ICML 2026 Papers — Page 30
International Conference on Machine Learning · 6554 papers
KIStego: Key-Independent Secure Image Distribution via Bipartite Structural Invariants
Ren Lijing, Denghui Zhang (Guangzhou University)
Safty and PrivacyConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningGaussian SplattingImageBenchmark
🎯 What it does: By first down-sampling and halftoning the secret image to obtain a binary invariant structure, then using binary sharing (XOR) to generate two zero-information shares, and mapping them through an orthogonal mapping to the Gaussian latent space of a diffusion model, the paper achieves keyless image steganography and secure transmission.
KITE: Knowledge-Guided Probabilistic Modeling for Time Series Forecasting with Exogenous Variables
Hanyin Cheng (East China Normal University), Chenjuan Guo (East China Normal University)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelFlow-based ModelContrastive LearningTabularTime SeriesFinance Related
🎯 What it does: Propose the KITE framework to achieve probabilistic time series forecasting with external variables, introducing historical conditions, knowledge guidance, and classifier-free guidance during the generation process.
Klein Hyperbolic Metric Learning
Yulin Zhang (Beihang University), Junlin Hu (Beihang University)
ClassificationRetrievalRepresentation LearningTransformerAuto EncoderContrastive LearningImage
🎯 What it does: This paper proposes applying the Klein model to hyperbolic metric learning and constructs a numerically stable framework based on Einstein's four-vector space, achieving the projection of Euclidean visual features onto the Klein hyperplane and its loss optimization.
Knapsack RL: Compute-Efficient Reinforcement Learning via Heterogeneous Rollout Allocation
Ziniu Li (Chinese University of Hong Kong), Zhi-Quan Luo (Chinese University of Hong Kong)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper studies how to heterogeneously allocate rollout budgets when fine-tuning large language models with reinforcement learning, and proposes the Knapsack RL framework;
KnapSpec: Self-Speculative Decoding via Adaptive Layer Selection as a Knapsack Problem
Seongjin Cha (KAIST), Insu Han (KAIST)
OptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the KnapSpec framework, which uses adaptive hierarchical selection to construct the draft model in self-speculative decoding (SSD) as a 0/1 knapsack problem, making it training-agnostic and applicable to long sequences;
Knothe-Rosenblatt Quantile Regression for Risk-sensitive Multi-objective Reinforcement Learning
Gwangpyo Yoo (Sungkyunkwan University), Honguk Woo (Sungkyunkwan University)
OptimizationTransformerReinforcement LearningTabularTime SeriesSequentialBenchmarkFinance Related
🎯 What it does: Propose a risk-sensitive multi-objective reinforcement learning framework based on Knothe-Rosenblatt (KR) quantile regression (KR-IQN)
Know More, Know Clearer: A Meta-Cognitive Framework for Knowledge Augmentation in Large Language Models
Hao Chen (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
Explainability and InterpretabilityKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a meta-cognitive knowledge-enhanced framework that can perform differentiated expansion for three types of knowledge states: mastered, confused, and missing, and can also calibrate the model's confidence and actual accuracy through a cognitive consistency alignment mechanism, enabling the model to better distinguish between known and unknown knowledge.
Know Thyself, Know Thy User: Intrinsic Dual-Perspective Reasoning for Role-Playing LLMs
Haotong Sun (Hangzhou Shenji Technology Co Ltd), Yinghui Jiang (Xiamen University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a model called KSKT that achieves dual-perspective reasoning from both self-awareness and user understanding in role-playing large language models, addressing the conflict between role authenticity and user satisfaction.
Knowing Bias, Doing Better: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancement
Jinhao Pan (George Mason University), Ziwei Zhu (George Mason University)
Federated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: To address social bias in large language models (LLMs), this paper proposes a lightweight method called KnowBias, which enhances the internal 'bias-aware' neurons (know-bias neurons) during inference, thereby suppressing biased outputs without compromising the model's overall capabilities.
Knowing the Unknown: Interpretable Open-World Object Detection via Concept Decomposition Model
Xueqiang Lv (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
Object DetectionExplainability and InterpretabilityConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelAuto EncoderContrastive LearningImageText
🎯 What it does: Propose an interpretable open-world object detection framework called IPOW, which decomposes RoI features into three categories of concepts: discriminative, shared, and background, and addresses the confusion between known and unknown classes through a concept-guided correction mechanism.
Knowing When to Quit: A Principled Framework for Dynamic Abstention in LLM Reasoning
Hen Davidov (University of Oxford), Patrick Rebeschini (University of Oxford)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A dynamic abandonment principle framework is proposed to improve the efficiency of large language models (LLMs) in reasoning tasks, especially when generating long and potentially incorrect responses.
Knowing Who, Not How Much: Learning-Augmented Mechanisms for Consumer Utility Maximization
Kira Goldner (Boston University), Thodoris Tsilivis (Boston University)
OptimizationFederated LearningReinforcement Learning from Human Feedback
🎯 What it does: Study the mechanism for maximizing consumer utility under an online random sequential model, and design a deterministic phased mechanism, introducing prediction to enhance performance.
Knowledge Diversion for Efficient Morphology Control and Policy Transfer
Fu Feng (Southeast University), Xin Geng (Southeast University)
Computational EfficiencyKnowledge DistillationRobotic IntelligenceTransformerReinforcement LearningMixture of ExpertsContrastive LearningTabularTime SeriesSequential
🎯 What it does: Propose DivMorph, which utilizes knowledge diversion on a Morphology-Aware Transformer to achieve a decomposable general morphological controller, supporting dynamic soft routing between tasks and morphologies and modular reuse.
