability and InterpretabilityTransformerLarge Language ModelDiffusion modelAuto EncoderText
🎯 What it does: Designed LaDiR, a text reasoning framework that utilizes latent diffusion models for self-correction in a continuous latent space.
LAMDA: A Longitudinal Android Malware Benchmark for Concept Drift Analysis
Md Ahsanul Haque (University of Texas at El Paso), Mohammad Saidur Rahman (Indian Institute of Technology Hyderabad)
TransformerTabularBenchmark
🎯 What it does: Created a large-scale Android malware longitudinal benchmark LAMDA spanning 12 years (2013-2025), and systematically evaluated concept drift and feature stability using this benchmark.
Landing with the Score: Riemannian Optimization through Denoising
Andrey Kharitenko (ETH Zurich), Florian Dorfler
OptimizationScore-based Model
🎯 What it does: This paper studies methods for Riemannian optimization on implicit data manifolds, proposing a denoising approach to achieve optimization, particularly in the context of generative AI.
Landscape of Thoughts: Visualizing the Reasoning Process of Large Language Models
Zhanke Zhou (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Propose LoT (Landscape of Thoughts), a visualization tool for 2D visualization of step-by-step reasoning trajectories of large language models (LLMs), and develop a lightweight validator based on this to improve reasoning accuracy.
LANE: Label-Aware Noise Elimination for Fine-Grained Text Classification
Tiberiu Sosea (University of Illinois Chicago), Cornelia Caragea (University of Illinois Chicago)
ClassificationTransformerContrastive LearningText
🎯 What it does: Proposes the Label-Aware Noise Elimination (LANE) method, which assigns dynamic weights to each sample based on semantic similarity between labels and training dynamics to mitigate the impact of noisy labels.
Language Agents for Hypothesis-driven Clinical Decision Making with Reinforcement Learning
David Bani-Harouni (Technical University of Munich), Matthias Keicher (Technical University of Munich)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextMultimodalityTabularBiomedical DataElectronic Health Records
🎯 What it does: Built and trained a dual-agent language model (LA-CDM) composed of a hypothesis generator and a decision maker, achieving iterative narrowing of diagnostic hypotheses through repeated requests and explanations of diagnostic tests, ultimately providing a diagnosis.
Language and Experience: A Computational Model of Social Learning in Complex Tasks
Cédric Colas (Massachusetts Institute of Technology), Joshua B. Tenenbaum
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelWorld ModelVideoText
🎯 What it does: Proposed a Bayesian framework that treats language guidance and direct experience as complementary evidence, inferring executable procedural world models, and verified its learning and social learning effectiveness on 10 VGDL video games.
Language Confusion Gate: Language-Aware Decoding Through Model Self-Distillation
Collin Zhang (Qwen Team, Alibaba Group), Junyang Lin (Qwen Team, Alibaba Group)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the Language Confusion Gate (LCG), a lightweight plugin that dynamically filters inappropriate language tokens during LLM decoding to reduce language confusion.
Language Identification in the Limit with Computational Trace
Binghui Peng (University of Maryland), Grigoris Velegkas (Google)
RecognitionChain-of-Thought
🎯 What it does: This paper introduces chain reasoning trajectories into Gold's language identification model, defining a learning paradigm with trajectories under extreme conditions and proving that all enumerable languages can be identified in the absence of noise.
Language in the Flow of Time: Time-Series-Paired Texts Weaved into a Unified Temporal Narrative
Zihao Li, Jingrui He (IBM Research)
Representation LearningData-Centric LearningTransformerLarge Language ModelTextMultimodalityTime Series
🎯 What it does: Proposes a plug-and-play framework called TaTS, which treats text as an auxiliary variable of time series. It first encodes text using pre-trained language models, then reduces dimensionality via MLP and concatenates it with the original time series, making it usable by any existing time series model.
Language Models are Injective and Hence Invertible
Giorgos Nikolaou (EPFL), Emanuele Rodolà (Sapienza University of Rome)
Safty 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 Models Use Lookbacks to Track Beliefs
Nikhil Prakash (Northeastern University), Atticus Geiger (Goodfire)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Analyze how large language models track characters' beliefs and reveal an intrinsic mechanism called 'lookback'.
Language-guided Open-world Video Anomaly Detection under Weak Supervision
Zihao Liu (Communication University of China), Linlin Yang (Communication University of China)
Data 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
Language-Instructed Vision Embeddings for Controllable and Generalizable Perception
Chengzhi Mao (Google), Wen-Sheng Chu (Google)
RetrievalKnowledge DistillationRepresentation LearningTransformerLarge Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the language-guided visual encoder LIVE, leveraging knowledge distillation from large language models (LLMs) to enable the visual encoder to dynamically generate task-specific embeddings during inference based on instructions, supporting zero-copy instruction-driven visual perception.
