ICML 2026 Papers — Page 32
International Conference on Machine Learning · 6554 papers
Learning to Extrapolate to New Tasks: A Relational Approach to Task Extrapolation
Adam Ousherovitch (University of Michigan), Yixin Wang (University of Michigan)
Meta LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextSequential
🎯 What it does: Propose a Relational Task Extrapolator (RTE) that leverages inter-task relationships for extrapolation, achieving predictions on new tasks by decomposing them into known anchor tasks and corresponding transformations.
Learning to Label: A Reinforced Self-Evolving Framework for Semi-supervised Referring Expression Segmentation
Runlong Cao (Nanjing University of Science and Technology), Chunyan Xu (Nanjing University of Science and Technology)
SegmentationTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Propose a semi-supervised reference expression segmentation framework called L2L, which transforms pseudo-label generation into a learnable decision process and continuously improves the reliability of pseudo-labels within a self-evolving closed-loop.
Learning to Memorize with Attributive and Associative Memory for Online Test-Time Adaptation of Vision-Language Models
Yuchao Zhang (Peking University), Zhouchen Lin (Peking University)
Domain AdaptationComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposed a new online test-time adaptation framework called A Memory, which improves test-time adaptation of vision-language models through attribute-oriented soft assignment and class-level associative memory.
Learning to Move Before Learning to Do: Task-Agnostic pretraining for VLAs
Junhao Shi (Fudan University), Xipeng Qiu (Fudan University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerVision-Language-Action ModelContrastive LearningVideoTextMultimodality
🎯 What it does: Proposes a two-stage framework called Task-Agnostic Pretraining (TAP), first pretraining through inverse dynamics self-supervised learning to learn 'how to move', and then aligning with language instructions using a small amount of expert demonstrations;
Learning to Perceive the World Through Control: Empowerment-Based Representation Learning
Mahsa Bastankhah (Princeton University), Benjamin Eysenbach (Princeton University)
Representation LearningReinforcement LearningContrastive LearningImagePoint CloudTabularTime Series
🎯 What it does: This paper investigates and theoretically analyzes whether state representations containing only control-related features can be learned without relying on reward signals, by using the empowerment (mutual information skill learning) objective in reinforcement learning, and experiments are conducted on multiple high-dimensional environments to validate the method.
Learning to Rank by Directly Optimizing Full-Order Probabilities
Yongxiang Tang (Kuaishou Technology), Peng Jiang (Kuaishou Technology)
Recommendation SystemOptimizationComputational EfficiencyScore-based ModelContrastive LearningTextTabularBenchmark
🎯 What it does: This paper proposes Full-Order Bound (FOB), which approximates and directly optimizes the probability of complete ranking events by introducing segmented thresholds to the latent Gaussian scores.
Learning to Rank from Incomplete Rankings
Cristiano Migali (Politecnico di Milano), Marco Mussi (Politecnico di Milano)
Recommendation SystemOptimizationData-Centric LearningSupervised Fine-TuningContrastive LearningGraphTabular
🎯 What it does: Designed a general framework and algorithm called PIRATE, which can learn the top-k items under incomplete rankings and satisfy the MNAR (Missing Not At Random) condition, and provided an approximate optimal sample complexity analysis.
Learning to Reason for Factuality
Xilun Chen (FAIR at Meta), Wen-tau Yih (FAIR at Meta)
GenerationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper designs a multi-dimensional reward function that simultaneously considers factual accuracy, answer detail, and relevance by applying online reinforcement learning on reasoning large language models (R-LLM), and implements a parallelizable VeriScore to improve long-form factual generation.
Learning to Reconfigure: Configuration-Control Co-optimization of Reconfigurable Robots for Heterogeneous Locomotion
Xiaoyu Xiong (Shanghai Qi Zhi Institute), Tao Du (Tsinghua University)
OptimizationRobotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningSimultaneous Localization and MappingPoint CloudTabularTime Series
🎯 What it does: Proposed and implemented a hierarchical collaborative optimization framework named Learning to Reconfigure, which automatically generates reconfigurable robot configurations and corresponding control strategies under multi-physical environments (land, air, water).
Learning to Refine: Spectral-Decoupled Iterative Refinement Framework for Precipitation Nowcasting
Yunlong Zhou (Nanjing University), Xiaotong Yuan
OptimizationRecurrent Neural NetworkTransformerDiffusion modelAuto EncoderContrastive LearningVideoTime SeriesPhysics Related
🎯 What it does: A deterministic, spectrum-decoupled iterative refinement framework called SDIR is proposed for high-resolution precipitation nowcasting.
Learning to Remember, Learn, and Forget in Attention-Based Models
Djohan Bonnet (Forschungszentrum Juelich), Emre Neftci (Forschungszentrum Juelich)
Computational EfficiencyKnowledge DistillationRepresentation LearningMeta LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextSequentialBenchmark
🎯 What it does: Propose an attention mechanism called Palimpsa based on Bayesian meta-plasticity to address the stability-plasticity dilemma in continuous learning with fixed-size memory.
Learning to Route Languages for Multilingual Policy Optimization
Geyang Guo (Georgia Institute of Technology), Wei Xu (Georgia Institute of Technology)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextMultimodalityBenchmark
🎯 What it does: This study proposes a language routing-based multilingual policy optimization framework called LRPO, which allows the model to actively select the response language during training and generate diverse training signals through multilingual rollouts.
Learning to Search and Searching to Learn for Generalization in Planning
Michael Aichmüller, Hector Geffner (RWTH Aachen University)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper proposes a self-improving search framework (GSP) based on weighted A* and relational graph neural networks, which learns a general Q-value by generating training data through search to achieve search-free solutions for different instances.