KODA: Contrastive Representation Comparison and Alignment for Vision-Language Foundation Models
Youqi WU, Farzan Farnia (Chinese University of Hong Kong)
Explainability and InterpretabilityRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposes the Contrastive Embedding Clustering task and designs the KODA (Kernel Optimization for Discrepancy Analysis) framework, which can identify subsets of samples that cluster strongly in one embedding space and weakly in another, thereby revealing structural differences between the two representations.
KORE: Enhancing Knowledge Injection for Large Multimodal Models via Knowledge-Oriented Controls
Kailin Jiang (University of Science and Technology of China), Qing Li (State Key Laboratory of General Artificial Intelligence, BIGAI)
Knowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the KORE method, which utilizes knowledge-oriented control to achieve knowledge injection and retention in large-scale multi-modal models.
Krause Synchronization Transformers
Jingkun Liu (Shanghai Qi Zhi Institute), Yue Song (Shanghai Qi Zhi Institute)
ClassificationGenerationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelContrastive LearningImageText
🎯 What it does: This paper proposes an attention mechanism based on the Krause consensus model—Krause Attention—which replaces the traditional dot product similarity with a distance metric, and limits the interaction between tokens through local neighborhood and top-k sparsification to suppress global synchronization and attention traps.
KromHC: Manifold-Constrained Hyper-Connections with Kronecker-Product Residual Matrices
Wuyang Zhou (Imperial College London), Danilo Mandic (Imperial College London)
OptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: A new framework called KromHC is proposed, which utilizes the Kronecker product of dual random matrices to address the training instability and parameter complexity issues of hyper-connection (HC) in neural networks.
Kronecker Generative Networks: A General Neural Architecture for Parameter-Efficient Learning Across Classification Tasks
Yang Yang (Wuhan University), Wei Xiang (La Trobe University)
ClassificationImageSequential
🎯 What it does: Propose a recursive generation rule based on the Kronecker product to construct a controllable network topology (KGN), treating topology as the primary design object;
KUMA: A Novel Framework with Koopman Separation and Efficient Multilevel Extraction in Time Series Forecasting
Sijie Xiong (Kyushu University), Atsushi Shimada (Kyushu University)
Computational EfficiencyTransformerTime SeriesBenchmark
🎯 What it does: Propose the KUMA framework, which separates time series into Koopman dynamics and residual dynamics, and further models the residual using a U-shaped multi-layer attention mechanism;
Kuramoto Oscillatory Phase Encoding: Neuro-inspired Synchronization for Improved Learning Efficiency
Mingqing Xiao (Microsoft Research Asia), Dongsheng Li (Microsoft Research Asia)
ClassificationSegmentationComputational EfficiencyRepresentation LearningTransformerAuto EncoderContrastive LearningImageMultimodality
🎯 What it does: Introduce a phase state that evolves with the hierarchy in Vision Transformer—Kuramoto oscillator phase encoding (KoPE)—and drive structural learning through phase synchronization and attention rotation.
L-CUBE: Isolating Long-Context Capacity from Knowledge with Controllable Mutual Information Scaling
Zhuo Chen (NSF AI Institute for Artificial Intelligence and Fundamental Interactions), Marin Soljacic
Data SynthesisComputational EfficiencyRepresentation LearningTransformerLarge Language ModelScore-based ModelContrastive LearningTextSequentialBenchmark
🎯 What it does: This paper proposes L-CUBE, a controllable information synthetic long sequence benchmark, used to separate the long context capture capability of language models from the confounding effects of semantic knowledge.
L-Drive: Beyond a Single Mapping—Latent Context Drives Time Series Forecasting
Fan Zhang (Shandong Technology and Business University), Hua Wang (Ludong University)
Recurrent Neural NetworkTransformerAuto EncoderContrastive LearningTime Series
🎯 What it does: Proposed the L-Drive framework, which achieves adaptive modeling of temporal changes by introducing latent context (L-Context) and patch-based relative position basis functions, thereby reducing prediction lag;
L-SR1: Learned Symmetric-Rank-One Preconditioning
Gal Lifshitz (Tel Aviv University), Dan Raviv (Tel Aviv University)
OptimizationMeshTabular
🎯 What it does: Proposed a learning-based second-order optimizer L-SR1, which achieves low-rank preconditioning of the inverse Hessian by combining the classical SR1 method with a learnable projection-guided secant mechanism.
L2G-NET: Local to Global Spectral Graph Neural Networks via Cauchy Factorizations
Samuel Fernandez (University of Southern California), Antonio Ortega (University of Southern California)
Computational EfficiencyRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Proposed L2G-Net, which achieves efficient spectral graph convolution by hierarchically partitioning the graph and utilizing Cauchy matrix decomposition.
LabBuilder: Protocol-Grounded 3D Layout Generation for Interactable and Safe Laboratory
Jianbao Cao (Shanghai Artificial Intelligence Laboratory), Dongzhan Zhou (Shanghai Artificial Intelligence Laboratory)
GenerationOptimizationRobotic IntelligenceTransformerLarge Language ModelTextPoint CloudMeshBenchmark
🎯 What it does: Propose a 3D laboratory layout auto-generation system called LabBuilder based on experimental protocols.
Label-Guided Representation Learning for Incomplete Multi-View Multi-Label Classification
Yang Li (National University of Defense Technology), Tingjin Luo (National University of Defense Technology)
ClassificationRepresentation LearningTransformerAuto EncoderContrastive LearningImageMultimodality
🎯 What it does: Propose Label‑Guided Representation Learning (LGRL) to address the multi-view multi-label classification problem with missing views and missing labels.
