ICLR 2025 Papers — Page 19
International Conference on Learning Representations · 3704 papers
Learning from weak labelers as constraints
Vishwajeet Agrawal (Carnegie Mellon University), Pradeep Kumar Ravikumar (Carnegie Mellon University)
ClassificationOptimizationTransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes an unsupervised learning framework directly based on the upper bound constraint of weak labeler error, using an alternating minimization method to learn classifiers on unlabeled data, avoiding the assumptions of traditional implicit generative models.
Learning Gain Map for Inverse Tone Mapping
Yinuo Liao (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
Image TranslationRestorationConvolutional Neural NetworkImageVideo
🎯 What it does: This paper proposes the Gain Map-based Inverse Tone Mapping (GM-ITM) task and designs a dual-branch network GMNet to learn the Gain Map corresponding to SDR images, achieving more efficient HDR up-conversion.
Learning General-purpose Biomedical Volume Representations using Randomized Synthesis
Neel Dey (Massachusetts Institute of Technology), Polina Golland (Massachusetts Institute of Technology)
SegmentationData SynthesisDomain AdaptationConvolutional Neural NetworkContrastive LearningImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A random synthesis engine based on medical shape templates was constructed, and a 3D UNet was pre-trained using multi-label contrastive learning, demonstrating good cross-domain generalization in multimodal registration and few-shot segmentation tasks.
Learning Generalizable Skills from Offline Multi-Task Data for Multi-Agent Cooperation
Sicong Liu (East China Normal University), Bin Yang (East China Normal University)
TransformerReinforcement LearningContrastive LearningSequential
🎯 What it does: Proposes the HiSSD framework, which combines common and task-specific skills in collaborative learning to enhance offline multi-task multi-agent cooperation strategy transfer.
Learning Geometric Reasoning Networks For Robot Task And Motion Planning
Smail Ait Bouhsain (National Center for Scientific Research), Thierry Simeon (National Center for Scientific Research)
Robotic IntelligenceGraph Neural NetworkGraph
🎯 What it does: A geometric reasoning network (GRN) based on graph neural networks is proposed to quickly predict the feasibility of robot operations (pick/put) and grasping types in a 3D environment, and to provide reasons for infeasibility.
Learning Graph Invariance by Harnessing Spuriosity
Tianjun Yao (Mohamed bin Zayed University of Artificial Intelligence), Zhiqiang Shen (Mohamed bin Zayed University of Artificial Intelligence)
Domain AdaptationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes the LIRS framework, which achieves graph invariant feature learning by first learning the outlier (spuriosity) characteristics in the graph and removing them from the features learned through ERM, significantly improving OOD generalization performance.
Learning Graph Quantized Tokenizers
Limei Wang (Meta AI), Bo Long (Meta AI)
Graph Neural NetworkTransformerContrastive LearningGraphBenchmarkPhysics Related
🎯 What it does: A graph quantization tokenizer (GQT) based on multi-task self-supervised learning and residual vector quantization is proposed, which maps graph nodes to discrete and compressed tokens.
Learning Harmonized Representations for Speculative Sampling
Lefan Zhang (Xiaohongshu Inc), Ruiwen Xu (Xiaohongshu Inc)
GenerationComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: The HASS (Harmonized Speculative Sampling) scheme is proposed, which achieves a higher acceptance rate by aligning the target distribution with the context during training and decoding, significantly accelerating the inference of LLMs;
Learning Hierarchical Polynomials of Multiple Nonlinear Features
Hengyu Fu (Peking University), Jason D. Lee (Princeton University)
🎯 What it does: This study explores how a three-layer neural network learns hierarchical polynomials formed by multiple nonlinear features, proposing a new analytical framework that proves a three-layer neural network trained through layer-wise gradient descent can effectively learn the target function.
Learning High-Degree Parities: The Crucial Role of the Initialization
Emmanuel Abbe (Ecole Polytechnique Federale de Lausanne), Donald Kougang-Yombi (African Institute for Mathematical Sciences)
OptimizationSupervised Fine-Tuning
🎯 What it does: This paper studies the learning of high-order parity functions, particularly the impact of initialization on learning effectiveness. By analyzing the performance of gradient descent on standard neural networks, it is found that the choice of initialization is crucial for learning almost complete parity functions.
Learning How Hard to Think: Input-Adaptive Allocation of LM Computation
Mehul Damani (Massachusetts Institute of Technology), Jacob Andreas (Massachusetts Institute of Technology)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A method is proposed to adaptively allocate computational resources based on input difficulty during language model inference, learning an input difficulty predictor to dynamically decide the number of samples or select the decoder accordingly.
Learning Interleaved Image-Text Comprehension in Vision-Language Large Models
Chenyu Zhou (Xiamen University), Rongrong Ji (Xiamen University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed the Interleaved Image-Text Understanding Task (IITC) and its corresponding VEGA dataset, and conducted post-training and evaluation on a multimodal large language model.
