NeurIPS 2024 Papers — Page 21
Conference on Neural Information Processing Systems · 4035 papers
Learning on Large Graphs using Intersecting Communities
Ben Finkelshtein (University of Oxford), Ron Levie (Technion Israel Institute of Technology)
ClassificationComputational EfficiencyGraph Neural NetworkGraphTime Series
🎯 What it does: This paper proposes a novel graph learning framework - Intercommunity Graph (ICG), which approximates any large graph as a linear combination of several intersecting cliques through a weak regularization lemma, thereby making the memory and time complexity of graph learning algorithms proportional only to the number of nodes.
Learning Optimal Lattice Vector Quantizers for End-to-end Neural Image Compression
Xi Zhang (Shanghai Jiao Tong University), Xiaolin Wu (Southwest Jiaotong University)
CompressionConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: Proposed a learnable optimal lattice vector quantization (OLVQ) and integrated it into an end-to-end neural image compression network to enhance bitrate-distortion performance.
Learning Optimal Tax Design in Nonatomic Congestion Games
Qiwen Cui (Paul G. Allen School of Computer Science Engineering University of Washington), Simon Shaolei Du
Optimization
🎯 What it does: This paper proposes an algorithm for learning optimal taxation through observing Nash equilibria without information feedback, applicable to non-atomic congestion games with countless players.
Learning Partitions from Context
Simon Buchholz (Max Planck Institute for Intelligent Systems)
🎯 What it does: This study investigates how to recover the partition structure of a set of tokens from the interaction information between the tokens, proposing an analytical framework that combines information theory and gradient descent dynamics.
Learning Place Cell Representations and Context-Dependent Remapping
Markus Pettersen (Simula Research Laboratory), Mikkel Elle Lepperød (Simula Research Laboratory)
Representation LearningRecurrent Neural NetworkSequential
🎯 What it does: A self-supervised objective function based on similarity is proposed and validated to learn joint spatial and contextual representations in neural networks, resulting in representations similar to hippocampal place cells, and its remapping behavior is explored.
Learning Plaintext-Ciphertext Cryptographic Problems via ANF-based SAT Instance Representation
Xinhao Zheng (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes a graph structure based on Algebraic Normal Form (ANF) and the corresponding CryptoANFNet neural network for efficient prediction of plaintext-ciphertext satisfiability in cryptography and key recovery.
Learning predictable and robust neural representations by straightening image sequences
Xueyan Niu (New York University), Eero P Simoncelli
Representation LearningAdversarial AttackConvolutional Neural NetworkContrastive LearningImageVideo
🎯 What it does: A self-supervised learning method aimed at 'straightening' has been designed and validated, training the network to produce representations with more linear temporal trajectories in image sequences, thereby achieving predictable and robust visual embeddings.
Learning Representations for Hierarchies with Minimal Support
Benjamin Rozonoyer (University of Massachusetts Amherst), Andrew McCallum (University of Massachusetts Amherst)
OptimizationRepresentation LearningGraph Neural NetworkGraphBiomedical Data
🎯 What it does: A framework is constructed to determine the unique distinguishing submatrix from known structural constraints (such as transitive closed DAGs) and is used for efficient training of node embeddings.
Learning rigid-body simulators over implicit shapes for large-scale scenes and vision
Yulia Rubanova (Google Deepmind), Tobias Pfaff (Google Deepmind)
Computational EfficiencyRobotic IntelligenceGraph Neural NetworkPoint CloudMesh
🎯 What it does: A rigid body simulator SDF-Sim based on graph networks has been developed, which uses learned implicit SDFs to represent object shapes instead of meshes, significantly reducing collision detection costs and supporting large-scale simulations of hundreds or even thousands of objects.
Learning Segmentation from Point Trajectories
Laurynas Karazija (Visual Geometry Group University of Oxford), Andrea Vedaldi (Visual Geometry Group University of Oxford)
Object TrackingSegmentationOptical FlowVideo
🎯 What it does: Under unsupervised conditions, an image segmentation network is trained using long-term point trajectories and optical flow information to achieve video object segmentation.
Learning Social Welfare Functions
Kanad Shrikar Pardeshi (Carnegie Mellon University), Aarti Singh (Carnegie Mellon University)
Tabular
🎯 What it does: This study investigates how to learn the implicit social welfare function of decision-makers from past decision data, focusing on the weighted power mean family. It provides theoretical results on its pseudo-dimension, VC dimension, and sample complexity in noisy scenarios, and proposes practical learning algorithms based on these theories.
Learning Spatially-Aware Language and Audio Embeddings
Bhavika Suresh Devnani (Georgia Institute of Technology), Miguel Sarabia (Apple)
Data SynthesisRetrievalConvolutional Neural NetworkContrastive LearningMultimodalityAudio
🎯 What it does: A multimodal model ELSA was trained, capable of mapping spatial audio and natural language descriptions to a shared embedding space, while supporting semantic retrieval and 3D sound source localization, and is compatible with non-spatial audio.