Learning to Self-Verify Makes Language Models Better Reasoners
Yuxin Chen (National University Of Singapore), Tat-Seng Chua (National University Of Singapore)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmarkChain-of-Thought
🎯 What it does: The study investigates the capability asymmetry between generation and self-verification in large language models, and proposes enhancing the generation performance by training models to self-verify;
Learning to Share: Selective Memory for Efficient Parallel Agentic Systems
Joseph Fioresi (University of Central Florida), Mubarak Shah (University of Central Florida)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented Learning to Share (LTS), a method that uses a global shared memory and a learnable controller in parallel agent systems to selectively share intermediate results, thereby reducing redundant computations and improving execution efficiency.
Learning to Theorize the World from Observation
Doojin Baek (KAIST), Sungjin Ahn (KAIST)
Explainability and InterpretabilityRepresentation LearningRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerAuto EncoderImageTextSequentialBenchmark
🎯 What it does: Studied a framework for learning executable interpretable theories (programs) from raw perceptual data, called Learning‑to‑Theorize (L2T) and its implementation model Neural Theorizer (NEO).
Learning to Think in Physics: Breaking Shortcut Learning in Scientific Diffusion via Representation Alignment
Haozhe Jia (Hong Kong University of Science and Technology (Guangzhou)), Yutao Yue (Jiangsu Industrial Technology Research Institute)
Convolutional Neural NetworkTransformerDiffusion modelScore-based ModelMeshPhysics Related
🎯 What it does: This paper inserts a lightweight projection head into the intermediate layer of a diffusion model, mapping the latent representation to the physical state and applying PDE residual constraints during training, allowing the model to 'think' about physical laws during the generation process.
Learning to Watch: Active Video Anomaly Understanding via Interleaved Policy Optimization
Mengjingcheng Mo (Chongqing University of Posts and Telecommunications), Xinbo Gao (Chongqing University of Posts and Telecommunications)
Anomaly DetectionTransformerReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningVideoText
🎯 What it does: By constructing the Anomπ framework, video anomaly understanding is modeled as an active closed-loop sequential decision-making process, combining interactive evidence acquisition and reasoning.
Learning to Watermark in the Latent Space of Generative Models
Sylvestre-Alvise Rebuffi (Meta FAIR), Alexandre Mourachko (Meta FAIR)
GenerationData SynthesisSafty and PrivacyKnowledge DistillationAdversarial AttackTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Learn and embed invisible watermarks in the latent spaces of diffusion models and autoregressive models, while migrating the watermark into the model through distillation, achieving watermarking without additional inference overhead within the model.
Learning Transferable Interaction Primitives from Game Videos for Humanoid Locomotion
Xiangming Zhu (Shanghai Jiao Tong University), Yunbo Wang (Shanghai Jiao Tong University)
Robotic IntelligenceTransformerReinforcement LearningVision-Language-Action ModelDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningVideoSequential
🎯 What it does: Extract and discretize human-environment interaction primitives from unlabeled game videos, and execute these primitives in target physical environments by leveraging a cross-domain aligned shared physical context space, thereby achieving reinforcement learning control for human-like dynamic walking tasks.
Learning Treatment Allocations with Risk Control Under Partial Identifiability
Sofia Ek (Uppsala University), Dave Zachariah (Uppsala University)
OptimizationFederated LearningExplainability and InterpretabilityReinforcement Learning from Human FeedbackTabularReview/Survey PaperBenchmark
🎯 What it does: Proposes a method for learning treatment allocation under partially identifiable scenarios, which can ensure that treatment risk does not exceed a preset tolerance level under limited sample conditions, while minimizing overall risk.
Learning Treatment Representations for Downstream Instrumental Variable Regression
Shiangyi Lin (Stanford University), Vasilis Syrgkanis (Stanford University)
Explainability and InterpretabilityRepresentation LearningDrug DiscoveryDiffusion modelScore-based ModelAuto EncoderContrastive LearningGaussian SplattingImageTabularBiomedical DataElectronic Health RecordsStochastic Differential Equation
🎯 What it does: Proposed a representation learning method guided by instrumental variables, which is used to identify causal effects and implement treatment interventions, even in the presence of high-dimensional treatment variables and unobserved confounding.
Learning Unanimously Acceptable Lotteries via Queries
Davin Choo (Harvard University), Nicholas Teh (University of Oxford)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackReinforcement LearningContrastive Learning
🎯 What it does: Proposes an algorithm and theoretical analysis for finding a randomized deployment scheme (i.e., 'uniformly acceptable lottery') that satisfies all stakeholders' thresholds under a query model that only receives binary feedback (accept/reject);
Learning Unmasking Policies for Diffusion Language Models
Metod Jazbec (University of Amsterdam), marco cuturi
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelText
🎯 What it does: This study proposes a lightweight sampling strategy trained using reinforcement learning, which automatically learns the unmasked order of masked diffusion language models (dLLMs), replacing heuristic sampling methods such as manually adjusted confidence thresholds.
Learning What to Generate: A Reinforcement Learning-based Closed-Loop Augmentation Framework for Person Re-identification
Xincheng Shi (Beihang University), Rongye Shi (Beihang University)
RecognitionGenerationData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringDiffusion modelContrastive LearningImageMultimodalityChain-of-Thought
🎯 What it does: This study proposes a closed-loop reinforcement learning framework called ReasonAug, which learns optimal editing instructions for existing frozen image generators, thereby generating more discriminative training samples without altering the generator;
Learning When to Act or Refuse: Guarding Agentic Reasoning Models for Safe Multi-Step Tool Use
Aradhye Agarwal (Microsoft Research), Ahmed Hassan Awadallah
Safty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Propose the MOSAIC framework, which divides the reasoning of the proxy model into a Plan → Check → Act/Refuse cycle, and explicitly performs safety evaluation and refusal decisions during the Check phase.