Learning Interpretable Hierarchical Dynamical Systems Models from Time Series Data
Manuel Brenner (Central Institute of Mental Health), Daniel Durstewitz (Central Institute of Mental Health)
Anomaly DetectionExplainability and InterpretabilityRepresentation LearningRecurrent Neural NetworkReinforcement LearningTime SeriesBiomedical DataMagnetic Resonance ImagingElectrocardiogram
🎯 What it does: This paper proposes a hierarchical dynamic system reconstruction (DSR) framework that utilizes multi-domain time series to learn individual-specific interpretable low-dimensional feature vectors and maps them to generative recursive dynamic models at the population level, thereby achieving unsupervised reconstruction, transfer learning, and few-shot learning of different dynamic laws.
Learning LLM-as-a-Judge for Preference Alignment
Ziyi Ye (Tsinghua University), Yiqun LIU
Recommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Using a pre-trained LLM to generate comparative judgments through self-sampling, training on preference data to obtain a decision model, Con-J, which can provide both binary preference judgments and explanations.
Learning local equivariant representations for quantum operators
Zhanghao Zhouyin (University of Science and Technology of China), Qiangqiang Gu (University of Science and Technology of China)
OptimizationComputational EfficiencyRepresentation LearningDrug DiscoveryGraph Neural NetworkGraphPhysics Related
🎯 What it does: A strict local equivariant message passing network (SLEM) is proposed for high-precision prediction of quantum operators (Hamiltonian, density matrix, overlap matrix).
Learning Long Range Dependencies on Graphs via Random Walks
Dexiong Chen (Max Planck Institute of Biochemistry), Karsten Borgwardt (Max Planck Institute of Biochemistry)
Graph Neural NetworkTransformerGraphBenchmark
🎯 What it does: This paper proposes a new architecture called NeuralWalker, which combines random walks with information propagation in graph neural networks. It uses random walk sequences to capture long-range dependencies, encodes them through a sequence model, and then performs local or global information propagation to generate graph and node representations.
Learning Mask Invariant Mutual Information for Masked Image Modeling
Tao Huang (University of Sydney), Chang Xu (University of Sydney)
Object DetectionSegmentationTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Proposed MI-MAE based on information bottleneck theory, improving Masked Autoencoders by maximizing mutual information and minimizing loss to enhance the quality of latent features, further improving pre-training effectiveness.
Learning mirror maps in policy mirror descent
Carlo Alfano (University of Oxford), Patrick Rebeschini (University of Oxford)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper conducts an empirical study on the Policy Mirror Descent (PMD) framework, learning and optimizing mirror mappings to enhance algorithm performance.
Learning Molecular Representation in a Cell
Gang Liu (University of Notre Dame), Shantanu Singh (Broad Institute of MIT and Harvard)
Representation LearningDrug DiscoveryGraph Neural NetworkMultimodalityGraphBiomedical Data
🎯 What it does: The InfoAlign method is proposed, which learns molecular representations on a cell context graph through an information bottleneck, allowing molecular embeddings to fully decode multimodal features such as cell morphology and gene expression while maintaining minimal redundant information.
Learning Multi-Index Models with Neural Networks via Mean-Field Langevin Dynamics
Alireza Mousavi-Hosseini (University of Toronto), Murat A Erdogdu
Stochastic Differential Equation
🎯 What it does: This study investigates the training of a two-layer neural network using Mean Field Langevin Dynamics (MFLD) to learn multi-index models, providing sample and computational complexity theory regarding effective dimension.
Learning Neural Networks with Distribution Shift: Efficiently Certifiable Guarantees
Gautam Chandrasekaran (University of Texas at Austin), Konstantinos Stavropoulos (University of Texas at Austin)
Domain AdaptationOptimization
🎯 What it does: This paper proposes the first effective testable learning (TDS) algorithm for regression problems in a distribution shift environment, which can provide low test error neural networks or rejection learning when the training distribution is known and the test distribution is unknown.
Learning on One Mode: Addressing Multi-modality in Offline Reinforcement Learning
Mianchu Wang (University of Warwick), Giovanni Montana (Alan Turing Institute)
Reinforcement LearningMultimodality
🎯 What it does: This paper proposes an offline reinforcement learning method called LOM, which models the behavior policy as a Gaussian mixture model and selects the single mode with the highest reward for weighted imitation learning, thereby achieving better policy learning on multimodal datasets.
Learning Partial Graph Matching via Optimal Partial Transport
Gathika Ratnayaka (Australian National University), Qing Wang (Australian National University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: Proposes an optimal partial transport-based partial graph matching optimization framework that automatically decides which nodes to match and provides the optimal partial mapping.
Learning Randomized Algorithms with Transformers
Johannes von Oswald (Google), Angelika Steger (ETH Zürich)
Transformer
🎯 What it does: This paper studies how to enable Transformers to learn randomized algorithms to enhance robustness and average performance in tasks such as associative recall, graph coloring, and grid exploration.
Learning Robust Representations with Long-Term Information for Generalization in Visual Reinforcement Learning
Rui Yang (University of Science and Technology of China), Bin Li (University of Science and Technology of China)
Representation LearningReinforcement LearningImage
🎯 What it does: A method for learning robust action value representations (ROUSER) under the information bottleneck framework is proposed, which utilizes first-order reward decomposition of long-range information to enhance the generalization ability of visual reinforcement learning.