Learning Structure-Aware Representations of Dependent Types
Konstantinos Kogkalidis (Aalto University), Jean-Philippe Bernardy (University of Gothenburg)
Representation LearningTransformerContrastive LearningText
🎯 What it does: A high-resolution dataset for Agda program proofs, AGDA2TRAIN, has been proposed, and a structured neural network, QUILL, has been developed based on it for premise selection;
Learning Structured Representations with Hyperbolic Embeddings
Aditya Sinha (University of Illinois), Han Zhao (University of Illinois)
Anomaly DetectionRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Design and implement a hierarchical structure regularization framework named HypStructure, which explicitly embeds the hierarchical information of the label tree into the feature space using hyperbolic space (Poincaré ball model), and incorporates two losses, HypCPCC and HypCenter, during training to reduce the distortion of the tree structure and enhance the tree-like nature of the features.
Learning Successor Features the Simple Way
Raymond Chua (McGill University), Doina Precup (McGill University)
Convolutional Neural NetworkReinforcement LearningImage
🎯 What it does: This paper proposes a simple method to directly learn Successor Features (SF) from pixels, without the need for pre-training or complex auxiliary losses, enabling real-time learning during task execution.
Learning symmetries via weight-sharing with doubly stochastic tensors
Putri A Van der Linden, Erik J Bekkers
ClassificationRecognitionOptimizationConvolutional Neural NetworkImage
🎯 What it does: Learn a weight sharing mechanism implemented through double compressible matrices to automatically discover and utilize symmetries in the data during the training process;
Learning the Expected Core of Strictly Convex Stochastic Cooperative Games
Nam Phuong Tran (University of Warwick), Long Tran-Thanh (University of Warwick)
🎯 What it does: This paper proposes an algorithm for learning the expected core in strictly convex stochastic cooperative games—Common-Points-Picking;
Learning the Infinitesimal Generator of Stochastic Diffusion Processes
Vladimir R Kostic, Massimiliano Pontil (University College London)
OptimizationComputational EfficiencyTime SeriesFinance RelatedPhysics RelatedStochastic Differential Equation
🎯 What it does: This paper proposes a kernel method based on the energy risk function to learn the infinitesimal generator (IG) of random diffusion processes from a single trajectory, and provides an upper bound on the learning error of its spectral decomposition.
Learning the Latent Causal Structure for Modeling Label Noise
Yexiong Lin (Sydney AI Centre, University of Sydney), Tongliang Liu (Sydney AI Centre, University of Sydney)
ClassificationData-Centric LearningAuto EncoderImage
🎯 What it does: This paper proposes a generative model based on latent causal structures, which automatically learns instance-dependent noise transition matrices using only noisy labeled data and a small number of clean samples, and uses them to correct the training of classifiers.
Learning the Optimal Policy for Balancing Short-Term and Long-Term Rewards
Qinwei Yang (Beijing Technology and Business University), Peng Wu (Beijing Technology and Business University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: This paper proposes a policy learning framework based on multi-objective optimization to learn optimal strategies while considering both short-term and long-term rewards.
Learning to Assist Humans without Inferring Rewards
Vivek Myers (University of California Berkeley), Anca Dragan (University of California Berkeley)
Reinforcement LearningContrastive LearningTabular
🎯 What it does: A reinforcement learning algorithm based on contrastive success representation is developed to train auxiliary agents to maximize human 'effective empowerment' without inferring human rewards.
Learning to Balance Altruism and Self-interest Based on Empathy in Mixed-Motive Games
Fanqi Kong (Peking University), Xue Feng (BIGAI)
Reinforcement LearningSequential
🎯 What it does: The LASE (Learning to Balance Altruism and Self-interest based on Empathy) algorithm is proposed, which calculates social relationships through an empathy mechanism in distributed multi-agent reinforcement learning and facilitates gift-giving to balance altruism and self-interest in mixed-motive games.
Learning to be Smooth: An End-to-End Differentiable Particle Smoother
Ali Younis (University of California), Erik B. Sudderth (University of California)
Object TrackingAutonomous DrivingSimultaneous Localization and MappingTime SeriesSequential
🎯 What it does: This paper proposes a differentiable bidirectional particle smoother (MDPS) that utilizes backward information from a combination of forward and backward particle filters to achieve a more accurate posterior distribution of states for complete time series.
Learning to compute Gröbner bases
Hiroshi Kera (Chiba University), Kazuhiro Yokoyama (Rikkyo University)
Transformer
🎯 What it does: This paper proposes a method for learning to compute Gröbner bases using Transformers and designs efficient algorithms for the random generation of Gröbner bases and the inverse Gröbner problem, addressing the challenges of dataset generation.
Learning to Cooperate with Humans using Generative Agents
Yancheng Liang (University of Washington), Natasha Jaques (University of Washington)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningAuto EncoderSequential
🎯 What it does: This paper proposes a generative agent (GAMMA) to train cooperative agents that can collaborate with unknown human partners in a zero-shot manner.