LABO: LLM-Accelerated Bayesian Optimization through Broad Exploration and Selective Experimentation
Zhuo Chen (Shanghai Artificial Intelligence Laboratory), Qinying Gu (Shanghai Artificial Intelligence Laboratory)
OptimizationLarge Language ModelGaussian SplattingTabularTime SeriesSequentialBenchmark
🎯 What it does: Proposed a Bayesian optimization framework called LABO that combines large language model (LLM) predictions with real experiments, using a gating strategy to dynamically allocate resources between global low-cost exploration and local high-precision experiments.
LAGEA: Language Guided Embodied Agents for Robotic Manipulation
Abdul Monaf Chowdhury (University of Dhaka), Rabeya Akter (University of Dhaka)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringVision Language ModelVision-Language-Action ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the LAGEA framework, which utilizes a vision-language model (VLM) to generate structured error reflection and converts it into a temporalized reward signal to guide reinforcement learning in robotic manipulation tasks.
LagLLM: LLM-empowered lead–lag dependency learning for spatial-temporal time series forecasting
Binqing Wu (Zhejiang University), Ling Chen (Zhejiang University)
Explainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringGraphTime SeriesBenchmark
🎯 What it does: Propose the LagLLM framework, which utilizes LLMs to generate lead-lag graphs through prompting and perform structured token sorting, enabling the model to explicitly capture spatial-temporal dependencies to improve the accuracy of time series forecasting.
Lagrangian Perturbation Diffusion Steering: Latent Reinforcement Learning for Generative Policies
Hikmet Simsir (Bilkent University), Ozgur S. Oguz (Bilkent University)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelScore-based ModelFlow-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageVideoTextMultimodalityPoint CloudTabularTime SeriesStochastic Differential EquationAudio
🎯 What it does: Proposes Lagrangian Perturbation Diffusion Steering (LPDS), a method for online adaptation by learning state-conditioned noise perturbations on a frozen generative control policy.
LakeQA: An Exploratory QA Benchmark over a Million-Scale Data Lake
Haonan Wang (Columbia University), Eugene Wu (Columbia University)
RetrievalExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTextMultimodalityTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed a heterogeneous data lake benchmark called LAKEQA, covering approximately 9.5TB of data and 40 million files, and evaluated the exploratory question-answering capabilities of LLMs on it.
LALM-as-a-Judge: Benchmarking Large Audio-Language Models for Safety Evaluation in Multi-Turn Spoken Dialogues
Amir Ivry (Technion-Israel Institute of Technology), Shinji Watanabe (Carnegie Mellon University)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-ThoughtAudio
🎯 What it does: A safety evaluation benchmark based on 24,000 multi-turn spoken dialogues was developed, and the safety judgment capabilities of six large-scale audio-language models (LALM) were assessed under text, audio, and multimodal inputs.
LAMP: Data-Efficient Linear Affine Weight-Space Models for Parameter-Controlled 3D Shape Generation and Extrapolation
Ghadi Nehme (Massachusetts Institute of Technology), Faez Ahmed (Massachusetts Institute of Technology)
GenerationData SynthesisOptimizationDiffusion modelScore-based ModelAuto EncoderContrastive LearningPoint CloudMesh
🎯 What it does: Construct a controllable and safe 3D mesh generation method by performing linear affine mixing in the aligned SDF decoder weight space from a small number of parameterized samples.
Landmark-Guided Policy Optimization for Multi-Objective Language Model Selection
Marcio Monteiro (RPTU University Kaiserslautern-Landau), Sophie Fellenz (RPTU University Kaiserslautern-Landau)
OptimizationHyperparameter SearchMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes LAMPS, a multi-objective AutoML framework designed to efficiently select pre-trained LLMs for fine-tuning.
Langevin Rollout Optimization for Modelic Reinforcement Learning
Tianyi Zhang (Tsinghua University), Shengbo Eben Li (Tsinghua University)
OptimizationReinforcement LearningScore-based ModelWorld ModelTabularTime SeriesSequentialStochastic Differential Equation
🎯 What it does: This paper proposes a gradient-guided online planning method based on Langevin dynamics, called LaRO, and combines it with MPPI to form MLAP, achieving more efficient planning and learning within the BOOM framework, ultimately proposing the improved algorithm BOOM-L.
LangForce: Bayesian Decomposition of Vision Language Action Models via Latent Action Queries
Shijie Lian (Huazhong University of Science and Technology), Kai Chen (Beijing Zhongguancun Academy)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerVision Language ModelVision-Language-Action ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the LangForce framework, which forces visual-language-action models to focus on language instructions through a dual-branch structure and potential action queries.
LangPrecip: Language-Aware Multimodal Precipitation Nowcasting
Xudong Ling (University of Electronic Science and Technology of China), Guiduo Duan (University of Electronic Science and Technology of China)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelRectified FlowAuto EncoderImageTextMultimodality
🎯 What it does: Utilizing natural language momentum descriptions to linguistically guide short-term precipitation forecasts, enabling the forecast model to achieve more accurate precipitation evolution predictions under limited historical radar windows.
Language as a Wave Phenomenon: Semantic Phase Locking and Interference in Neural Networks
Alper YILDIRIM, İbrahim Yücedağ (Düzce University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelAuto EncoderContrastive LearningText
🎯 What it does: This paper proposes the PRISM model, which replaces attention with unit norm constraints and phase-guided Fourier filtering to explore the computational role of phase in language sequence modeling;
Language Bias in LVLMs: From In-Depth Analysis to Simple and Effective Mitigation
Yangneng Chen (Harbin Institute of Technology), Jing Li (Harbin Institute of Technology)
OptimizationExplainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Study and alleviate language bias in large vision-language models (LVLM), improving the model's visual consistency and credibility.