Learning When to Attend: Conditional Memory Access for Long-Context LLMs
Sakshi Choudhary (Purdue University), Stefano Soatto (AWS Agentic AI)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Designed the L2A layer, which adaptively decides whether to invoke global attention in long-context LLMs by using token-level routing, thus significantly reducing computational cost while maintaining performance.
Learning with Admissibility: Robust Fuzzy Hashing for Cross-Modal Retrieval with Noisy Labels
Xincheng Sun (Sichuan University), Yuan Sun (Sichuan University)
RetrievalContrastive LearningMultimodality
🎯 What it does: Proposes a Robust Fuzzy Cross-Modal Hashing (RFCMH) framework based on fuzzy set theory to address the label noise problem in cross-modal retrieval;
Learning-Augmented Online Covering Problems
Afrouz Jabal Ameli (Utrecht University), Moritz Venzin (Bocconi University)
OptimizationTabularBenchmark
🎯 What it does: Propose a general black-box framework (ICE) that incorporates predictions of requests for online covering problems into algorithms, achieving a smooth improvement in competitive ratio with prediction error η through an offline decomposition and online charging mechanism.
Learning-Augmented Online Minimization with Dual Predictions
Christian Coester (University of Oxford), Alexander Turoczy (University of Oxford)
OptimizationTabularTime Series
🎯 What it does: This paper studies the learning-enhanced online minimization problem, proposing an algorithm that uses dual prediction. For layered set cover and metric task systems, it proves the stability and learnability of the algorithm, and verifies its effectiveness on the k-server and parking permit problems through experiments.
Learning-augmented Rent-or-Buy with a Sample
Davidson Zhu (Duke University), Debmalya Panigrahi (Duke University)
OptimizationTabularFinance Related
🎯 What it does: Proposed and analyzed two online algorithms (Algorithm 5.2 and Algorithm 5.3) for solving the lease-or-buy (Bahncard) problem under the premise of having random samples, significantly reducing the competitive ratio from the traditional 2 down to 1.5, and even approaching 1;
Learning-Augmented Scalable Linear Assignment Problem Optimization via Neural Dual Warm-Starts
Ilay Yavlovich (Technion Israel Institute of Technology), Jose Yallouz (Technion Israel Institute of Technology)
OptimizationGraph Neural NetworkTransformerReinforcement LearningAuto EncoderContrastive LearningGraphTabular
🎯 What it does: Use deep learning models to predict dual variables for linear assignment problems to warm start traditional exact solvers.
Learning-Guided Integration Contours Construction for Fast Large-Scale Generalized Eigensolvers
Yeqiu Chen (University of Science and Technology of China), Lei Liu (University of Science and Technology of China)
OptimizationComputational EfficiencyGraph Neural NetworkTransformerDiffusion modelScore-based ModelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningGraphTabularTime SeriesPhysics Related
🎯 What it does: Proposed a framework called Deepcontour, which combines deep learning-based spectral prediction with KDE (kernel density estimation) for automatically constructing integral contours, to accelerate the solution of large-scale generalized eigenvalue problems.
Learning-To-Measure: In-Context Active Feature Acquisition
Yuta Kobayashi (Columbia University), Shalmali Joshi (Columbia University)
Representation LearningData-Centric LearningMeta LearningTransformerReinforcement LearningAuto EncoderContrastive LearningImageTabularBiomedical DataElectronic Health Records
🎯 What it does: Proposes L2M (a sequence model framework based on meta-learning) for active feature acquisition in scenarios with missing values, capable of performing without retraining for each task during testing.
Learning-to-Optimize via Deep Unfolded Flows
Augustinos D Saravanos, Chuchu Fan (Massachusetts Institute of Technology)
OptimizationFlow-based ModelTabularTime SeriesSequentialBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a learning-based optimization framework called FlowOptimizer, which utilizes flow models to continuously transform the candidate solution set by learning a velocity field, achieving self-supervised iterative optimization.
Learning, Solving and Optimizing PDEs with TensorGalerkin: an efficient high-performance Galerkin assembly algorithm
Shizheng Wen (ETH Zurich), Siddhartha Mishra (ETH Zurich)
OptimizationAuto EncoderMeshPhysics Related
🎯 What it does: Proposed and implemented a Galerkin assembly framework called TENSORGALERKIN based on tensorized Map-Reduce, which can be used for efficient solution, physics-informed learning, and PDE-constrained optimization.
Least-Loaded Expert Parallelism: Load Balancing An Imbalanced Mixture-of-Experts
Xuan-Phi Nguyen (Salesforce AI Research), Shafiq Joty (Salesforce AI Research)
OptimizationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Proposed the Least-Loaded Expert Parallelism (LLEP) algorithm to dynamically balance the load and address routing imbalance issues in Mixture-of-Experts models.
LEC: Linear Expectation Constraints for Selection-Conditioned Risk Control in Selective Prediction and Routing Systems
Zhiyuan Wang (University of Electronic Science and Technology of China), Kaidi Xu (City University of Hong Kong)
OptimizationFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelReinforcement LearningVision Language ModelContrastive LearningTextMultimodalityBenchmark
🎯 What it does: A framework based on linear expectation constraints (LEC) is proposed to ensure that the error probability of accepted predictions does not exceed a user-specified risk threshold in both single-model selection prediction and multi-model routing systems.