Learning Shape-Independent Transformation via Spherical Representations for Category-Level Object Pose Estimation
Huan Ren (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
Pose EstimationTransformerPoint Cloud
🎯 What it does: Proposes the SpherePose method, which utilizes spherical representation to learn shape-agnostic transformations for category-level object pose estimation.
Learning Spatial-Semantic Features for Robust Video Object Segmentation
Xin Li (Harbin Institute of Technology), Ming-Hsuan Yang
SegmentationTransformerVideo
🎯 What it does: A robust video object segmentation framework (S3) based on spatial-semantic feature learning and discriminative query propagation is proposed.
Learning Spatiotemporal Dynamical Systems from Point Process Observations
Valerii Iakovlev (Aalto University), Harri Lähdesmäki (Aalto University)
GenerationData SynthesisComputational EfficiencyTransformerTime SeriesSequentialPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes a generative model that can learn the spatiotemporal dynamics of systems from randomly sampled observation point processes.
Learning Splitting Heuristics in Divide-and-Conquer SAT Solvers with Reinforcement Learning
Shumao Zhai (Beihang University), Ning Ge (State Key Laboratory of Complex and Critical Software Environment)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A segmentation heuristic optimization method based on reinforcement learning and graph neural networks (RDC-SAT) is proposed and implemented to improve the splitting decisions of the Divide-and-Conquer SAT solver.
Learning stochastic dynamics from snapshots through regularized unbalanced optimal transport
Zhenyi Zhang (Peking University), Peijie Zhou (Peking University)
OptimizationTime SeriesBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a Regularized Unbalanced Optimal Transport (DeepRUOT) method based on deep learning, which can learn continuous non-equilibrium stochastic dynamics from sparse temporal snapshot data.
Learning Structured Representations by Embedding Class Hierarchy with Fast Optimal Transport
Siqi Zeng (University of Illinois), Han Zhao (University of Illinois)
ClassificationRetrievalRepresentation LearningFlow-based ModelImage
🎯 What it does: This paper builds upon the existing CPCC regularization framework by introducing Optimal Transport (OT) distance to measure the similarity between class distributions, thereby achieving more fine-grained and accurate hierarchical embeddings.
Learning Structured Universe Graph with Outlier OOD Detection for Partial Matching
Zetian Jiang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Object DetectionAnomaly DetectionConvolutional Neural NetworkGraph Neural NetworkImageGraph
🎯 What it does: A structured universe graph matching method is proposed, which learns latent universe graphs and combines energy-based OOD detection to address issues of partial matching, occlusion, and annotation errors in graph matching.
Learning Successor Features with Distributed Hebbian Temporal Memory
Evgenii Aleksandrovich Dzhivelikian (Moscow Institute of Physics and Technology), Aleksandr Panov (Moscow Institute of Physics and Technology)
Recurrent Neural NetworkReinforcement LearningWorld ModelSequential
🎯 What it does: A distributed Hebbian temporal memory (DHTM) based on factor graphs and distributed Hebbian learning is proposed for online sequence learning and reinforcement learning based on successor features.
Learning system dynamics without forgetting
Xikun ZHANG, Dacheng Tao (Nanyang Technological University)
Graph Neural NetworkTime SeriesBiomedical DataBenchmarkPhysics RelatedOrdinary Differential Equation
🎯 What it does: This study investigates the Continuous Dynamic Learning (CDL) task and proposes the MS-GODE model, which can continuously learn across multiple systems without forgetting, and establishes the biological cell system benchmark Bio-CDL.
Learning Task Belief Similarity with Latent Dynamics for Meta-Reinforcement Learning
Menglong Zhang (Southern University of Science and Technology), Quanying Liu (Southern University of Science and Technology)
Robotic IntelligenceMeta LearningReinforcement LearningSequential
🎯 What it does: The SimBelief framework is proposed, which achieves rapid adaptation in sparse reward environments and efficient transfer to out-of-distribution (OOD) tasks through learning latent task belief similarity in Bayes-Adaptive MDP.
Learning the Complexity of Weakly Noisy Quantum States
Yusen Wu (Beijing Normal University), Jingbo Wang
Physics Related
🎯 What it does: Using classical shadows and quantum learning algorithms to predict the finite structure complexity (LS-complexity) of weakly noisy quantum states.
Learning the Optimal Stopping for Early Classification within Finite Horizons via Sequential Probability Ratio Test
Akinori F. Ebihara (NEC Corporation), Hitoshi Imaoka (NEC Corporation)
ClassificationOptimizationComputational EfficiencyRecurrent Neural NetworkTransformerSupervised Fine-TuningVideoTabular
🎯 What it does: A SPRT-based early classification framework called FIRMBOUND is proposed, which can efficiently solve the optimal stopping boundary within a finite time domain without the need for expensive backward induction calculations.
Learning to Adapt Frozen CLIP for Few-Shot Test-Time Domain Adaptation
Zhixiang Chi (University of Toronto), Konstantinos N Plataniotis (University of Toronto)
Domain AdaptationTransformerContrastive LearningImage
🎯 What it does: This paper proposes a method (L2C) for achieving few-shot temporal adaptation during testing with a small number of unlabeled samples by learning input space features in parallel outside the CLIP framework.