LapFlow: Laplacian Multi-scale Flow Matching for Generative Modeling
Zelin Zhao (Georgia Institute of Technology), Yongxin Chen (Georgia Institute of Technology)
GenerationTransformerFlow-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;
Laplacian Kernelized Bandit
Shuang Wu (UCLA), Arash A. Amini (UCLA)
OptimizationReinforcement LearningGraph
🎯 What it does: Studied the multi-user contextual bandit problem, where users are connected through a known graph structure, and the reward function is non-linear and satisfies graph homophily; proposed a joint penalty combining graph smoothness and individual RKHS regularization, proved its equivalence to the norm of a single multi-user RKHS, and explicitly derived the corresponding reproducing kernel; based on this unified kernel design, proposed two Gaussian Process bandit algorithms, LK-GP-UCB and LK-GP-TS, and provided high-probability regret upper bounds based on effective dimension; demonstrated their superiority over multiple baselines in synthetic experiments.
LaplacianFormer:Rethinking Linear Attention with Laplacian Kernel
Zhe Feng (University of Chinese Academy of Sciences), Xiaopeng Zhang (Chinese Academy of Sciences)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: Propose a linear attention Transformer called LaplacianFormer, achieving injective and efficient attention mechanisms by introducing the Laplacian kernel.
Large Depth Completion Model from Sparse Observations
Zhu Yu (Zhejiang University), Hui-liang Shen
Depth EstimationTransformerPrompt EngineeringImagePoint CloudBenchmark
🎯 What it does: Propose a Transformer-based LDCM model to generate dense, metric-scale depth/point clouds from sparse depth observations and RGB images;
Large Language Model Compression with Global Rank and Sparsity Optimization
Changhai Zhou (Fudan University), Cheng Jin (Fudan University)
CompressionComputational EfficiencyTransformerReinforcement LearningText
🎯 What it does: This paper proposes a two-stage LLM compression framework called CAP, which first decomposes weights into low-rank and sparse subspaces using RPCA, and then learns the retention ratio under a global budget through Bayesian sampling and policy gradient.
Large Scale Diffusion Distillation via Score-Regularized Continuous-Time Consistency
Kaiwen Zheng (Tsinghua), Qinsheng Zhang (NVIDIA)
GenerationKnowledge DistillationTransformerDiffusion modelScore-based ModelImageVideoTextStochastic Differential Equation
🎯 What it does: Proposes a Score-Regularized Continuous-Time Consistency Model (rCM), achieving few-step distillation for large-scale image and video models
Larger Datasets Can Be Repeated More: A Theoretical Analysis of Multi-Epoch Scaling in Linear Regression
Tingkai Yan (Peking University), Kaifeng Lyu (Tsinghua University)
OptimizationData-Centric LearningText
🎯 What it does: Proposed the definition of effective reuse rate E_{K,N}, and conducted theoretical derivation and experimental validation under the linear regression setting with multi-round SGD training.
LaSeR: Reinforcement Learning with Last-Token Self-Rewarding
Wenkai Yang (Renmin University of China), Yankai Lin (Tencent)
TransformerLarge 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;
Late-to-Early Training: LET LLMs Learn Earlier, So Faster and Better
Ji Zhao (Hong Kong University of Science and Technology Guangzhou), Zeke Xie (Hong Kong University of Science and Technology Guangzhou)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Proposes the Late-to-Early Training (LET) paradigm, which uses representations from later layers of a small pre-trained model to guide the learning of earlier layers in a large model, accelerating LLM pre-training and improving performance.
Latent Adaptation of Foundation Policies for Sim-to-Real Transfer
Longchao Da (Arizona State University), Hua Wei (Arizona State University)
Domain AdaptationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: Propose a transferable framework called Found-adapt based on foundation policies, which learns a general state encoder and potential controller in offline simulations, and achieves zero-shot or few-shot transfer from simulation to real-world environments by performing lightweight adaptation of the potential space with only a small amount of data in real environments.
Latent Concept Disentanglement in Transformer-based Language Models
Guan Zhe Hong (University of Oxford), Rina Panigrahy (Google Research)
Explainability and InterpretabilityRepresentation LearningTransformerTextSequential
🎯 What it does: Investigate how Transformer models separate and utilize hidden concepts in ICL tasks, demonstrating that models achieve two-hop reasoning and low-dimensional representation of numerical parameters through sparse attention heads.