LECTOR: Joint Learning of Scientific Reasoning Graphs and Introduction Generation
Jiabei Xiao (Chinese University of Hong Kong), SHIXIANG TANG
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextGraphReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the Content-based Scientific Paper Introduction Generation task (CCIG), generating introductions that are logically consistent with the main chapters and cite reliable sources by constructing logical reasoning graphs and using logic-expression co-reinforcement learning;
Left–Right Symmetry Breaking in CLIP-Style Vision-Language Models Trained on Synthetic Spatial-Relation Data
Takaki Yamamoto (InfoTech Toyota Motor Corporation), Toshihiro Tanizawa (InfoTech Toyota Motor Corporation)
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: By training a CLIP-style Transformer on controllable 1D synthetic image-text data, the study investigates how the model learns and generalizes left-right spatial relationships;
LEGO-FL: Learning Heterogeneous Federated Models as a LEGO Assembly Games
Zeqi Leng (Jilin University), Bo Yang (Jilin University)
OptimizationFederated LearningKnowledge DistillationRepresentation LearningContrastive LearningImage
🎯 What it does: Reimagine Heterogeneous Federated Learning by splitting and recombining models in the form of LEGO components, proposing the LEGO-FL framework
LEGO: An LLM-Enabled Hierarchical Optimizer for Tensor Computation Graphs with Structure-Aware Search and Compositional Synthesis
Ruiyuan Xu (Chinese Academy of Sciences), Huimin Cui (Chinese Academy of Sciences)
OptimizationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringMixture of ExpertsImageTextGraph
🎯 What it does: Construct a hierarchical optimization framework called LEGO based on large language models, used to automatically synthesize and optimize end-to-end tensor computation graphs on GPUs.
LEMUR: Learned Multi-Vector Retrieval
Elias Jääsaari (University of Helsinki), Teemu Roos (University of Helsinki)
RetrievalComputational EfficiencyTransformerSupervised Fine-TuningContrastive LearningImageText
🎯 What it does: Proposed the LEMUR framework, which transforms multi-vector retrieval tasks into supervised learning and then remaps them to single-vector retrieval, thereby significantly reducing retrieval latency.
Length Generalization Bounds for Transformers
Andy Yang (University of Notre Dame), Anthony Widjaja Lin
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelContrastive Learning
🎯 What it does: Demonstrated that the length generalization of Transformer does not have a computable upper bound in general cases, and provided the length generalization complexity of C RASP‑ and Transformer.
LERD: Latent Event-Relational Dynamics for Neurodegenerative Classification
Yicheng Feng (University of Copenhagen), Hengguan Huang (University of Copenhagen)
ClassificationRecurrent Neural NetworkTransformerAuto EncoderContrastive LearningTime SeriesBiomedical DataAlzheimer's DiseaseStochastic Differential EquationOrdinary Differential EquationAudio
🎯 What it does: Propose an unsupervised Bayesian latent event relational dynamics model, LERD, for simultaneously recovering implicit event temporal sequences, event-to-event timing relational graphs (ERG), and a predictor for Alzheimer's disease (AD) classification from multi-channel EEG.
Less Data, Faster Training: repeating smaller datasets speeds up learning via sampling biases
Jingwen Liu (Columbia University), Bingbin Liu (Harvard University)
OptimizationComputational EfficiencyData-Centric LearningTransformer
🎯 What it does: Investigated and verified the 'few-data fast training' phenomenon, which indicates that under a fixed computational budget, repeatedly using a smaller dataset can converge faster than using a larger dataset, and explored the reasons behind this.
Less Diverse, Less Safe: The Indirect But Pervasive Risk of Test-Time Scaling in Large Language Models
Shahriar Kabir Nahin (University of South Florida), Anshuman Chhabra (University of South Florida)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringGenerative Adversarial NetworkContrastive LearningText
🎯 What it does: This paper proposes a reference-guided diversity compression protocol called REFIV, which compresses the diversity of candidate answers in Test-Time Scaling (TTS), thereby inducing LLMs to generate unsafe outputs during the TTS process.
Less is Enough: Synthesizing Diverse Data in LLM Feature Space with Sparse Autoencoders
Zhongzhi Li (University of Georgia), Ninghao Liu (Hong Kong Polytechnic University)
Data SynthesisExplainability and InterpretabilityRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringAuto EncoderContrastive LearningText
🎯 What it does: This paper proposes a feature activation coverage (FAC) metric based on sparse autoencoders (SAE) to measure the diversity of data in the internal feature space of LLMs, and designs the FAC Synthesis framework. It first identifies task-related features missing in the seed data, then generates synthetic samples that explicitly activate these features, thereby constructing diverse and high-quality post-training data.
Less Is More in Federated Continual Learning: RieSelect for Conflict-Aware Layer Selection in LLMs
Wenqi Qiu (Shenzhen University), Laizhong Cui (Shenzhen University)
Federated LearningComputational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Study the federated continual learning of large language models on edge devices, proposing the RieSelect method to achieve sparse, conflict-aware layer uploading;
Less Is More: Elevating RAG via Performance-Driven Context Compression
Ziqiang Cui (City University of Hong Kong), Chen Ma (City University of Hong Kong)
RetrievalCompressionComputational EfficiencyKnowledge DistillationTransformerReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a context compression framework named CORE-RAG for RAG, which directly optimizes the compressor using performance-driven reinforcement learning to generate compressed documents that align with task objectives.
Less Is More: Fast and Accurate Reasoning with Cross-Head Unified Sparse Attention
Lijie Yang (Princeton University), Ravi Netravali (Princeton University)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsText
🎯 What it does: Proposes LessIsMore, a training-agnostic sparse attention mechanism that improves decoding efficiency in long reasoning models by leveraging cross-head unified token selection and stable nearest window.