Learning to Clarify: Multi-turn Conversations with Action-Based Contrastive Self-Training
Maximillian Chen (Columbia University), Sercan O Arik
Recommendation SystemOptimizationTransformerLarge Language ModelContrastive LearningTextTabular
🎯 What it does: This paper proposes Action-Based Contrastive Self-Training (ACT), a method for preference optimization and action planning in multi-turn dialogue based on actions, enabling LLMs to better identify and handle ambiguities with limited data, thereby improving dialogue success rates.
Learning to Communicate Through Implicit Communication Channels
Han Wang (Chinese University of Hong Kong), Baoxiang Wang (Chinese University of Hong Kong)
Robotic IntelligenceReinforcement Learning
🎯 What it does: This paper proposes an Implicit Channel Protocol (ICP) framework that utilizes probing actions to construct implicit communication channels for multi-agent collaboration.
Learning to Contextualize Web Pages for Enhanced Decision Making by LLM Agents
Dongjun Lee (Korea Advanced Institute of Science and Technology), Kimin Lee (Korea Advanced Institute of Science and Technology)
Recommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: The LCoW framework is proposed, which utilizes a contextualization module to transform complex web pages into an understandable form, enhancing the decision-making quality of LLM agents in web automation tasks.
Learning to Discover Regulatory Elements for Gene Expression Prediction
Xingyu Su (Texas A&M University), Shuiwang Ji (Texas A&M University)
Biomedical Data
🎯 What it does: The Seq2Exp framework is proposed to predict gene expression by learning and extracting regulatory elements from DNA sequences and epigenetic signals.
Learning to Discretize Denoising Diffusion ODEs
Vinh Tong (University of Stuttgart), Mathias Niepert (University of Bern)
GenerationOptimizationComputational EfficiencyKnowledge DistillationDiffusion modelImageOrdinary Differential Equation
🎯 What it does: This study proposes a lightweight framework (LD3) for pre-trained diffusion models, which significantly reduces the number of neural network evaluations while maintaining generation quality by learning optimal time discretization strategies.
Learning to engineer protein flexibility
Petr Kouba (Czech Technical University), Josef Sivic (Czech Technical University)
Protein Structure PredictionTransformerLarge Language ModelSupervised Fine-TuningBiomedical Data
🎯 What it does: Using a pre-trained protein language model and limited MD data, we propose Flexpert-Seq and Flexpert-3D to predict local protein flexibility, and fine-tune ProteinMPNN based on Flexpert-3D (Flexpert-Design) to achieve control over the increase and decrease of protein sequence flexibility.
Learning to Explore and Exploit with GNNs for Unsupervised Combinatorial Optimization
Utku Umur ACIKALIN, Carla P Gomes
OptimizationGraph Neural NetworkGraph
🎯 What it does: An unsupervised graph neural network framework named X2GNN is proposed for simultaneously exploring and exploiting in combinatorial optimization problems (maximum clique, maximum independent set, maximum cut) to generate high-quality solutions.
Learning to Generate Diverse Pedestrian Movements from Web Videos with Noisy Labels
Zhizheng Liu (University of California), Bolei Zhou (University of California)
GenerationAutonomous DrivingTransformerDiffusion modelVideoPoint Cloud
🎯 What it does: A large-scale urban pedestrian dataset called CityWalkers is constructed, and a diffusion model-based method called PedGen is proposed to learn and generate diverse, context-aware pedestrian motion trajectories from online videos.
Learning to Help in Multi-Class Settings
Yu Wu (Rutgers University), Anand D. Sarwate (Rutgers University)
ClassificationOptimizationFederated LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: A 'Learning to Help' (L2H) framework for multi-class classification is proposed, enabling resource-constrained local devices to push hard-to-classify samples to the server via a rejector to enhance overall performance;
Learning to Plan Before Answering: Self-Teaching LLMs to Learn Abstract Plans for Problem Solving
Jin Zhang (Tsinghua University), Chongjie Zhang (Moonshot AI)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A new self-supervised training method called LEPA is proposed, which first allows a large language model to generate an abstract 'expected plan', and then answers questions based on the plan, continuously optimizing the plan and answers through self-reflection.
Learning to Search from Demonstration Sequences
Dixant Mittal (Moovita), Wee Sun Lee (National University of Singapore)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningWorld ModelSequential
🎯 What it does: This paper proposes the Differentiable Tree Search Network (D-TSN), a neural network architecture that learns to construct differentiable search trees from demonstration sequences, enabling the joint learning of submodules such as world models and value functions for end-to-end search and planning.
Learning to Select Nodes in Branch and Bound with Sufficient Tree Representation
Sijia Zhang (University of Science and Technology of China), Xiangyang Li
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A TRGNN node selection framework based on ternary graph representation and reinforcement learning is proposed to address the node selection problem in the branch-and-bound tree of mixed-integer linear programming (MILP).
Learning to Solve Differential Equation Constrained Optimization Problems
Vincenzo Di Vito Francesco, Ferdinando Fioretto (University Of Virginia)
OptimizationTabularTime SeriesFinance RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A dual-network-based proxy optimization framework DE-OP is proposed for near real-time solving of optimization problems constrained by differential equations.