Language Generation in the Limit: Complexity Barriers and Implications for Learning
Marcelo Arenas (Pontifical Catholic University of Chile), Alexander Kozachinskiy (National Center for Artificial Intelligence of Chile)
GenerationTextReview/Survey Paper
🎯 What it does: This paper studies the sample complexity of achieving language generation on a finite language family at the limit, exploring the feasibility and difficulty of various classical formal language classes such as context-free, regular, locally testable threshold (LTT), and non-erasing pattern languages.
Language Generation with Feedback: Queries and Mistakes
Steve Hanneke (Purdue University), Grigoris Velegkas (Google Research)
GenerationReinforcement Learning from Human FeedbackLarge Language ModelTextReview/Survey Paper
🎯 What it does: This paper studies the problem of language generation with feedback, and provides necessary and sufficient conditions for the language set to be generable under two models: error feedback and query feedback.
Language Generation with Replay: A Learning-Theoretic View of Model Collapse
Giorgio Racca (University of Copenhagen), Amartya Sanyal (University of Copenhagen)
GenerationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextReview/Survey Paper
🎯 What it does: This paper constructs a replay model for language generation from the perspective of learning theory, analyzing the feasibility of language generation under the scenario of model self-replay (model collapse);
Language Model Augmented Semi-Supervised Statistical Inference
Xinrui Ruan (University of California, Berkeley), Jingshen Wang (University of California, Berkeley)
Federated LearningExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelContrastive LearningTextBiomedical DataAlzheimer's DiseaseRetrieval-Augmented GenerationChain-of-ThoughtAudio
🎯 What it does: Propose a framework (LASS) that utilizes large language models (LLMs) for semi-supervised statistical inference, improving the efficiency of parameter estimation and maintaining statistical validity by generating pseudo-labels through calibrated LLM predictions when labels are missing.
Language Model Circuits Are Sparse in the Neuron Basis
Aryaman Arora (Transluce), Sarah Schwettmann (Transluce)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringAuto EncoderContrastive LearningText
🎯 What it does: This paper investigates the use of MLP activations as a sparse interpretable feature basis in language models and proposes a circuit tracing method based on gradients, verifying its effectiveness on models such as Llama3.1‑8B‑Instruct.
Language Model Networks: Supervision-Efficient Learning through Dense Communication
Shiguang Wu (Tsinghua University), Quanming Yao (Tsinghua University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: The paper proposes a differentiable language model network called LMNet, which achieves dense vector communication between models by trimming pre-trained language models into Transformer nodes without embedding/decoding layers, and adding trainable seq2seq edges between nodes, enabling model-to-model communication guided by supervision signals from the final task.
Language-based Trial and Error Falls Behind in the Era of Experience
Haoyu Wang (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
Computational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextSequentialRetrieval-Augmented Generation
🎯 What it does: Propose the SCOUT framework, which separates a lightweight scout network from a large language model (LLM). The scout efficiently explores the task first, and then the environment dynamics are injected into the LLM through supervised fine-tuning and multi-round reinforcement learning, enhancing its performance on unseen symbolic/spatial tasks.
Laplacian Representations for Decision-Time Planning
Dikshant Shehmar (University of Alberta), Marlos C. Machado (University of Alberta)
Reinforcement LearningContrastive LearningWorld ModelOptical FlowGraphTabularTime Series
🎯 What it does: This paper proposes a decision-time planning algorithm called ALPS that utilizes Laplacian representations, achieving efficient planning and control in offline goal-conditional reinforcement learning tasks.
LAPRAS : Learning-Augmented PRivate Answering for linear query Streams.
Pranay Mundra (Yale University), Quanquan C. Liu (Yale University)
Safty and PrivacyTabular
🎯 What it does: This paper proposes a learning-enhanced online differential privacy linear query answer framework, LAPRAS, which precomputes high-quality answers using a prediction set and dynamically allocates the remaining budget;
LaRA-Fusion: Latent-Robust Adaptation via Dual-Loop Constraints for Infrared and Visible Image Fusion
Yaru Su (Fuzhou University), Xiao Ke (Fuzhou University)
Image TranslationImage HarmonizationRestorationConvolutional Neural NetworkDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: By introducing an inner-loop geometric reversible and an outer-loop adversarial alignment dual-cycle Manifold constraint, potential space robust learning for infrared and visible light image fusion is achieved.
LARA: Latent Action Representation Alignment for Vision-Language-Action Models
Mengya Liu, Siyuan Huang
Representation LearningRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelFlow-based ModelAuto EncoderContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes the LARA framework, which jointly trains the Latent Action Model and a diffusion-based Vision-Language-Action model, achieving complementarity through representation alignment.
LARFT: Closing the Cognition-Action Gap for Length Instruction Following in Large Language Models
Wei Zhang (Beijing University of Posts and Telecommunications), Sen Su (Beijing University of Posts and Telecommunications)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose a length-aware reinforcement fine-tuning framework called LARFT, which enables large language models to internally recognize length while meeting length instructions and precisely output.
Large Language Model Agents Are Not Always Faithful Self-Evolvers
Weixiang Zhao (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
Explainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper conducts a systematic study on the faithfulness of self-evolving large language model (LLM) agents, employing causal intervention techniques to perturb both raw and condensed experiences, and performing comprehensive evaluations across four mainstream frameworks, 13 LLM backbones, and nine task environments.