Less is More: Geometric Unlearning for LLMs with Minimal Data Disclosure
Chenchen Tan (Monash University), Longxiang Gao (Qilu University of Technology)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Propose a geometry-based LLM unlearning method (Geometric Unlearning), which eliminates specific knowledge by projecting and aligning target entities in the prompt-conditioned hidden state space, without needing access to the original training corpus.
Less Precise Can Be More Reliable: A Systematic Evaluation of Quantization’s Impact on VLMs Beyond Accuracy
Aymen Bouguerra (Universit´ e Paris-Saclay), Fabio Arnez (Universit´ e Paris-Saclay)
ClassificationCompressionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper conducts a systematic quantitative evaluation of Vision-Language Models (VLMs), designing over 700k evaluation runs covering five reliability dimensions (robustness, calibration, OOD detection, distribution drift, and pulse correlation), and analyzes the impact of different quantization strategies on these metrics.
Less Token, More Signal: MoE Expert Pruning via Critical Token Selection
Zeliang Zong (Hikvision Research Institute), Jilin Hu (East China Normal University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose a Mixture-of-Experts (MoE) expert pruning method called STEP based on key token selection, which uses attention-guided token importance to filter noise, and combines dual-factor expert importance scoring and expert-to-bias knowledge preservation techniques to achieve efficient compression;
Let EEG Models Learn EEG
Yifan Wang (Stony Brook University), Chenyu You (Stony Brook University)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelFlow-based ModelTime SeriesBiomedical Data
🎯 What it does: Designed a Transformer model based on flow matching (JET) for high-fidelity EEG signal generation.
Let Language Constrain Geometry: Vision–Language Models as Semantic and Spatial Critics for 3D Generation
Weimin Bai (Peking University), He Sun (Peking University)
GenerationData SynthesisOptimizationTransformerReinforcement LearningVision Language ModelDiffusion modelScore-based ModelNeural Radiance FieldGaussian SplattingImageTextPoint CloudMeshBenchmark
🎯 What it does: This paper proposes the VLM3D framework, which utilizes a large vision-language model (VLM) as a differentiable semantic and spatial dual evaluator to guide the 3D generation process;
Let the Prototype Guide You: Robust Aggregation of Sparse Multi-Class Annotations via Annotator Prototype Learning
Ju Chen (Hohai University), Shenyu Zhang (Southeast University)
ClassificationData-Centric LearningImageTextTabular
🎯 What it does: Propose a class-specific prototype-based Bayesian classifier combination (CPBCC) for robust aggregation of sparse multi-class annotations.
Letting Trajectories Spread: Quality-Preserving Control for Diverse Flow Matching
Jingxuan Wu (University of North Carolina at Chapel Hill), Yang You (National University of Singapore)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowImageTextMultimodality
🎯 What it does: Propose a training-free, inference-time control mechanism that utilizes orthogonal noise and feature space volume expansion to enhance the generative diversity of flow matching models;
Leveraging Evidence Priors for Robust Prompt Learning under Noisy Supervision in Vision-Language Models
Junnan Zou (Beijing Jiaotong University), Jianping Fan (Lenovo)
ClassificationRecognitionDomain AdaptationData-Centric LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: The study investigates using evidence priors to enhance the robustness of prompt learning in vision-language models under noisy labels.
Leveraging Gauge Freedom for Learning Non-Gradient Population Dynamics of Stochastic Systems
Jules Berman (New York University), Benjamin Peherstorfer (New York University)
Diffusion modelScore-based ModelAuto EncoderContrastive LearningPoint CloudTabularTime SeriesPhysics RelatedStochastic Differential Equation
🎯 What it does: Propose an algorithm called NGIF that learns non-gradient population dynamics by utilizing the weak continuity equation and gauge freedom
Leveraging Lineage Barcodes as Natural Augmentations for Contrastive Learning of Cell Fate in scRNA-seq Data
Shizhao Joshua Yang (Dartmouth College), Kevin Z. Lin (University of Washington)
Explainability and InterpretabilityRepresentation LearningTransformerAuto EncoderContrastive LearningBiomedical Data
🎯 What it does: Leverage single-cell lineage barcodes as natural data augmentation to train a semi-supervised contrastive learning model, LCL, to extract gene expression signals that determine cell fate;
Leveraging Machine Unlearning for Cost-Efficient Preference Alignment
XiaoHua Feng, Chaochao Chen (Zhejiang University)
Computational EfficiencyData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: This paper proposes a U2A framework that combines Machine Unlearning with Preference Alignment, utilizing a two-layer optimization to precisely select and weight negative samples for unlearning, thereby significantly improving the preference alignment of LLMs without requiring a large number of positive samples.
LFQ: Logit-aware Final-block Quantization for Boosting the Generation Quality of Low-Bit Quantized LLMs
Jung Hyun Lee (KAIST), Eunho Yang (KAIST)
GenerationOptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: An improved method for post-training quantization (PTQ) of large language models (LLMs) after low-bit weight quantization is proposed, called Logit-aware Final-block Quantization (LFQ). This method better preserves the token probability distribution during generation by fine-tuning scalars using cross-entropy loss on the final Transformer block.
Lie-Algebraic Acceleration of Neural Koopman Dynamics
Jongwon Lee (Korea University), Sungwoo Park (Korea University)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningAuto EncoderGenerative Adversarial NetworkContrastive LearningTabularTime SeriesBiomedical DataPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A Koopman dynamics framework combining Lie algebra structure with neural Magnus expansion was constructed, and prefix-scan parallel acceleration on the time axis was implemented; by restricting the generator to a finite-dimensional Lie subalgebra, numerical operations maintain structural consistency and improve interpretability.