Learning to Steer Markovian Agents under Model Uncertainty
Jiawei Huang (ETH Zurich), Niao He (ETH Zurich)
OptimizationRobotic IntelligenceReinforcement LearningAgentic AI
🎯 What it does: This paper addresses the 'Steering' problem of how to guide Markovian multi-agent systems to learn the desired policy by designing additional rewards in the presence of model uncertainty, and provides the corresponding theoretical and algorithmic framework.
Learning Transformer-based World Models with Contrastive Predictive Coding
Maxime Burchi (University of Wurzburg), Radu Timofte (University of Wurzburg)
TransformerReinforcement LearningContrastive LearningWorld ModelTime SeriesSequentialBenchmark
🎯 What it does: This paper proposes a Transformer-based world model called TWISTER, which uses action-conditioned contrastive predictive coding (AC-CPC) to learn high-quality temporal features, thereby enhancing the performance of model-based reinforcement learning.
Learning under Temporal Label Noise
Sujay Nagaraj (University of Toronto), Thomas Hartvigsen (University of Virginia)
ClassificationAnomaly DetectionRecurrent Neural NetworkTime SeriesBiomedical Data
🎯 What it does: This paper introduces the concept of Temporal Label Noise in time series classification and provides a formal definition.
Learning vector fields of differential equations on manifolds with geometrically constrained operator-valued kernels
Daning Huang (Pennsylvania State University), Yan Li (Pennsylvania State University)
Time SeriesOrdinary Differential Equation
🎯 What it does: Study the vector field of ordinary differential equations on manifolds and propose a geometric constraint operator-valued kernel regression method along with a numerical integrator that preserves the manifold.
Learning Video-Conditioned Policy on Unlabelled Data with Joint Embedding Predictive Transformer
Hao Luo (Peking University), Zongqing Lu (Beijing Academy of Artificial Intelligence)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerOptical FlowVideo
🎯 What it does: A Joint Embedding Prediction Transformer (JEPT) has been developed, capable of learning video conditional policies on mixed data that includes expert demonstrations and unlabeled expert videos, in order to reduce the demand for action labels.
Learning View-invariant World Models for Visual Robotic Manipulation
Jing-Cheng Pang (Nanjing University), Yang Yu (Nanjing University)
Robotic IntelligenceTransformerReinforcement LearningContrastive LearningWorld ModelImage
🎯 What it does: The ReViWo method is proposed, which utilizes multi-view data to learn a viewpoint-invariant world model, achieving robust control of visual robotic operations under viewpoint disturbances.
Learning-Augmented Frequent Directions
Anders Aamand (University of Copenhagen), Hao WU (University of Waterloo)
Recurrent Neural NetworkVideo
🎯 What it does: This paper proposes two learning-enhanced streaming algorithms, Misra-Gries and Frequent Directions, which guide memory allocation using predictors to achieve lower errors in frequency estimation and high-dimensional direction estimation tasks.
Learning-Augmented Search Data Structures
Chunkai Fu (Texas A and M University), Samson Zhou (Texas A and M University)
OptimizationComputational EfficiencyPoint CloudTabular
🎯 What it does: A search structure based on machine learning prediction, specifically skip lists and KD-trees, is proposed, utilizing predicted frequency to guide level elevation and dimension splitting, thereby achieving expected search times close to optimal.
Learning-Guided Rolling Horizon Optimization for Long-Horizon Flexible Job-Shop Scheduling
Sirui Li (Massachusetts Institute of Technology), Cathy Wu (Massachusetts Institute of Technology)
OptimizationAuto EncoderTabular
🎯 What it does: A learning-based rolling horizon optimization framework L-RHO is proposed to accelerate and improve the solution of long-slot combinatorial optimization problems, such as flexible job shop scheduling.
Leave-One-Out Stable Conformal Prediction
Kiljae Lee (Ohio State University), Yuan Zhang (Ohio State University)
Tabular
🎯 What it does: A leave-one-out stable conformal prediction (LOO-StabCP) method is proposed, which accelerates full conformal prediction with algorithm stability and achieves efficient inference for multiple prediction requests.
LeFusion: Controllable Pathology Synthesis via Lesion-Focused Diffusion Models
Hantao Zhang (Beihang University), Pascal Fua (Swiss Federal Institute of Technology Lausanne)
SegmentationGenerationData SynthesisDiffusion modelImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A lesion-focused diffusion model named LeFusion has been developed, capable of synthesizing images/annotations with lesions from normal medical images. It achieves fine control over lesion size, location, texture, and category through texture histogram control, multi-channel decomposition, and lesion mask diffusion.
Less is More: Masking Elements in Image Condition Features Avoids Content Leakages in Style Transfer Diffusion Models
Lin Zhu (Shanghai Jiao Tong University), Nanyang Ye
Image TranslationGenerationDiffusion modelImageBenchmark
🎯 What it does: In text-driven image generation, a training-free method is proposed to achieve content and style decoupling by masking the features of style reference images.