Latent Denoising Makes Good Tokenizers
Jiawei Yang (University Of Southern California), Yue Wang
GenerationTransformerDiffusion 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)
GenerationTransformerDiffusion 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 Fourier Transform
Mason Long Wang (Massachusetts Institute of Technology), Cheng-Zhi Anna Huang (Massachusetts Institute of Technology)
GenerationDiffusion modelAuto EncoderAudio
🎯 What it does: Proposes the Latent Fourier Transform framework, achieving precise control over different time scales in music generation and mixing by applying Fourier transforms and performing frequency domain masking in the latent space.
Latent Geometry-Driven Network Automata for Complex Network Dismantling
Thomas Adler (Tsinghua University), Carlo Vittorio Cannistraci
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: Propose a network automaton framework named LGD-NA based on potential geometry, which estimates the geometric distance between nodes using local rules and decomposes the network according to this distance.
Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling
Tal Daniel (Carnegie Mellon University), David Held (Carnegie Mellon University)
Auto 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)
Explainability 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)
Computational 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 Stochastic Interpolants
Saurabh Singh (Poetiq AI), Dmitry Lagun (Google DeepMind)
GenerationDiffusion modelFlow-based ModelImageStochastic Differential Equation
🎯 What it does: Propose Latent Stochastic Interpolants (LSI), jointly training an encoder, decoder, and continuous-time stochastic interpolation generative model in the latent space to achieve end-to-end efficient learning.
Latent Thinking Optimization: Your Latent Reasoning Language Model Secretly Encodes Reward Signals in Its Latent Thoughts
Hanwen Du (Ohio State University), Xia Ning (Ohio State University)
OptimizationTransformerLarge Language ModelText
🎯 What it does: Investigate the potential thought trajectories of Huginn-3.5B, discovering distinguishable patterns between correct and incorrect trajectories in the latent space, and build a Latent Reward Model (LRM) and Latent Thinking Optimization (LTO) algorithm, optimizing the reasoning process via probabilistic sampling without updating model parameters.
Latent Veracity Inference for Identifying Errors in Stepwise Reasoning
Minsu Kim (Mila Quebec AI Institute), Yoshua Bengio (Mila Quebec AI Institute)
Anomaly 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 Visual Reasoning
Bangzheng Li (University of California, Davis), Zicheng Liu (Advanced Micro Devices, Inc)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the Latent Visual Reasoning (LVR) model, enabling multimodal large language models to perform autoregressive reasoning in visual embedding space and alternate with text generation;
Latent Wasserstein Adversarial Imitation Learning
Siqi Yang (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)
Reinforcement 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.
Latent Wavelet Diffusion For Ultra High-Resolution Image Synthesis
Luigi Sigillo (Sapienza University of Rome), Danilo Comminiello (Singapore Management University)
GenerationData SynthesisDiffusion modelAuto EncoderImage
🎯 What it does: Proposed the Latent Wavelet Diffusion (LWD) framework, introducing frequency-aware spatial adaptive masks into latent diffusion models to enhance ultra-high-resolution (2K–4K) image synthesis quality.
Latent-Guided Reasoning: Empowering Small LLMs with Large-Model Thinking
Hanzhu Chen (University of Science and Technology of China), Jianye HAO (Tianjin University)
Knowledge DistillationRepresentation LearningLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed the Latent Guidance framework, which decouples high-level cognitive planning of large language models from low-level text generation of small language models, enabling small models to acquire the reasoning capabilities of large models;
Latent-to-Data Cascaded Diffusion Models for Unconditional Time Series Generation
Lifeng Shen (Chongqing University of Posts and Telecommunications), James Kwok
GenerationData SynthesisDiffusion modelScore-based ModelTime Series
🎯 What it does: Propose a dual-space diffusion framework L2D-Diff, which first generates global representations in the latent space and then refines temporal data in the data space, achieving unconditional time series generation.
LatentQA: Teaching LLMs to Decode Activations Into Natural Language
Alexander Pan (University Of California Berkeley), Jacob Steinhardt (University Of California Berkeley)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the LATENTQA framework, training a decoder LLM to read and manipulate activations of large language models (LLMs) through natural language question answering; simultaneously constructing a large-scale activation-question answering dataset and fine-tuning via the LIT method;
LaTo: Landmark-tokenized Diffusion Transformer for Fine-grained Human Face Editing
Zhenghao Zhang (Alibaba Cloud Computing), Weizhi Wang (Alibaba Cloud Computing)
Image 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.