LieStoNet: Learning Lie Symmetries from Spatiotemporal Data for Stochastic Dynamical Systems
Shida Liu (Harvard University), L Mahadevan
OptimizationExplainability and InterpretabilityTransformerScore-based ModelAuto EncoderContrastive LearningTabularTime SeriesSequentialFinance RelatedPhysics RelatedStochastic Differential Equation
🎯 What it does: Proposed and implemented LieStoNet, an end-to-end machine learning framework for automatically discovering Lie-point continuous symmetries of stochastic differential equations (SDEs) from spatiotemporal trajectory data.
LieWarper: Geometry-Aware Motion Transfer via Lie Algebra
Linsong Shan, Yixuan Geng
Image TranslationGenerationPose EstimationDepth EstimationOptimizationTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingOptical FlowImageVideo
🎯 What it does: Propose a geometry-aware motion transfer framework called LieWarper based on the Sim(2) Lie group, which can precisely transfer the motion trajectory of the reference video to the target video while maintaining global camera motion consistency.
LIF Recurrent Memory Enables Long-Horizon Spiking Computation
Fenghao Liu (Zhejiang University), Gang Pan (Zhejiang University)
Computational EfficiencyRepresentation LearningRecurrent Neural NetworkSpiking Neural NetworkTime SeriesSequential
🎯 What it does: Proposed a Local Recurrent Memory Module (LRMM) based on standard LIF neurons, which constructs local memory circuits with a small number of neurons to achieve long-sequence modeling and gradient stability;
LIFT: A Novel Framework for Enhancing Long-Context Understanding of LLMs via Long Input Fine-Tuning
Yansheng Mao (Peking University), Muhan Zhang (Peking University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the LIFT framework, which enables short-context LLMs to answer questions without full input in long-context tasks by generating synthetic QA and parameter fine-tuning during testing on long texts.
Lifting Traces to Logic: Programmatic Skill Induction with Neuro-Symbolic Learning for Long-Horizon Agentic Tasks
Jie-Jing Shao (Nanjing University), Yu-Feng Li (Nanjing University)
Autonomous DrivingOptimizationExplainability and InterpretabilityRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIWorld ModelTextSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Neuro-Symbolic Skill Induction (NSI) framework, which elevates interaction trajectories into executable, logic-driven programs. It separates perception from symbolic execution using neuro-symbolic methods and automatically corrects and expands skills through reflective planning at runtime.
LiftQuant: Continuous Bit-Width LLM via Dimensional Lifting and Projection
Liulu He (Nanjing University), Li Du (Nanjing University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Propose the LiftQuant framework to achieve continuous bit-width control for LLM weight quantization.
Light Forcing: Accelerating Autoregressive Video Diffusion via Sparse Attention
Chengtao Lv (Nanyang Technological University), Wenya Wang (Nanyang Technological University)
GenerationTransformerDiffusion modelVideoBenchmark
🎯 What it does: Propose LIGHT FORCING for autoregressive video diffusion models, achieving improved generation efficiency through sparse attention while maintaining or even enhancing visual quality.
Light Up Your Face: A Physically Consistent Dataset and Diffusion Model for Face Fill-Light Enhancement
Jue Gong (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
Image TranslationRestorationGenerationData SynthesisTransformerPrompt EngineeringDiffusion modelScore-based ModelImage
🎯 What it does: This paper proposes a Face Fill Lighting Enhancement (FFE) method that improves facial brightness and shadow quality using controllable virtual fill light sources while keeping the original background illumination unchanged.
LightAVSeg: Lightweight Audio-Visual Segmentation
Qing Zhong (Huazhong Agricultural University), Angela Yao (National University of Singapore)
SegmentationRecurrent Neural NetworkTransformerVision Language ModelAuto EncoderContrastive LearningVideoMultimodalityAudio
🎯 What it does: Propose LightAVSeg, a lightweight audio-visual segmentation framework that can efficiently and in real-time perform pixel-level localization of sound-emitting objects in videos on mobile devices.
LightningRL: Breaking the Accuracy–Parallelism Trade-off of Block-wise dLLMs via Reinforcement Learning
Yanzhe Hu (Shanghai Jiao Tong University), Zhijie Deng (Shanghai Jiao Tong University)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: Post-training a pre-trained block-wise diffusion language model (dLLM) using reinforcement learning, aiming to simultaneously improve parallel generation speed and generation quality.
Lightweight and Interpretable Transformer via Unrolling of Mixed Graph Algorithms for Traffic Forecast
Ji Qi (Tsinghua University), H. Vicky Zhao (Tsinghua University)
OptimizationExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTransformerGraphTime Series
🎯 What it does: This paper constructs a lightweight and interpretable Transformer-like network, achieving traffic flow prediction by expanding the hybrid graph algorithm of spatial undirected graphs and temporal directed graphs into ADMM neural layers.
Lightweight Federated Incremental Learning via Decoupled Replay
Xiuying Wang (Beijing University of Posts and Telecommunications), Imran Razzak (Mohamed bin Zayed University of Artificial Intelligence)
Federated LearningSafty and PrivacyComputational EfficiencyRepresentation LearningSupervised Fine-TuningMixture of ExpertsDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: Propose a lightweight federated incremental learning framework, Li-FIL, which uses the server to generate dense features and performs decoupled replay on the client side to alleviate catastrophic forgetting.
LightWM: Training-Free Hierarchical Working Memory for Small Language Model Agents
Ziyi Wang (Beijing University of Posts and Telecommunications), Lanshan Zhang (Beijing University of Posts and Telecommunications)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringWorld ModelTextSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes LightWM, a training-agnostic hierarchical working memory framework that provides goal-conditioned global and local memory for small language models in long-sequence tasks, while maintaining information consistency through rule-constrained patch updates.