Learning to Decouple the Lights for 3D Face Texture Modeling
Tianxin Huang (National University of Singapore), Gim Hee Lee (National University of Singapore)
RestorationSegmentationConvolutional Neural NetworkContrastive LearningImageVideo
🎯 What it does: A 3D facial texture reconstruction framework is proposed that utilizes neural representation to separate ambient light, capable of recovering clear textures under unnatural lighting caused by external occlusions.
Learning to Discuss Strategically: A Case Study on One Night Ultimate Werewolf
Xuanfa Jin (Institute of Automation Chinese Academy of Sciences), Jun Wang (UCL)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This study investigates a social reasoning game called One Night Ultimate Werewolf (ONUW), proposing and validating a discussion strategy based on RL training, which is embedded in an LLM agent, achieving significantly improved performance in three-player and five-player games.
Learning to Edit Visual Programs with Self-Supervision
R. Kenny Jones (Brown University), Daniel Ritchie (Brown University)
Convolutional Neural NetworkTransformerImage
🎯 What it does: Learn how to edit visual programs and propose a self-supervised joint fine-tuning framework.
Learning to Embed Distributions via Maximum Kernel Entropy
Oleksii Kachaiev (Università degli Studi di Genova), Stefano Recanatesi (Technion Israel Institute of Technology)
ClassificationOptimizationImageTextMultimodalityBiomedical Data
🎯 What it does: A framework for unsupervised learning of data-dependent distribution kernels is proposed, utilizing the maximization of the quantum Renyi entropy of distribution embeddings to automatically construct kernel functions for distribution regression and classification tasks.
Learning to grok: Emergence of in-context learning and skill composition in modular arithmetic tasks
Tianyu He (University of Maryland), Andrey Gromov (Meta)
TransformerLarge Language ModelTabular
🎯 What it does: Exploring the contextual learning and task combination capabilities of modular arithmetic tasks on GPT-style Transformers, and analyzing their impact on O.O.D. generalization.
Learning to Handle Complex Constraints for Vehicle Routing Problems
Jieyi Bi (Nanyang Technological University), Jie Zhang (Nanyang Technological University)
OptimizationReinforcement LearningTabular
🎯 What it does: A proactive infeasibility prevention (PIP) framework is proposed to assist neural network constructors in generating feasible and high-quality solutions for vehicle routing problems under complex constraints.
Learning to Merge Tokens via Decoupled Embedding for Efficient Vision Transformers
Dong Hoon Lee (Korea Advanced Institute of Science and Technology), Seunghoon Hong (Korea Advanced Institute of Science and Technology)
ClassificationSegmentationTransformerImage
🎯 What it does: A decoupled Token Embedding (DTEM) is proposed to improve Token merging in visual Transformers.
Learning to Mitigate Externalities: the Coase Theorem with Hindsight Rationality
Antoine Scheid (Centre de Mathématiques Appliquées - CNRS - École polytechnique), Alain Oliviero Durmus
OptimizationReinforcement LearningTabular
🎯 What it does: In a two-party learning-based multi-armed bandit game, the impact of externalities on social welfare is studied, and it is proven that social welfare is irrecoverable in the absence of property rights. Subsequently, an online gaming strategy (BELGIC) is designed, allowing the downstream party to cooperate with the upstream party through learning optimal transfers (incentives), ultimately achieving an online version of the Coase theorem and reaching social welfare optimality.
Learning to Predict Structural Vibrations
Jan van Delden (University of Gttingen), Timo Lüddecke
Convolutional Neural NetworkTransformerPhysics Related
🎯 What it does: This paper proposes and evaluates a new Frequency-Query Operator (FQO) model for predicting the vibration modes and frequency responses of plates under different excitation frequencies by constructing a dataset of 12,000 vibrating plate samples with varying boundaries, materials, and detailed structures (beading).
Learning to Price Homogeneous Data
Keran Chen (University of Wisconsin Madison), Kirthevasan Kandasamy (University of Wisconsin Madison)
OptimizationReinforcement Learning
🎯 What it does: The study focuses on maximizing long-term revenue for sellers in a homogeneous data market by learning the optimal pricing curve when the types of buyers are unknown.
Learning to Reason Iteratively and Parallelly for Complex Visual Reasoning Scenarios
Shantanu Jaiswal (Carnegie Mellon University), Cheston Tan (Centre for Frontier AI Research)
OptimizationExplainability and InterpretabilityTransformerVision Language ModelImageVideo
🎯 What it does: A fully neural iterative and parallel reasoning mechanism named IPRM is proposed to handle complex visual question answering scenarios, capable of generating multiple operations in parallel at each reasoning step and updating memory iteratively.
Learning to Reason via Program Generation, Emulation, and Search
Nathaniel Weir (Johns Hopkins University), Peter Clark (Allen Institute for Artificial Intelligence)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Train the language model to generate pseudo-programs and have the model imitate execution, using program search (COTACS) to find the optimal program to complete diverse reasoning tasks.