Large Language Model Teaches Visual Students: Cross-Modality Transfer of Fine-Grained Conceptual Knowledge
Thomas Liang, Yong Jae Lee (University of Wisconsin Madison)
ClassificationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningImageText
🎯 What it does: Propose a cross-modal knowledge distillation framework called LaViD, which utilizes a language model to generate multiple-choice questions (MCQs) to extract fine-grained conceptual knowledge and distills it to a pure visual student model;
Large Language Models as Topological Thinkers: A Benchmark on Graph Persistent Homology
Hao Li (Wuhan University), Hao Jiang (Wuhan University)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmark
🎯 What it does: Proposed the LLM4PH benchmark to evaluate the multi-scale structural reasoning capabilities of large language models in graph persistent homology (PH), covering tasks from simple simplex identification to complex filtering strategy design and real-world graph classification.
Large Language Models Develop Novel Social Biases Through Adaptive Exploration
Addison J. Wu (Princeton University), Thomas L. Griffiths (Princeton University)
Recommendation SystemFederated LearningExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: In multi-round decision-making tasks, the study investigates how LLMs spontaneously generate new social biases through an exploration-exploitation balance in scenarios without prior differences, and validates this through recruitment experiments.
Large Language Models Explore by Latent Distilling
Yuanhao Zeng (ShanghaiTech University), Kan Ren (ShanghaiTech University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose an online lightweight potential distiller that combines exploratory sampling (ESamp) to encourage semantic diversity during LLM decoding
Large Scale Manifold Balanced Clustering
Fangfang Li (Xidian University), Xingyu Xue (Xidian University)
OptimizationComputational EfficiencyRepresentation LearningAuto EncoderContrastive LearningImagePoint CloudTabular
🎯 What it does: Proposed a large-scale manifold balanced clustering method (LMBC), which clusters directly in the anchor space using anchor-induced distance and achieves clustering balance by maximizing the Schatten-p norm.
Large Vision-Language Models Get Lost in Attention
Gongli Xi (Beijing University of Posts and Telecommunications), Wendong Wang (Beijing University of Posts and Telecommunications)
Information TheoryExplainability and InterpretabilityRepresentation LearningTransformerVision Language ModelMultimodality
🎯 What it does: Propose a unified framework based on information theory and geometry to quantify innovation and reorganization in Transformer residual flows, analyze the functions of attention and FFN in large vision-language models, and verify the redundancy of visual attention through alternative experiments.
Large-capacity and Receiver Authenticable Generative Image Steganography
Jiannian Wang (Harbin Institute of Technology (Shenzhen)), Guangming Lu (Harbin Institute of Technology (Shenzhen))
GenerationSafty and PrivacyConvolutional Neural NetworkDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Proposes a receiver-authenticable, high-capacity image steganography framework based on diffusion models, which can embed multiple secret images into a single generated image without using a cover image and achieve receiver isolation.
Large-Scale Molecular Dynamics Simulations: Direct Interatomic Modeling with Dilated Message Passing
Haokai Hong (Hong Kong Polytechnic University), KC Tan
Drug DiscoveryProtein Structure PredictionGraph Neural NetworkTransformerMixture of ExpertsContrastive LearningGraph
🎯 What it does: Propose a molecular dynamics simulation framework DKMP* based on dilated star-shaped message passing, enabling large-scale molecular dynamics simulations at the all-atom level
Large-Scale Notification Dispatch with Bundle Treatments and Multi-Outcome Uplift Optimization
Jiajing Xu (Kuaishou Technology), Yanan Niu (Kuaishou Technology)
Recommendation SystemOptimizationReinforcement LearningMixture of ExpertsContrastive LearningTabularTime Series
🎯 What it does: Proposed the BUOPLR method for bundled processing and multi-result optimization of large-scale push notifications, addressing the notification scheduling problem under platform budget and device manufacturer quota constraints.
Large-Scale Terminal Agentic Trajectory Generation from Dockerized Environments
Siwei Wu (University of Manchester), Chenghua Lin (University of Manchester)
Data SynthesisAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningAgentic AIScore-based ModelTextSequential
🎯 What it does: This paper proposes an expandable TerminalTraj data generation pipeline, using a Dockerized environment to construct and verify terminal interaction trajectories.
Large-scale Uncertainty Quantification for Latent Variable Models Using Subsampling Markov Chain Monte Carlo
Xiaoyu Wang (Boston University), Jonathan H. Huggins (Boston University)
OptimizationExplainability and InterpretabilityComputational EfficiencyHyperparameter SearchScore-based ModelContrastive LearningImageTextTabularBenchmarkStochastic Differential Equation
🎯 What it does: For models with local latent variables, the joint scaling limit of the SGLD-Gibbs algorithm is studied, and a method for adjusting hyperparameters that ensures statistical validity is provided.
LaRI: Layered Ray Intersections for Single-view 3D Geometric Reasoning
Rui Li (King Abdulaziz University), Peter Wonka
GenerationData SynthesisDepth EstimationConvolutional Neural NetworkTransformerNeural Radiance FieldAuto EncoderGaussian SplattingPoint CloudMesh
🎯 What it does: Proposes a single-view Layered Ray Intersections (LaRI) model that can predict multi-layer point maps of all camera rays intersecting the scene surface in a single forward pass, achieving complete reconstruction of both visible and occluded geometry.