Likelihood Matching for Diffusion Models
Lei Qian (Peking University), Song Xi Chen (Tsinghua University)
GenerationData SynthesisDiffusion modelScore-based ModelContrastive LearningGaussian SplattingImageTabularTime SeriesStochastic Differential Equation
🎯 What it does: This paper proposes a direct maximum likelihood training framework based on quasi-maximum likelihood estimation (QMLE) called Likelihood Matching (LM). It approximates the inverse transition density by matching the score function with the Hessian (covariance) during the reverse process of diffusion models, thereby achieving direct optimization of the likelihood of the original data distribution.
Likelihood over Estimation: Robust Quadratic Discriminant Analysis for Heavy-Tailed Distributions with Theory and Evidence
Niranjana Ambadi (LawMate Private Limited), eugene pinsky
ClassificationTabularTime SeriesSequentialBenchmarkAgriculture RelatedFinance RelatedPhysics Related
🎯 What it does: Proposes Stable‑QDA, a quadratic discriminant analysis method that replaces the Gaussian distribution with symmetric α‑stable distribution in high-dimensional data.
LILO: Bayesian Optimization with Natural Language Feedback
Kasia Kobalczyk, Eytan Bakshy (Meta)
OptimizationReinforcement Learning from Human FeedbackNeural Architecture SearchTransformerLarge Language ModelPrompt EngineeringImageTextTabular
🎯 What it does: Proposes a framework called LILO (Language-in-the-Loop Optimization) based on Bayesian optimization, which leverages large language models (LLMs) to convert free-text feedback from decision-makers into structured pairwise preferences, and performs Bayesian optimization using a Gaussian process (GP) surrogate.
LiME: Lightweight Mixture of Experts for Efficient Multimodal Multi-task Learning
Md Kowsher (University Of Central Florida), Chen Chen
Federated LearningComputational EfficiencyRepresentation LearningTransformerMixture of ExpertsVision Language ModelImageVideoTextMultimodality
🎯 What it does: LiME proposes a lightweight Mixture of Experts framework, achieving efficient parameterized fine-tuning for multi-modal multi-task learning by sharing PEFT modules and using lightweight expert vectors for modulation.
LimiX-2M: Mitigating Low-Rank Collapse and Attention Bottlenecks in Tabular Foundation Models
Yuanrui Wang (Tsinghua University), Peng Cui (Stable AI)
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerMixture of ExpertsContrastive LearningTabularBenchmark
🎯 What it does: Designed a sparse embedding layer based on RBF called RaBEL and a reordered bidirectional attention structure, constructing the LimiX-2M model with 2M parameters.
LIMMT: Less Is More for Motion Tracking
Yu Guan (Tsinghua University), Li Yi (Tsinghua University)
Object TrackingData-Centric LearningRobotic IntelligenceTransformerReinforcement LearningAuto EncoderContrastive LearningOptical FlowPoint CloudSequential
🎯 What it does: Proposed a motion tracking framework for data centers called LIMMT, which uses a three-stage General Quality Selection (GQS) to evaluate and filter a large amount of MoCap data based on physical feasibility, behavioral diversity, and action complexity, generating a high-quality training subset.
LIMSSR: LLM-Driven Sequence-to-Score Reasoning under Training-Time Incomplete Multimodal Observations
Huangbiao Xu (Fuzhou University), Yuxin Peng (Peking University)
RestorationRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsAuto EncoderContrastive LearningVideoTextMultimodality
🎯 What it does: Propose a framework called LIMSSR based on large language models to address learning tasks where multi-modal missing data exists during the training phase, converting missing multi-modal reasoning into a conditional sequence-to-score reasoning process.
LiMuon: Light and Fast Muon Optimizer for Large Models
Feihu Huang (Nanjing University of Aeronautics and Astronautics), Songcan Chen (Nanjing University of Aeronautics and Astronautics)
OptimizationComputational EfficiencyTransformerLarge Language ModelContrastive LearningImageText
🎯 What it does: Proposed a lightweight and fast Muon optimizer called LiMuon, aiming to efficiently train large-scale models while balancing low memory usage and low sample complexity.
LineageFlow: Flow Matching for High-Fidelity Family-Aware Protein Sequence Generation
Langzhang Liang (Fudan University), Zenglin Xu (Fudan University)
GenerationProtein Structure PredictionTransformerDiffusion modelScore-based ModelFlow-based ModelBiomedical Data
🎯 What it does: A discrete flow matching model based on lineage priors, called LineageFlow, was constructed. It is initialized with a Dirichlet distribution reconstructed from ancestral sequences and incorporates a single 'mutation-selection-amplification' (rerouting) operation during the flow process to achieve target guidance.
Linear Bandits beyond Inner Product Spaces, the case of Bandit Optimal Transport
Lorenzo Croissant (CREST)
OptimizationReinforcement Learning
🎯 What it does: This paper studies the online optimal transport problem, demonstrating that online linear bandit learning is also feasible in non-Hilbert spaces (i.e., non-inner product spaces), and presents an algorithm called EntUCB, which can achieve sublinear cumulative regret under this setting;
Linear Causal Representation Learning by Topological Ordering, Pruning, and Disentanglement
Hao Chen (Shanghai Jiao Tong University), Yu Guang Wang (Shanghai Jiao Tong University)
Representation LearningAuto EncoderContrastive LearningTextTabular
🎯 What it does: Proposes a new linear causal representation learning algorithm, CREATOR, which utilizes topological sorting, pruning, and disentanglement to recover latent causal features and graph structures in heterogeneous environments.