Learning to Shape In-distribution Feature Space for Out-of-distribution Detection
Yonggang Zhang (Hong Kong Baptist University), Yiu-ming Cheung (Hong Kong Baptist University)
Anomaly DetectionRepresentation LearningImage
🎯 What it does: Proposes a Distributional Representation Learning (DRL) framework that actively shapes the ID feature space into a predefined mixed distribution during the pre-training of classification models, thereby achieving out-of-distribution (OOD) detection without distributional assumptions.
Learning to Solve Quadratic Unconstrained Binary Optimization in a Classification Way
Ming Chen (National University of Defense Technology), Yingwu Chen (National University of Defense Technology)
OptimizationGraph Neural NetworkBenchmark
🎯 What it does: A classification-based neural solver VCM is proposed, which utilizes a Deep Value Network (DVN) to extract value features of QUBO, and then outputs a global binary solution in one go through a Value Classification Network (VCN), training the model using a Greedy-guided Self Trainer (GST) without prior optimal labels.
Learning to Understand: Identifying Interactions via the Möbius Transform
Justin Singh Kang (University of California Berkeley), Kannan Ramchandran (University of California Berkeley)
Explainability and InterpretabilityComputational EfficiencyTransformerSupervised Fine-TuningTextTabular
🎯 What it does: A fully non-adaptive Sparse Möbius Transform (SMT) algorithm is proposed for efficiently reconstructing sparse, low-order interactions in machine learning models, enhancing model interpretability.
Learning Transferable Features for Implicit Neural Representations
Kushal Vyas (Rice University), Guha Balakrishnan (Rice University)
RestorationDomain AdaptationImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A new training framework called STRAINER is proposed, which uses implicit neural representations (INR) with shared initial layers to learn transferable features, enabling rapid fitting of similar signals and improving reconstruction quality.
Learning Truncated Causal History Model for Video Restoration
Amirhosein Ghasemabadi (University of Alberta), Di Niu (University of Alberta)
RestorationSuper ResolutionConvolutional Neural NetworkVideo
🎯 What it does: The TURTLE framework is proposed, utilizing a learnable Causal History Model (CHM) to perform various restoration tasks such as denoising, deraining, desnowing, deblurring, and super-resolution using only past frames.
Learning Versatile Skills with Curriculum Masking
Yao Tang (Shanghai Jiao Tong University), Shuai Li (Shanghai Jiao Tong University)
Robotic IntelligenceTransformerReinforcement LearningSequential
🎯 What it does: This paper proposes a curriculum-based masking method for unsupervised reinforcement learning pre-training called CurrMask, which combines block masking with automatic curriculum learning to enhance the learning and transfer of skills across different time scales.
Learning via Surrogate PAC-Bayes
Antoine Picard, Benjamin Guedj (Inria and University College London)
OptimizationMeta LearningBiomedical DataOrdinary Differential Equation
🎯 What it does: A framework (SuPAC) is proposed to iteratively optimize the PAC-Bayes generalization bounds by constructing a low-dimensional approximate risk, achieving a reduction in computational complexity for expensive empirical risk queries.
Learning Where to Edit Vision Transformers
Yunqiao Yang (City University of Hong Kong), Ying Wei (Zhejiang University)
Meta LearningTransformerImageBenchmark
🎯 What it does: A meta-learning based hypernetwork method is proposed for single-sample editing of pre-trained Vision Transformers (ViT), which first locates the parameters to be modified and then fine-tunes them.
Learning with Fitzpatrick Losses
Seta Rakotomandimby (Ecole des Ponts), Mathieu Blondel (Google DeepMind)
MultimodalityBenchmark
🎯 What it does: This paper proposes a novel family of loss functions based on the Fitzpatrick function—Fitzpatrick loss, which serves as a tight upper bound for the Fenchel-Young loss while maintaining the same link function;
Learning World Models for Unconstrained Goal Navigation
Yuanlin Duan (Rutgers University), He Zhu (Rutgers University)
Robotic IntelligenceReinforcement LearningWorld ModelSequential
🎯 What it does: This paper proposes a goal-oriented exploration algorithm named MUN, which can construct state transitions between arbitrary sub-goals in the replay buffer, thereby training a more reliable world model and goal-conditioned policy.
Learning-Augmented Algorithms for the Bahncard Problem
Hailiang Zhao (Zhejiang University), Shuiguang Deng (Zhejiang University)
OptimizationTime SeriesSequential
🎯 What it does: A new learning-enhanced algorithm PFSUM is proposed to solve the Bahncard problem, providing a competitive ratio under any prediction error; extensive experimental validation is also presented.
Learning-Augmented Algorithms with Explicit Predictors
Marek Elias, Shay Moran (Technion)
OptimizationTabularSequential
🎯 What it does: This paper studies online algorithm problems and proposes a learning-enhanced algorithm that utilizes predictive information obtained from historical and current data through machine learning models. It designs new algorithms to improve performance, particularly in fundamental issues such as caching and scheduling.