LaVCa: LLM-assisted Visual Cortex Captioning
Takuya Matsuyama (University of Osaka), Yu Takagi (University of Osaka)
Explainability and InterpretabilityTransformerLarge Language ModelVision Language ModelMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposed LaVCa, a visual cortex lexical description method based on large language models, which automatically generates natural language annotations for each voxel.
Lavida-O: Elastic Large Masked Diffusion Models for Unified Multimodal Understanding and Generation
Shufan Li (University of California Los Angeles), Jason Kuen (Adobe)
GenerationTransformerMixture 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)
GenerationDiffusion 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.
Layerwise Federated Learning for Heterogeneous Quantum Clients using Quorus
Jason Han (Rice University), Tirthak Patel (Rice University)
ClassificationFederated LearningImagePhysics Related
🎯 What it does: Proposed the Quorus framework, which realizes hierarchical federated learning for quantum clients with varying depths, allowing each client to select variational quantum circuits of different depths based on their hardware capabilities.
LazyDrag: Enabling Stable Drag-Based Editing on Multi-Modal Diffusion Transformers via Explicit Correspondence
Zixin Yin (Hong Kong University of Science and Technology), Heung-Yeung Shum (Hong Kong University of Science and Technology)
GenerationTransformerDiffusion modelImage
🎯 What it does: This paper proposes a training-free drag-based editing method called LazyDrag, which achieves full inversion strength on multi-modal diffusion Transformers through explicit correspondence mapping, eliminating the need for optimization during testing and enabling stable image editing and natural inpainting.
LCA: Local Classifier Alignment for Continual Learning
Tung Tran (Kyushu University), Khoat Than (Hanoi University of Science and Technology)
ClassificationTransformerImage
🎯 What it does: Propose a local classifier alignment (LCA) loss, which is combined with incremental model merging (IM) to form a complete method for continual learning, addressing the mismatch between the merged classifier and feature extractor.
LD-EnSF: Synergizing Latent Dynamics with Ensemble Score Filters for Fast Data Assimilation with Sparse Observations
Pengpeng Xiao (Georgia Institute of Technology), Peng Chen (Georgia Institute of Technology)
Computational EfficiencyRecurrent Neural NetworkScore-based ModelTime SeriesPhysics Related
🎯 What it does: Proposed the LD-EnSF method, achieving fast Bayesian data assimilation for high-dimensional sparse observations using LDNet and LSTM encoders in a low-dimensional latent space;
LD-MoLE: Learnable Dynamic Routing for Mixture of LoRA Experts
Yuan Zhuang (University of Connecticut), Fei Miao (University of Connecticut)
Computational 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)
ClassificationSegmentationSuper 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.
Lean Finder: Semantic Search for Mathlib That Understands User Intents
Jialin Lu (Simon Fraser University), Wuyang Chen (University Of Texas At Austin)
Data SynthesisRetrievalAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose Lean Finder, a semantic retrieval engine for Lean and mathlib, capable of understanding mathematicians' intentions and retrieving relevant theorems.
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)
Large 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)
ClassificationRepresentation 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)
Data-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 the Ropes, Then Trust the Wins: Self-imitation with Progressive Exploration for Agentic Reinforcement Learning
Yulei Qin (Tencent Youtu Lab), Xing Sun (Tencent Youtu Lab)
Reinforcement LearningAgentic AIText
🎯 What it does: Propose SPEAR—a reinforcement learning (RL) framework that combines self-mimicry with progressive exploration for training large language model (LLM)-driven agents;
Learn to Guide Your Diffusion Model
Alexandre Galashov (Google DeepMind), Valentin De Bortoli (Google DeepMind)
GenerationDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes a method to learn time and condition-related guidance weights (ω_c(s,t)) in diffusion models, utilizing self-consistency constraints to minimize the difference between the true and guided distributions, thereby achieving more accurate distribution matching during conditional sampling.
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)
Computational 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)
Anomaly 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.
Learnability and Privacy Vulnerability are Entangled in a Few Critical Weights
Xingli Fang (North Carolina State University), Jung-Eun Kim (North Carolina State University)
Safty and PrivacyConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: Studied a mechanism for estimating member privacy vulnerability at the weight level, and proposed the CWRF method to enhance model robustness against membership inference attacks by resetting and freezing vulnerable weights during fine-tuning.
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)
ClassificationRecognitionTransformerAudio
🎯 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.