LASER: Learning Active Sensing for Continuum Field Reconstruction
Huayu Deng (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
OptimizationTransformerReinforcement LearningDiffusion modelAuto EncoderWorld ModelImageTabularTime SeriesPhysics Related
🎯 What it does: This paper proposes the LASER framework, which achieves active sensing and high-precision reconstruction of continuous physical fields under sparse observations;
LassoFlexNet: a Flexible Neural Architecture for Tabular Data
Kry Yik-Chau Lui, Yanshuai Cao (Rbc Borealis)
ClassificationOptimizationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningMixture of ExpertsAuto EncoderContrastive LearningTabularTime SeriesSequentialBenchmark
🎯 What it does: LassoFlexNet is a deep network architecture specifically designed for tabular data, combining Per-Feature Embedding, Piecewise Linear Encoding (PLE), Tied Group Lasso, and MLP-Mixer, and equipped with the Seq-Hier-Prox-Adam-EMA optimizer, achieving end-to-end nonlinear feature selection and efficient learning.
Last-iterate Convergence of ADMM on Multi-affine Quadratic Equality Constrained Problem
Yutong Chao (Technical University of Munich), Majid Khadiv (Technical University of Munich)
OptimizationRobotic IntelligenceTabularTime Series
🎯 What it does: This paper investigates the sequential iterative convergence of ADMM for non-convex optimization problems with multiple transformed quadratic equality constraints, and provides theoretical guarantees for sublinear and linear convergence rates.
Last-Iterate Convergence of Regularized Gradient Methods for Stochastic Monotone Variational Inequalities
Shinji Ito (University of Tokyo), Kenshi Abe (CyberAgent)
Optimization
🎯 What it does: This paper studies the last-iteration convergence properties of randomly smooth monotone variational inequalities (VIs), proposes and analyzes two monotone methods: regularized gradient (RG) and regularized optimistic gradient (ROG), and provides convergence rates under anytime settings.
LAST: Bridging Vision-Language and Action Manifolds via Gromov-Wasserstein Alignment
Huaihai Lyu (MAIS Institute of Automation Chinese Academy of Sciences), Changsheng Xu (MAIS Institute of Automation Chinese Academy of Sciences)
Representation LearningRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelContrastive LearningMultimodalityPoint Cloud
🎯 What it does: Propose LAST (Lie-Algebraic Action Space Tokenizer), which discretizes robot action streams by performing global linearization on Lie-algebra (SE(3)) and local covariance correction, aligning robot action flows with visual-language semantic spaces geometrically and statistically, thereby achieving learnable representations for VLA models.
LaST$_{0}$: Latent Spatio-Temporal Chain-of-Thought for Robotic Vision-Language-Action Model
Zhuoyang Liu, Shanghang Zhang (Peking University)
Robotic IntelligenceTransformerMixture of ExpertsVision-Language-Action ModelFlow-based ModelImageVideoTextMultimodalityPoint CloudChain-of-Thought
🎯 What it does: Proposes LaST 0, a dual-system vision-language-action model that achieves efficient foresight reasoning and execution through implicit spatiotemporal chaining reasoning.
Latent Collaboration in Multi-Agent Systems
Jiaru Zou, Ling Yang
Computational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIAuto EncoderTextBenchmark
🎯 What it does: Proposed a multi-agent system framework called LatentMAS, which allows large language models (LLMs) to collaborate directly in a continuous latent space, rather than relying on text for reasoning and communication.
Latent Diffusion Pretraining for Crystal Property Prediction
Shrimon Mukherjee (Indian Association for the Cultivation of Science), Niloy Ganguly (Indian Institute of Technology Kharagpur)
Drug DiscoveryGraph Neural NetworkTransformerSupervised Fine-TuningDiffusion modelAuto EncoderContrastive LearningGraphTabularPhysics Related
🎯 What it does: Designed and implemented the CrysLDNet pre-training and fine-tuning framework, which maps crystal structures to a smooth latent space using VAE, then trains a diffusion model in this space, and finally fine-tunes on limited labeled data for crystal property prediction.
Latent Forcing: Reordering the Diffusion Trajectory for Pixel-Space Image Generation
Alan Baade (Stanford University), Li Fei-Fei (Stanford University)
GenerationTransformerDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: Propose Latent Forcing, which first generates latent representations obtained from self-supervised encoders (such as DINOv2) in the diffusion trajectory, then gradually diffuses pixels to build an end-to-end pixel space diffusion model; meanwhile, during training, independent time variables are used for different modalities and multi-step scheduling is applied.
Latent Guided Sampling for Combinatorial Optimization
Sobihan Surendran (Sorbonne Université and Université Paris Cité), Sylvain Le Corff (Sorbonne Université and Université Paris Cité)
OptimizationGraph Neural NetworkReinforcement LearningAuto EncoderGraphTabularBenchmark
🎯 What it does: LGS-Net proposes an instance-conditioned latent space model and achieves efficient inference through Latent Guided Sampling.
Latent Laplace Diffusion for Irregular Multivariate Time Series
Zinuo You (University of Bristol), John Cartlidge (University of Bristol)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelAuto EncoderTabularTime SeriesFinance RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose Latent Laplace Diffusion (LLapDiff), a continuous-time generative framework that performs diffusion modeling on low-dimensional latent trajectories, enabling long-horizon irregular multivariate time series prediction and missing value imputation without requiring integration over physical time steps.
Latent Reasoning in TRMs is Secretly a Policy Improvement Operator
Arip Asadulaev (Mohamed bin Zayed University of Artificial Intelligence), Martin Takáč
Computational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningReinforcement LearningDiffusion modelImageTextChain-of-Thought
🎯 What it does: Deep reasoning on small-scale cyclic Transformer (TRM) reveals that it implicitly implements policy improvement, based on which the Deep Improvement Supervision (DIS) training framework is proposed. This framework uses discrete diffusion/erosion strategies to generate progressive goals, explicitly providing improvement supervision at each step.