Let Me Grok for You: Accelerating Grokking via Embedding Transfer from a Weaker Model
Zhiwei Xu (University of Michigan), Wei Hu (University of Michigan)
OptimizationComputational EfficiencyTransformerTabular
🎯 What it does: The GrokTransfer method is proposed, which accelerates the grokking process of neural networks by transferring embeddings learned from weak models to larger target models. Theoretical proof is provided for the XOR task, and experimental validation is conducted on various algorithmic tasks (modular addition, modular multiplication, parity, Transformer).
Let SSMs be ConvNets: State-space Modeling with Optimal Tensor Contractions
Yan Ru Pei (Brainchip Inc.)
RecognitionOptimizationComputational EfficiencyConvolutional Neural NetworkAudio
🎯 What it does: This paper proposes the Centaurus network, which employs a deep state space model (SSM) block with a variable connection structure and optimizes training efficiency through tensor contraction.
Let the Code LLM Edit Itself When You Edit the Code
Zhenyu He (Peking University), Di He (Peking University)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: Proposes an efficient method for updating the KV cache of large language models in real-time code editing scenarios.
Let Your Features Tell The Differences: Understanding Graph Convolution By Feature Splitting
Yilun Zheng (Nanyang Technological University), Lihui Chen (Nanyang Technological University)
ClassificationOptimizationGraph Neural NetworkGraphBenchmark
🎯 What it does: This paper proposes a graph feature selection method called GFS, which uses a new metric TFI to distinguish between beneficial and detrimental feature dimensions for graph convolution, significantly improving the node classification performance of various GNNs.
LevAttention: Time, Space and Streaming Efficient Algorithm for Heavy Attentions
Ravindran Kannan (Simons Institute UC Berkeley), David Woodruff
Computational EfficiencyTransformerImage
🎯 What it does: This paper proposes a new attention mechanism called LevAttention, which constructs a 'universal key set' independent of context length by utilizing the Leverage Score of key vectors (or more generally, f-sensitivity). This allows each query to retrieve only this set to obtain all attention scores above a threshold, with a time complexity of O(n·poly(d/ε)) and can be implemented in streaming, parallel, or distributed environments.
Leveraging Driver Field-of-View for Multimodal Ego-Trajectory Prediction
M. Eren Akbiyik (ETH Zurich), Xi Wang (TU Munich)
Autonomous DrivingTransformerVideoMultimodality
🎯 What it does: A multi-modal autoregressive network called RouteFormer is proposed, which integrates the driver's perspective (head video + gaze points) with environmental scenes to predict the future trajectory of driving vehicles. It also introduces the Path Complexity Index (PCI) as a trajectory complexity metric and a new GEM dataset.
Leveraging Flatness to Improve Information-Theoretic Generalization Bounds for SGD
Ze Peng (Nanjing University), Yang Gao (Nanjing University)
OptimizationConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: This paper proposes the technique of 'omniscient trajectory' to derive an explicit mutual information theoretical generalization bound that utilizes the flatness of SGD, and based on this, obtains a lower bound of O(1/√n) for GD on the CLB problem;
Leveraging Sub-Optimal Data for Human-in-the-Loop Reinforcement Learning
Calarina Muslimani (University of Alberta), Matthew E. Taylor (University of Alberta)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: A Sub-optimal Data Pre-training (SDP) method was designed and validated, which uses pseudo-labels of sub-optimal data to pre-train the reward model in human-robot interaction reinforcement learning and initializes the replay buffer, significantly improving feedback efficiency.
Leveraging Submodule Linearity Enhances Task Arithmetic Performance in LLMs
Rui Dai (National Engineering Laboratory for Brain Inspired Intelligence Technology and Application, University of Science and Technology of China), Jieping Ye (Independent Researcher)
TransformerLarge Language ModelText
🎯 What it does: This paper finds through statistical analysis that, although the overall model lacks linearity, its sub-modules (layers, attention, MLP, etc.) exhibit high linearity. It then proposes a training-free task arithmetic model merging method: first, the model is split into sub-modules, and the closed-form optimal merging weights are derived using the linear characteristics of the sub-modules. Subsequently, the sub-modules are linearly merged to enhance the performance of multi-task models.
Leveraging Variable Sparsity to Refine Pareto Stationarity in Multi-Objective Optimization
Zeou Hu (University of Waterloo), Yaoliang Yu (University of Waterloo)
OptimizationFederated LearningImageTabular
🎯 What it does: This paper proposes the concept of Refined Pareto Stability (RPS) using a function-variable sparse structure and designs the RP-MGDA algorithm based on this to solve multi-objective optimization problems.
LICO: Large Language Models for In-Context Molecular Optimization
Tung Nguyen (University of California), Aditya Grover (University of California)
OptimizationDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringGraph
🎯 What it does: A general black-box optimization surrogate model LICO is proposed, which adds input and output embedding layers and a prediction layer on a pre-trained LLM, and achieves few-shot inference of molecular properties through task prompting.
LICORICE: Label-Efficient Concept-Based Interpretable Reinforcement Learning
Zhuorui Ye (Tsinghua University), Fei Fang (Carnegie Mellon University)
Explainability and InterpretabilityReinforcement LearningVision Language ModelImage
🎯 What it does: Under a limited budget for concept annotation, a new training framework called LICORICE is proposed, enabling reinforcement learning agents to learn interpretable concept bottleneck policies from a small amount of annotated data, achieving performance that is comparable to or even better than traditional baselines.