Learning-Augmented Approximation Algorithms for Maximum Cut and Related Problems
Vincent Cohen-Addad (Google Research), Debmalya Panigrahi (Duke University)
Optimization
🎯 What it does: This paper proposes a machine learning-based prediction method to enhance the approximation algorithms for offline NP-hard problems, particularly the maximum cut and more broadly, 2-CSP.
Learning-Augmented Dynamic Submodular Maximization
Arpit Agarwal (Indian Institute of Technology Bombay), Eric Balkanski (Columbia University)
OptimizationGraph
🎯 What it does: This paper studies the dynamic submodular maximization problem, maintaining an approximately optimal solution on sequences of dynamic insertions/deletions, and proposes an algorithm that accelerates updates using predictive information.
Learning-Augmented Priority Queues
Ziyad Benomar (ENSAE Ecole Polytechnique FairPlay joint team), Christian Coester (University of Oxford)
GraphBenchmark
🎯 What it does: This paper proposes a learning-enhanced priority queue under three prediction models (dirty comparison, pointer prediction, ranking prediction), implemented using skip lists and proves optimal comparison complexity;
Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching
Xinyin Ma (National University of Singapore), Xinchao Wang (National University of Singapore)
GenerationComputational EfficiencyTransformerDiffusion modelImageOrdinary Differential Equation
🎯 What it does: A learning layer caching mechanism (Learning-to-Cache) is proposed, which dynamically decides which layers can be cached during the denoising process of the diffusion Transformer, reducing computational load without updating model parameters.
Least Squares Regression Can Exhibit Under-Parameterized Double Descent
Xinyue Li (Yale University), Rishi Sonthalia (Boston College)
Supervised Fine-TuningTabular
🎯 What it does: Two linear regression models are proposed, which exhibit a bimodal descent phenomenon in the under-parameterized region, and it is proven that the peak positions are influenced by the sample covariance spectrum and the alignment of the target vector with the singular vectors.
LeDex: Training LLMs to Better Self-Debug and Explain Code
Nan Jiang (Purdue University), Anoop Deoras (AWS AI Labs)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: By constructing an automated data collection and validation pipeline, combined with supervised fine-tuning and reinforcement learning, the self-debugging capability of open-source LLMs is significantly enhanced.
Length Optimization in Conformal Prediction
Shayan Kiyani (University of Pennsylvania), Hamed Hassani (University of Pennsylvania)
OptimizationImageText
🎯 What it does: A length-optimal compliant prediction framework (CPL) is proposed under the constraint of conditional validity.
LESS: Label-Efficient and Single-Stage Referring 3D Segmentation
Xuexun Liu (Shenzhen University), Lin Ma (University of Trento)
SegmentationTransformerContrastive LearningPoint Cloud
🎯 What it does: A single-stage label-efficient referential 3D segmentation method called LESS is proposed, which only uses binary mask supervision.
Lever LM: Configuring In-Context Sequence to Lever Large Vision Language Models
Xu Yang (Southeast University), Hanwang Zhang (Nanyang Technological University)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a micro language model named Lever-LM, designed to generate high-quality context demonstration (ICD) sequences for large visual-language models (LVLM) to enhance their reasoning capabilities in visual-language tasks.
Leveraging an ECG Beat Diffusion Model for Morphological Reconstruction from Indirect Signals
Lisa Bedin (Ecole Polytechnique), Eric Moulines (Ecole Polytechnique)
GenerationAnomaly DetectionDiffusion modelTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: This paper proposes BeatDiff, a lightweight diffusion model for generating 12-lead heartbeat morphologies, and designs EM-BeatDiff based on this model to perform multiple ECG processing tasks such as denoising, missing lead reconstruction, and anomaly detection within a Bayesian inverse problem framework without requiring additional training.
Leveraging Catastrophic Forgetting to Develop Safe Diffusion Models against Malicious Finetuning
Jiadong Pan (Chinese Academy of Sciences), Liang Li (Chinese Academy of Sciences)
GenerationSafty and PrivacyDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes a method to induce catastrophic forgetting through contrastive learning and latent/noise transformation, allowing diffusion models to maintain safe generation of images without harmful content even after being maliciously fine-tuned.
Leveraging Contrastive Learning for Enhanced Node Representations in Tokenized Graph Transformers
Jinsong Chen (Huazhong University of Science and Technology), Kun He (Huazhong University of Science and Technology)
ClassificationRepresentation LearningGraph Neural NetworkTransformerContrastive LearningGraph
🎯 What it does: GCFormer is proposed, using both positive and negative token sequences along with a Transformer backbone for node classification.
Leveraging Drift to Improve Sample Complexity of Variance Exploding Diffusion Models
Ruofeng Yang (Shanghai Jiao Tong University), Shuai Li (Shanghai Jiao Tong University)
GenerationData SynthesisDiffusion modelScore-based ModelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A drifted variance diffusion model (Drifted VESDE) is proposed, and a unified analysis and sampling complexity proof for inverse SDE and inverse PFODE is provided based on this model.