Latent Reasoning VLA: Latent Thinking and Prediction for Vision-Language-Action Models
Shuanghao Bai (Xi'an Jiaotong University), Shanghang Zhang (Peking University)
Computational EfficiencyRepresentation LearningRobotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelFlow-based ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Propose a Vision-Language-Action model called LaRA-VLA, which internalizes chain-of-thought (CoT) reasoning into continuous latent representations, eliminating the overhead of explicitly generating CoT during inference;
Latent Representation Alignment for Offline Goal-Conditioned Reinforcement Learning
Hyungkyu Kang (Seoul National University), Min-hwan Oh (Seoul National University)
Reinforcement LearningContrastive LearningTabularTime SeriesSequentialBenchmark
🎯 What it does: This paper studies offline goal-oriented reinforcement learning, proposing the LAVL algorithm to address the performance bottleneck caused by overgeneralization of the value function;
Latent Space Robust Optimization of Neural Processes with Aligned Stratified Order-Statistic Loss Reduction
Qi Tao (National University of Defense Technology), Qi Wang (National University of Defense Technology)
OptimizationRepresentation LearningData-Centric LearningRobotic IntelligenceDiffusion modelScore-based ModelFlow-based ModelRectified FlowContrastive LearningGaussian SplattingImageTabularTime SeriesSequentialBenchmark
🎯 What it does: Propose OS-NPs, which achieve robust learning in the latent space by stratifying the sampling particles of IWNP according to quantiles, and by using regularized extreme case optimization, balancing average performance and tail risk;
Latent Spherical Flow Policy for Reinforcement Learning with Combinatorial Actions
Lingkai Kong (Harvard University), Milind Tambe (Harvard University)
OptimizationReinforcement LearningFlow-based ModelGraphTabularBenchmark
🎯 What it does: Propose a stochastic policy based on Latent Spherical Flow, using a solver to map sampled cost directions to feasible actions that satisfy combinatorial constraints, combining expressiveness and feasibility;
Latent Thoughts Tuning: Bridging Context and Reasoning with Fused Information in Latent Tokens
Weihao Liu (University of Illinois Chicago), Lu Cheng (University of Illinois Chicago)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the Latent Thoughts Tuning (LT-Tuning) framework, enabling large language models to perform stable and dynamic reasoning in a continuous latent space.
Latent-Guided Cooperative Energy-Based Models
Cong Geng (China Mobile Jiutian Artificial Intelligence Technology Company Limited), Junlan Feng (China Mobile Jiutian Artificial Intelligence Technology Company Limited)
GenerationData SynthesisAnomaly DetectionTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningImage
🎯 What it does: A jointly trained latent-guided collaborative energy-based model (LGCEBM) was constructed, where potential information is provided by a pre-trained self-supervised encoder to guide the energy function, and an auxiliary generator is used to achieve efficient MCMC initialization.
LatentChem: From Textual CoT to Latent Thinking in Chemical Reasoning
Xinwu Ye (University of Hong Kong), Xiangru Tang (Yale University)
Explainability and InterpretabilityComputational EfficiencyDrug DiscoveryReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextGraphChain-of-Thought
🎯 What it does: Proposed a new chemical reasoning interface called LatentChem, which transforms traditional text-based chain-of-thought (CoT) reasoning into continuous latent space reasoning, eliminating the problem of modality mismatch;
LatentLens: Revealing Highly Interpretable Visual Tokens in LLMs
Benno Krojer (Mila - Quebec AI Institute), Marius Mosbach
Explainability and InterpretabilityTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Develop and evaluate a new method for interpreting visual tokens — LatentLens, which compares the representations of visual tokens in large language models (LLMs) with contextualized word embeddings from a large text corpus, providing sentence-level interpretable descriptions.
LATMiX: Learnable Affine Transformations for Microscaling Quantization of LLMs
Ofir Gordon (Arm), Hai Victor Habi (Arm)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Propose a learnable affine transformation (LATMiX) for micro-scale quantization in LLMs, reducing activation outliers and improving inference accuracy at low bit precision.
LATO: 3D Mesh Flow Matching with Structured TOpology Preserving LAtents
Tianhao Zhao (Huazhong University of Science and Technology), Wei Yang (Huazhong University of Science and Technology)
GenerationData SynthesisTransformerDiffusion modelFlow-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningPoint CloudMesh
🎯 What it does: Propose LATO, a topology-preserving 3D mesh generation framework combining sparse voxel VAE and flow matching, which directly learns and recovers mesh topology and geometry in the latent space.
LaTtE-Flow: Layerwise Timestep-Expert Flow-based Transformer
Ying Shen (University of Illinois Urbana Champaign), Lifu Huang (UC Davis)
GenerationComputational EfficiencyRepresentation LearningTransformerMixture of ExpertsVision Language ModelFlow-based ModelAuto EncoderImageTextMultimodality
🎯 What it does: Proposed a unified multimodal model called LaTtE-Flow, which can simultaneously perform image understanding and generation under the same architecture.
LAVA: A Unified Framework for Finetuning Language and Vision Models
Daorui Ding (Tianjin University), Hongying Liu (Tianjin University)
ClassificationSegmentationGenerationDepth EstimationRepresentation LearningConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkAgriculture Related
🎯 What it does: Propose a unified parameter-efficient fine-tuning framework called LAVA, which can be simultaneously applied to the attention modules of language models and the convolution modules of vision models.