Lie Algebra Canonicalization: Equivariant Neural Operators under arbitrary Lie Groups
Zakhar Shumaylov (University of Cambridge), Carola-Bibiane Schönlieb (Harvard University)
OptimizationConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: A new energy-based normalization method (LieLAC) is proposed, which achieves the equivariance of any Lie group by minimizing the energy on the Lie group orbit, allowing existing pre-trained models (such as CNN, POSEIDON) to be transformed into equivariant/invariant networks corresponding to the symmetry group.
LIFe-GoM: Generalizable Human Rendering with Learned Iterative Feedback Over Multi-Resolution Gaussians-on-Mesh
Jing Wen (University of Illinois Urbana-Champaign), Shenlong Wang (University of Illinois Urbana-Champaign)
GenerationPose EstimationTransformerGaussian SplattingImageVideoMesh
🎯 What it does: This paper proposes a generalizable and animatable human avatar reconstruction and rendering framework called LIFE-GOM, which can generate a 3D representation in T-pose from sparse images in one go and supports high-quality rendering from arbitrary poses and viewpoints.
Lift Your Molecules: Molecular Graph Generation in Latent Euclidean Space
Mohamed Amine Ketata (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
GenerationData SynthesisOptimizationDrug DiscoveryGraph Neural NetworkDiffusion modelAuto EncoderGraph
🎯 What it does: This paper proposes the Synthetic Coordinate Embedding (SYCO) framework, which maps two-dimensional molecular graphs to three-dimensional Euclidean point clouds, and uses diffusion models to generate molecular graphs in this latent space, resulting in a new method for molecular graph generation.
LiFT: Learning to Fine-Tune via Bayesian Parameter Efficient Meta Fine-Tuning
Minyoung Kim (Samsung AI Center), Timothy Hospedales
Meta LearningSupervised Fine-TuningText
🎯 What it does: A parameter-efficient fine-tuning method based on a hierarchical Bayesian model, called LiFT, is proposed for cross-task learning on how to fine-tune pre-trained models.
Lightning-Fast Image Inversion and Editing for Text-to-Image Diffusion Models
Dvir Samuel (OriginAI), Rami Ben-Ari (OriginAI)
GenerationOptimizationComputational EfficiencyDiffusion modelImageStochastic Differential Equation
🎯 What it does: Designed and implemented an efficient image inversion and editing framework based on the guided Newton-Raphson method for text-to-image diffusion models;
Lightweight Neural App Control
Filippos Christianos (Huawei Noah's Ark Lab), Kun Shao (Huawei Noah's Ark Lab)
Computational EfficiencyRobotic IntelligenceTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: A lightweight multimodal mobile application control framework LiMAC has been developed, combining a lightweight Action Transformer with a fine-tuned Vision-Language model to achieve low-latency and low-memory interactions on mobile devices.
Lightweight Predictive 3D Gaussian Splats
Junli Cao (University of California), Jian Ren (Snap)
CompressionComputational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: A lightweight 3D Gaussian point rendering representation is designed, utilizing a parent-child tree structure and neural networks to significantly reduce storage while maintaining high-quality rendering.
Limits of Deep Learning: Sequence Modeling through the Lens of Complexity Theory
Nikola Zubic (University of Zurich), Davide Scaramuzza (University of Zurich)
TransformerChain-of-Thought
🎯 What it does: This study investigates and proves the inherent computational limitations of Structured State Space Models (SSM) and Transformers in tasks involving function composition, multi-step reasoning, and operations requiring combinations over large domains, supported by theoretical proofs and experimental validation of these limitations.
Limits to scalable evaluation at the frontier: LLM as judge won’t beat twice the data
Florian E. Dorner (Max Planck Institute for Intelligent Systems), Moritz Hardt (ETH Zurich)
Large Language ModelTextBenchmark
🎯 What it does: The study investigates the theoretical limits and empirical performance of using large language models as evaluators in model assessment, proving that they can achieve at most twice the sample efficiency when evaluating cutting-edge models.
Linear Combination of Saved Checkpoints Makes Consistency and Diffusion Models Better
Enshu Liu (Tsinghua University), Yu Wang (Tsinghua University)
GenerationOptimizationComputational EfficiencyDiffusion modelImage
🎯 What it does: This paper proposes a linear combination of checkpoints during the training of diffusion models and consistency models to improve generation quality and significantly reduce training costs.
Linear combinations of latents in generative models: subspaces and beyond
Erik Bodin (University of Cambridge), Henry Moss (Lancaster University)
GenerationData SynthesisOptimizationDiffusion modelImage
🎯 What it does: The Latent Optimal Linear combinations (LOL) method is proposed to construct linear combinations that satisfy the prior distribution in generative models, thereby achieving more reliable latent space interpolation and subspace definition.
Linear Mode Connectivity in Differentiable Tree Ensembles
Ryuichi Kanoh (National Institute of Informatics), Mahito Sugiyama (National Institute of Informatics)
TabularBenchmark
🎯 What it does: The research and implementation of the linear mode connectivity (LMC) of the soft tree ensemble model has been conducted, verifying that models with different random initializations can maintain performance through parameter interpolation.