Leveraging Environment Interaction for Automated PDDL Translation and Planning with Large Language Models
Sadegh Mahdavi (University of British Columbia), Yanshuai Cao (Borealis AI)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper automatically generates PDDL domain and problem files through interaction between a large language model and the environment, completely eliminating human intervention.
Leveraging Hallucinations to Reduce Manual Prompt Dependency in Promptable Segmentation
Jian Hu (Queen Mary University of London), Shaogang Gong (Queen Mary University of London)
Object DetectionSegmentationLarge Language ModelPrompt EngineeringDiffusion modelImageBiomedical DataChain-of-Thought
🎯 What it does: This paper proposes a training-free Prompt-Mask cyclic generation method called ProMaC, which utilizes multi-scale thinking chain prompts and visual contrast reasoning to automatically generate instance-level prompts from a single task general prompt and iteratively improve masks, achieving semantic segmentation without manual prompts.
Leveraging partial stragglers within gradient coding
Aditya Ramamoorthy (Iowa State University), Vrinda S Girimaji
OptimizationFederated Learning
🎯 What it does: A gradient encoding protocol is proposed that utilizes partial stragglers, supporting both exact gradient reconstruction and approximate reconstruction.
Leveraging Separated World Model for Exploration in Visually Distracted Environments
Kaichen Huang (Nanjing University), De-Chuan Zhan (Nanjing University)
Robotic IntelligenceReinforcement LearningWorld ModelVideo
🎯 What it does: A dual-layer optimization framework named SeeX is proposed, utilizing a separated world model for unsupervised reinforcement learning exploration in visually distracting environments.
Leveraging Tumor Heterogeneity: Heterogeneous Graph Representation Learning for Cancer Survival Prediction in Whole Slide Images
Junxian Wu (Anhui University), Lizhi Shao (Anhui University)
Representation LearningGraph Neural NetworkImageBiomedical Data
🎯 What it does: ProtoSurv utilizes heterogeneous graph neural networks combined with pathological prior knowledge to model tumor heterogeneity in whole slide images, achieving cancer survival risk prediction.
Leveraging Visual Tokens for Extended Text Contexts in Multi-Modal Learning
Alex Jinpeng Wang (National University of Singapore), Mike Zheng Shou
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposes the VisInContext method, which utilizes a visual encoder to convert long texts into image-based visual tokens, enabling efficient context length expansion in multimodal large models.
Lexicon3D: Probing Visual Foundation Models for Complex 3D Scene Understanding
Yunze Man (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)
Object DetectionSegmentationMixture of ExpertsDiffusion modelImageVideoMultimodality
🎯 What it does: This paper constructs a unified detection framework to systematically evaluate seven visual foundation models on four types of tasks in complex 3D scene understanding (visual language reasoning, visual alignment, semantic segmentation, and geometric registration).
LFME: A Simple Framework for Learning from Multiple Experts in Domain Generalization
Liang Chen (MBZUAI), Lingqiao Liu (University of Adelaide)
ClassificationSegmentationDomain AdaptationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A simple and effective domain generalization framework LFME is proposed, which utilizes expert models for each source domain to guide the target model through logit regularization during the training phase, allowing the target model to perform as an expert across all source domains, thereby achieving better generalization performance on unknown target domains.
LG-CAV: Train Any Concept Activation Vector with Language Guidance
Qihan Huang (Zhejiang University), Mingli Song (Zhejiang University)
ClassificationExplainability and InterpretabilityVision Language ModelImage
🎯 What it does: This paper proposes a method to train arbitrary Concept Activation Vectors (LG-CAV) using concept descriptions provided by the pre-trained vision-language model CLIP, allowing for the generation of high-quality CAVs without the need for extensive manual image labeling, which can then be used for model calibration.
LG-VQ: Language-Guided Codebook Learning
Liang Guotao, luolinfeng
GenerationData SynthesisTransformerGenerative Adversarial NetworkContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a language-guided codebook learning framework called LG-VQ, which utilizes pre-trained text semantics (CLIP) to guide VQ codebook learning, aligning it with text to enhance performance on multimodal tasks such as text-to-image, image captioning, and VQA.
Light Unbalanced Optimal Transport
Milena Gazdieva (Skolkovo Institute of Science and Technology), Alexander Korotin (Skolkovo Institute of Science and Technology)
Image TranslationOptimizationAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A lightweight and theoretically feasible continuous unbalanced entropy regularized optimal transport (UEOT) solver, called U-LightOT, is proposed;
LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS
Zhiwen Fan (University of Texas at Austin), Zhangyang Wang (Xiamen University)
CompressionComputational EfficiencyKnowledge DistillationGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes the LightGaussian method, which achieves a 15-fold compression and improves rendering FPS through pruning 3D Gaussian splines, gamma decomposition, SH reduction, and vector quantization.