Linear Ensembles Wash Away Watermarks: On the Fragility of Distributional Perturbations in LLMs
Zhihao Wu (King's College London), Runcong Zhao (King's College London)
Federated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Through theoretical proof and large-scale experiments, it is shown that linear integration (averaging the output distributions of independent watermarks) can almost completely eliminate distributed watermark perturbations, and the WASH framework is proposed to achieve efficient integration across models and tokenizers.
Linear Regression with Unknown Truncation Beyond Gaussian Features
Alexandros Kouridakis (University of Texas at Austin), Constantine Caramanis (University of Texas at Austin)
OptimizationComputational EfficiencyRepresentation LearningContrastive LearningTabular
🎯 What it does: Proposes an algorithm that can efficiently estimate the linear regression parameter w⋆ in a multi-dimensional feature space under unknown truncation sets (composed of at most k intervals).
Linear-Core Surrogates: Smooth Loss Functions with Linear Rates for Classification and Structured Prediction
Mehryar Mohri (Google Research), Yutao Zhong (Google Research)
ClassificationSupervised Fine-TuningContrastive LearningImageTextSequential
🎯 What it does: Designed and theoretically proved a class of smooth convex losses named Linear-Core Surrogates, which can achieve linear H-consistency in binary classification, multi-class classification, and structured prediction tasks, and provided the corresponding unbiased stochastic gradient estimation methods.
Linearizing Vision Transformer with Test-Time Training
Yining Li (Tsinghua University), Gao Huang (Tsinghua University)
ClassificationGenerationTransformerSupervised Fine-TuningDiffusion modelImage
🎯 What it does: By introducing a Test-Time Training (TTT) framework on a pre-trained Softmax Transformer, linear complexity attention is achieved, and fast fine-tuning is completed while preserving the original weights.
Linguistic Nepotism: Trading-off Quality for Language Preference in Multilingual RAG
Dayeon Ki (University of Maryland), Kevin Duh (Johns Hopkins University)
RetrievalExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes a controlled method to measure language preference in multilingual retrieval-augmented generation (mRAG) using internal model metrics. The method constructs parallel multilingual documents, ensures the authenticity of citations through a two-stage filtering process, and evaluates the model's citation accuracy across different languages using next-word prediction, further revealing the impact of language and position on citation behavior.
Linguistic Properties and Model Scale in Brain Encoding: From Small to Compressed Language Models
SUBBA REDDY OOTA, Bapi Raju Surampudi (IIIT Hyderabad)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelContrastive LearningTextBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Systematically compared the predictive ability of different-scale (1B–14B) Transformer models and their compressed versions (quantization and pruning) on natural language fMRI recordings, evaluating their linguistic capabilities using FlashHolmes.
Linguistic Relative Policy Optimization for Video Anomaly Reasoning
Jiaxu Leng (Chongqing University of Posts and Telecommunications), Xinbo Gao (Chongqing University of Posts and Telecommunications)
Anomaly DetectionOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes a language-based relative policy optimization (LRPO) framework, which utilizes pre-trained multi-modal large language models for parameter-free fine-tuning in video anomaly detection, and guides the inference process through editable language-based anomaly experiences.
Lions and Muons: Optimization via Stochastic Frank-Wolfe under Heavy-Tailed Noise
Maria-Eleni Sfyraki (University of California San Diego), Jun-Kun Wang (University of California San Diego)
OptimizationText
🎯 What it does: This paper proposes a unified framework that views two modern deep learning optimizers, Lion and Muon, as Stochastic Frank-Wolfe (FW) methods, and based on this, designs FW variants for light-tailed and heavy-tailed noise, providing theoretical convergence analysis.
LipoPU: Pocket-level Prediction of Lipid-Protein Interactions via Positive-Unlabeled Learning
Yuxing Wang (Westlake University), Jing Huang (Westlake University)
Drug DiscoveryTransformerSupervised Fine-TuningContrastive LearningBiomedical Data
🎯 What it does: Propose a PU learning framework called LipoPU based on attention-based multi-instance learning, used for predicting the lipid binding capacity and category of protein binding pockets;
Listening Through the Noise: Cauchy-Driven Diffusion Bridges for Robust Gastrointestinal Auscultation and Clinical Benchmarking
Dian Ding (Shanghai Jiao Tong University), Guangtao Xue (Shanghai Jiao Tong University)
RestorationAnomaly DetectionTransformerDiffusion modelScore-based ModelBiomedical DataBenchmarkStochastic Differential EquationAudio
🎯 What it does: A framework for denoising and diagnosing gastrointestinal sounds based on a Cauchy-driven diffusion bridge is studied, and a large-scale clinical gastrointestinal sound dataset, CLINBS, is constructed.
LiteVSR: Lightweight Adaptation of Frozen Diffusion Transformers for Video Super-Resolution
Yu Cao (Queen Mary University Of London), Jifei Song (Huawei Darwin Research Center)
Super ResolutionTransformerDiffusion modelScore-based ModelFlow-based ModelVideo
🎯 What it does: LiteVSR achieves super-resolution reconstruction of low-quality videos by building a lightweight State-Aware Adapter on a frozen Diffusion Transformer (DiT);
LithoDreamer: A Physics-Informed World Model for Multi-Stage Computational Lithography
Yuqi Jiang (Zhejiang University), Cheng Zhuo (Zhejiang University)
OptimizationTransformerDiffusion modelAuto EncoderContrastive LearningWorld ModelImageTabularPhysics Related
🎯 What it does: Propose LithoDreamer, a physics-informed world model that continuously simulates the four-stage lithography process of Layout-Mask-Resist Image-ADI, and achieves forward evolution and reverse planning.