Linear Multistep Solver Distillation for Fast Sampling of Diffusion Models
Yuchen Liang (Peking University), Yunhe Wang (Huawei)
GenerationComputational EfficiencyKnowledge DistillationDiffusion modelImageOrdinary Differential Equation
🎯 What it does: A linear multi-step solver distillation framework is proposed, allowing the student solver to approximate the teacher solver's sampling trajectory with very few function evaluations (NFE), thus achieving fast and high-quality diffusion model sampling.
Linear Partial Gromov-Wasserstein Embedding
Yikun Bai (Vanderbilt University), Soheil Kolouri (Vanderbilt University)
RetrievalOptimizationPoint Cloud
🎯 What it does: This paper proposes a linearized partial Gromov-Wasserstein (LPGW) embedding for fast computation of similarity between different metric spaces.
Linear Representations of Political Perspective Emerge in Large Language Models
Junsol Kim (University of Chicago), Aaron Schein (University of Chicago)
TransformerLarge Language ModelText
🎯 What it does: This study investigates whether there are linear representations of political positions within large language models and predicts and adjusts the model's political bias by probing attention heads.
Linear SCM Identification in the Presence of Confounders and Gaussian Noise
Vahideh Sanjaroonpouri, Pouria Ramazi (Brock University)
🎯 What it does: This paper studies the identifiability of linear structural causal models with Gaussian noise and common factors under observed data, providing sufficient conditions for finite identifiability and an explicit construction of equivalence classes.
Linear Spherical Sliced Optimal Transport: A Fast Metric for Comparing Spherical Data
Xinran Liu (Vanderbilt University), Soheil Kolouri (Vanderbilt University)
OptimizationComputational EfficiencyAuto EncoderPoint CloudBiomedical DataAlzheimer's Disease
🎯 What it does: A linear spherical slice optimal transport (LSSOT) framework is proposed and implemented for rapid comparison of spherical probability distributions, applied to brain surface registration and point cloud interpolation.
Linear Transformer Topological Masking with Graph Random Features
Isaac Reid (Google Research), Krzysztof Marcin Choromanski (University of Cambridge)
Computational EfficiencyRobotic IntelligenceTransformerImagePoint Cloud
🎯 What it does: A linear Transformer topology masking method based on Graph Random Features (GRF) is designed, capable of achieving O(N) time and space complexity attention masks under any graph structure.
Lines of Thought in Large Language Models
Raphaël Sarfati (Cornell University), Christopher Earls (Cornell University)
TransformerLarge Language ModelTextStochastic Differential Equation
🎯 What it does: By studying the internal token trajectories of large language models, it was found that they evolve along low-dimensional manifolds and can be approximated using a linear + noise model.
LiNeS: Post-training Layer Scaling Prevents Forgetting and Enhances Model Merging
Ke Wang (École Polytechnique Fédérale de Lausanne), Pascal Frossard (École Polytechnique Fédérale de Lausanne)
ClassificationOptimizationTransformerLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: A post-training hierarchical scaling method called LiNeS is proposed, which significantly restores the generalization ability of pre-trained models while maintaining the performance of fine-tuning tasks, and further enhances the effectiveness of multi-task model merging, single-task model merging, and different reward LLM strategy merging.
Lipschitz Bandits in Optimal Space
Xiaoyi Zhu (Fudan University), Zengfeng Huang (Fudan University)
🎯 What it does: The performance of Lipschitz bandits under limited memory is studied, and the Log-Li algorithm is proposed, which requires only O(log T) bits of memory to achieve optimal regret;
LiveBench: A Challenging, Contamination-Limited LLM Benchmark
Colin White (Abacus.AI), Micah Goldblum (Columbia)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A sustainable and pollution-resistant evaluation benchmark for LLMs, named LiveBench, has been constructed and released. It includes multiple categories (mathematics, coding, reasoning, language, instruction following, data analysis) tasks and achieves evaluation without LLM/human judgment through an automated, objective truth-based scoring system.
LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code
Naman Jain (University of California Berkeley), Ion Stoica (University of California Berkeley)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper presents LiveCodeBench, a real-time, pollution-free, comprehensive benchmark for evaluating the coding capabilities of large language models.
LiveXiv - A Multi-Modal live benchmark based on Arxiv papers content
Nimrod Shabtay (Tel Aviv University), Raja Giryes (Tel Aviv University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelTextMultimodalityBenchmark
🎯 What it does: Introducing LiveXiv, a real-time, sustainably growing multimodal benchmark that automatically extracts visual content such as charts and tables from arXiv scientific papers, generates VQA/TQA questions using GPT-4o, and ensures data quality through multi-layer automatic filtering.
LLaMA-Omni: Seamless Speech Interaction with Large Language Models
Qingkai Fang (Institute of Computing Technology Chinese Academy of Sciences), Yang Feng (Institute of Computing Technology Chinese Academy of Sciences)
TransformerLarge Language ModelSupervised Fine-TuningTextAudio
🎯 What it does: LLaMA-Omni is proposed, an end-to-end low-latency speech-to-text interaction model that can generate both text and speech responses simultaneously without relying on ASR.