Lighting Every Darkness with 3DGS: Fast Training and Real-Time Rendering for HDR View Synthesis
Xin Jin (Nankai University), Bo Ren (Nankai University)
RestorationData SynthesisComputational EfficiencyNeural Radiance FieldGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: Reconstruct HDR scenes from multi-perspective noisy RAW images, achieving real-time rendering and post-processing (exposure changes, HDR gamma mapping, focus transformation, etc.)
Lightweight Frequency Masker for Cross-Domain Few-Shot Semantic Segmentation
Jintao Tong (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
SegmentationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: In the cross-domain few-shot semantic segmentation task, a lightweight frequency mask (Amplitude-Phase Masker and Adaptive Channel Phase Attention) is introduced to reduce the channel correlation of feature maps, thereby enhancing the model's robustness to inter-domain differences and significantly improving mIoU.
Limits of Transformer Language Models on Learning to Compose Algorithms
Jonathan Thomm (IBM Research), Abbas Rahimi (IBM Research)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Study the ability and sample efficiency of Transformer language models in learning combinatorial algorithm tasks.
Linear Causal Bandits: Unknown Graph and Soft Interventions
Zirui Yan (Rensselaer Polytechnic Institute), Ali Tajer (Rensselaer Polytechnic Institute)
Graph
🎯 What it does: This paper proposes an implementable algorithm GA-LCB for the linear causal bandit problem with soft random interventions in an unknown causal graph, addressing exploration and optimal intervention selection when the graph structure is unknown.
Linear Causal Representation Learning from Unknown Multi-node Interventions
Burak Varıcı (Carnegie Mellon University), Ali Tajer (Rensselaer Polytechnic Institute)
Representation LearningGraph
🎯 What it does: This paper proposes a causal representation learning (CRL) framework for the scenario of unknown multi-node interventions (UMN) and provides the corresponding identifiability theory and implementation algorithm UMNI-CRL.
Linear Regression using Heterogeneous Data Batches
Ayush Jain (Granica Computing), Alon Orlitsky (University of California San Diego)
OptimizationTabular
🎯 What it does: Under multi-source small batch data, learn a multiple linear regression model and recover the regression vector for each subgroup.
Linear Time Approximation Algorithm for Column Subset Selection with Local Search
YuanBin Zou, Qilong Feng (Central South University)
OptimizationTabular
🎯 What it does: A column subset selection (CSS) algorithm based on local search is proposed, which obtains an approximate solution using a two-step mixed sampling method while ensuring linear time complexity.
Linear Transformers are Versatile In-Context Learners
Max Vladymyrov (Google Research), Rong Ge (Duke University)
OptimizationTransformerTabular
🎯 What it does: This paper studies the implicit optimization behavior of linear Transformers in context learning, proving that each layer maintains a linear regression model and can be viewed as a variant of preconditioned gradient descent; after training on a mixed noise linear regression task, the linear Transformer can automatically discover complex optimization algorithms that include momentum and adaptive re-scaling.
Linear Uncertainty Quantification of Graphical Model Inference
Chenghua Guo (Beijing University of Posts and Telecommunications), Xi Zhang (Beijing University of Posts and Telecommunications)
OptimizationExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A linear propagation-based uncertainty quantification method, LinUProp, is proposed for graphical model inference, which can quickly and analytically provide the interval width of posterior confidence and ensures convergence and linear scalability.
Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models
Shuxia Lin (Southeast University), Xin Geng (Southeast University)
CompressionKnowledge DistillationTransformerImage
🎯 What it does: A linear decomposition and recombination framework called 'learngene' is proposed, which allows for the generation of ViT models with different layers and scales from a large ViT model with just one training session.
Linguistic Collapse: Neural Collapse in (Large) Language Models
Robert Wu (University of Toronto), Vardan Papyan (University of Toronto)
GenerationTransformerLarge Language ModelText
🎯 What it does: This study investigates the phenomenon of neural collapse in autoregressive language models, conducting experiments with GPT-Neo models of varying sizes and training settings on the TinyStories dataset, and quantifying the relationship between four NC attributes and generalization.
Linking In-context Learning in Transformers to Human Episodic Memory
Li Ji-An (University of California San Diego), Marcelo G Mattar
TransformerLarge Language ModelText
🎯 What it does: This study investigates the relationship between 'induction heads' in the Transformer model and the human context memory (CMR) model, quantifying their similarities and assessing their causal impact on in-context learning (ICL).
LinNet: Linear Network for Efficient Point Cloud Representation Learning
Hao Deng (Northwest University), Lin Wang (Northwest University)
ClassificationSegmentationAutonomous DrivingRepresentation LearningPoint Cloud
🎯 What it does: A novel point cloud representation learning framework called LinNet is proposed, which employs linear complexity neighborhood search and downsampling, and introduces the Disassembled Set Abstraction (DSA) module to achieve lightweight and efficient local feature aggregation.
LION: Linear Group RNN for 3D Object Detection in Point Clouds
Zhe Liu (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
Object DetectionAutonomous DrivingRecurrent Neural NetworkPoint Cloud
🎯 What it does: A window-based 3D detection framework LION based on linear RNN is proposed, which enables long-range feature interaction in large groups, significantly improving the detection performance of sparse point clouds.