Lavida-R1: Advancing Reasoning for Unified Multimodal Diffusion Language Models
Shufan Li (Adobe), Jason Kuen (Adobe)
GenerationData SynthesisRepresentation LearningReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkImageTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a unified post-training framework LaViDa-R1 to enhance the reasoning capabilities of multi-modal diffusion language models, covering multiple tasks such as visual math reasoning, visual question answering, image editing, and object localization;
Layer-Centric Factors of Variation Disentanglement for Task- and Model-Agnostic Generalization
Hee-Jun Jung (Gwangju Institute of Science and Technology), Kangil Kim (Gwangju Institute of Science and Technology)
ClassificationObject DetectionSegmentationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningImageText
🎯 What it does: Propose a pluggable Orthogonal Subspaces Projection (OSP) layer, which utilizes orthogonal projection in the middle layers of the network to achieve factor separation, and verifies its transferability and generality across multiple tasks and models.
Layer-wise Gradient Disentanglement: Decoupling Semantics and Preferences in Direct Preference Optimization
Mengyang Li (Tianjin Normal University), Zhong Zhang (Tianjin Normal University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes GDO-DPO, a curriculum learning framework based on hierarchical gradient states, used to decouple two learning objectives—semantic understanding and preference discrimination—in direct preference optimization (DPO);
LayerT2V: A Unified Multi-Layer Video Generation Framework
Guangzhao Li (Shanghai Jiao Tong University), Xiaohong Liu (Shanghai Jiao Tong University)
GenerationData SynthesisTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelFlow-based ModelRectified FlowAuto EncoderImageVideoText
🎯 What it does: Propose LayerT2V, a unified multi-layer video generation framework that can generate complete videos, background layers, and multiple foreground layers along with their corresponding alpha masks in a single inference.
LazyAttention: Efficient Retrieval-Augmented Generation with Deferred Positional Encoding
Haocheng Xia (University of Illinois Urbana-Champaign), Yongjoo Park (University of Illinois Urbana-Champaign)
RetrievalComputational EfficiencyTransformerTextRetrieval-Augmented Generation
🎯 What it does: Propose LazyAttention, achieving efficient retrieval-augmented generation with position-independent KV cache through delayed position encoding
LC-QAT: Data-Efficient 2-Bit QAT for LLMs via Linear-Constrained Vector Quantization
Haoyu Wang (Tsinghua University), Xu Han (Tsinghua University)
Computational EfficiencyKnowledge DistillationData-Centric LearningTransformerLarge Language ModelAuto EncoderContrastive LearningText
🎯 What it does: Propose LC-QAT, a 2-bit weight quantized vector quantization QAT framework, which utilizes a linearly constrained codebook to achieve lookup-free, differentiable end-to-end training, and combines strong PTQ initialization with differentiable gradient estimation to achieve data-efficient fine-tuning.
LDARNet: DNA Adaptive Representation Network with Learnable Tokenization for Genomic Modeling
Daria Ledneva (Moscow Independent Research Institute of Artificial Intelligence), Denis Kuznetsov (Moscow Independent Research Institute of Artificial Intelligence)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningBiomedical DataBenchmark
🎯 What it does: Propose LDARNet, a hierarchical DNA foundation model with 120M parameters, which uses self-learning dynamic chunking (adaptive tokenization) for masked language modeling (MLM) pre-training and is fine-tuned on 27 genomic downstream tasks.
Leaderboard Incentives: Model Rankings under Strategic Post-Training
Yatong Chen (Max Planck Institute for Intelligent Systems), Moritz Hardt (Max Planck Institute for Intelligent Systems)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper models benchmark evaluation as a leader-follower (Stackelberg) game, analyzing the incentive structure generated when model submitters perform targeted post-training on leaderboards. It proves that traditional benchmarks often lead to no Nash equilibrium and leaderboard distortion, while the proposed Tune-before-test mechanism can achieve a unique equilibrium and rank models according to their potential capabilities under certain conditions.
Leak@$k$: Unlearning Does Not Make LLMs Forget Under Probabilistic Decoding
Hadi Reisizadeh (University of Minnesota), Mingyi Hong (University of Minnesota)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Evaluate knowledge leakage of large language models under probabilistic decoding conditions after unlearning, and propose a new evaluation metric leak@k and the corresponding robust unlearning algorithm RULE.
LeakGFN: Robust Molecular Generation in Generative Flow Networks via Flow Decomposition
Hwanhee Kim (Yonsei University), Sanghyun Park (Yonsei University)
Drug DiscoveryGraph Neural NetworkScore-based ModelFlow-based ModelGenerative Adversarial NetworkGraphBiomedical Data
🎯 What it does: Propose LeakGFN, a dual-headed generative flow network structure, which solves the flow leakage problem caused by trajectory truncation in molecular generation by decomposing the flow into chemical flow and effective flow.
LEAP: Zone-Aware MCTS for LLM Self-Speculative Decoding
LeiQuan Zheng, Yuan Liu (South China University of Technology)
Computational EfficiencyReinforcement Learning from Human FeedbackNeural Architecture SearchTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Construct a draft model through self-speculative decoding, and accelerate LLM inference by performing online layer configuration search using Monte Carlo Tree Search (MCTS) based on zone-awareness and layer grouping.
Learn from A Rationalist: Distilling Intermediate Interpretable Rationales
Jiayi Dai (University of Alberta), Randy Goebel (University of Alberta)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningImageText
🎯 What it does: This paper proposes the REKD (Rationale Extraction with Knowledge Distillation) framework, which utilizes the interpretability and verifiable intermediate 'rationale' from the teacher model to guide the student model's feature selection and prediction, thus addressing the 'chicken and egg' dilemma faced by lightweight models during rationale extraction.