Lips Are Lying: Spotting the Temporal Inconsistency between Audio and Visual in Lip-Syncing DeepFakes
Weifeng Liu (Wuhan University), Run Wang (Nanyang Technological University)
ClassificationRecognitionTransformerVideoMultimodalityAudio
🎯 What it does: Proposes the LipFD method, which utilizes the temporal inconsistency between audio and lip movements to detect Lip-Syncing deepfake videos.
LISA: Layerwise Importance Sampling for Memory-Efficient Large Language Model Fine-Tuning
Rui Pan (University of Illinois Urbana-Champaign), Tong Zhang (University of Illinois Urbana-Champaign)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposes the Layerwise Importance Sampled AdamW (LISA) algorithm, which randomly freezes certain layers to achieve efficient fine-tuning of LLMs;
Lisa: Lazy Safety Alignment for Large Language Models against Harmful Fine-tuning Attack
Tiansheng Huang (Georgia Institute of Technology), Ling Liu (Georgia Institute of Technology)
OptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a method to reduce security drift caused by harmful fine-tuning attacks during the fine-tuning phase of large language models (LLMs) through Dual-State Optimization (BSO) and its improved version, Lisa.
Listenable Maps for Zero-Shot Audio Classifiers
Francesco Paissan (Fondazione Bruno Kessler), Cem Subakan (Concordia University)
ClassificationExplainability and InterpretabilityContrastive LearningAudio
🎯 What it does: This paper proposes LMAC-ZS, a post-hoc explanation method that generates audible attention masks through a decoder to explain the decisions of zero-shot audio classifiers like CLAP.
LiT: Unifying LiDAR "Languages" with LiDAR Translator
Yixing Lao (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
Object DetectionDomain AdaptationAutonomous DrivingNeural Radiance FieldPoint Cloud
🎯 What it does: This paper proposes the LiDAR Translator (LiT), which unifies data from different LiDAR domains into the same 'language' through scene reconstruction and ray casting.
LiteVAE: Lightweight and Efficient Variational Autoencoders for Latent Diffusion Models
Seyedmorteza Sadat (ETH Zurich), Romann M. Weber (Disney Research Studios)
GenerationData SynthesisComputational EfficiencyDiffusion modelAuto EncoderImage
🎯 What it does: A lightweight VAE architecture called LiteVAE is designed, which preprocesses images using 2D discrete wavelet transform, constructs a multi-scale feature extraction and aggregation module, significantly reducing encoder parameters while maintaining or improving reconstruction quality.
LIVE: Learnable In-Context Vector for Visual Question Answering
Yingzhe Peng (Southeast University), Xu Yang (Southeast University)
RecognitionOptimizationComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: A learnable In-Context Vector (LIVE) is proposed to simulate traditional In-Context Learning (ICL) in large multimodal models, enabling the completion of visual question answering tasks without the need for a large number of demonstrations.
LiveScene: Language Embedding Interactive Radiance Fields for Physical Scene Control and Rendering
Delin Qu (Fudan University), Xuelong Li (Fudan University)
GenerationData SynthesisNeural Radiance FieldContrastive LearningVideoBenchmark
🎯 What it does: This paper proposes LiveScene, a language-embedded interactive radiance field that can reconstruct and control multiple interactive objects from monocular video, enabling free control of object states through natural language.
LLaMo: Large Language Model-based Molecular Graph Assistant
Jinyoung Park (Korea University), Hyunwoo J. Kim (Korea University)
GenerationDrug DiscoveryGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: LLaMo is proposed, an end-to-end trained large molecular graph language model that can simultaneously perform tasks such as molecular description generation, IUPAC name prediction, and property prediction.
LLaNA: Large Language and NeRF Assistant
Andrea Amaduzzi (University of Bologna), Luigi di Stefano
ClassificationGenerationData SynthesisTransformerLarge Language ModelNeural Radiance FieldTextMultimodalityPoint Cloud
🎯 What it does: LLaNA is constructed, a multimodal large language model that can directly encode NeRF (Neural Radiance Fields) weights and combine with LLM to perform NeRF captioning, question answering, and zero-shot classification tasks.
LLM Circuit Analyses Are Consistent Across Training and Scale
Curt Tigges (EleutherAI), Stella Biderman (EleutherAI)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Track and analyze the formation, evolution, and algorithmic stability of circuits (mechanisms) in the Pythia series of large language models during 3×10¹¹ training tokens across parameter scales from 300 million to 1.2 billion.
LLM Dataset Inference: Did you train on my dataset?
Pratyush Maini (Carnegie Mellon University), Adam Dziedzic (CISPA Helmholtz Center for Information Security)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper investigates the reasons for the failure of membership inference attacks (MIA) on large-scale language models and proposes a 'dataset inference' method based on multiple MIA aggregations to detect whether a certain dataset has been used for training LLMs.