ICML 2025 Papers — Page 14
International Conference on Machine Learning · 3257 papers
Harmonizing Geometry and Uncertainty: Diffusion with Hyperspheres
Muskan Dosi (Indian Institute of Technology Jodhpur), Richa Singh (Indian Institute of Technology Jodhpur)
GenerationData SynthesisDiffusion modelScore-based ModelImage
🎯 What it does: This paper proposes HyperSphereDiff, a diffusion model that uses von Mises–Fisher noise on the hypersphere to maintain class geometric structures and capture directional uncertainty.
Harnessing Heterogeneous Statistical Strength for Personalized Federated Learning via Hierarchical Bayesian Inference
Mahendra Singh Thapa (Rochester Institute of Technology), Rui Li (Rochester Institute of Technology)
Federated LearningTabular
🎯 What it does: A hierarchical Bayesian inference framework is proposed for simultaneously learning global and personalized models in federated learning.
HashAttention: Semantic Sparsity for Faster Inference
Aditya Desai (University of California Berkeley), Ion Stoica (University of California Berkeley)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: By mapping key-value pairs and query vectors to a hash space, sparse attention is achieved by utilizing a learned mapping function and Hamming distance retrieval to identify important tokens in the attention.
Haste Makes Waste: A Simple Approach for Scaling Graph Neural Networks
Rui Xue (North Carolina State University), Xiaorui Liu (North Carolina State University)
ClassificationOptimizationGraph Neural NetworkGraph
🎯 What it does: A training framework named REST is proposed, which significantly reduces feature staleness and improves the training effectiveness and convergence speed of large-scale graph neural networks by decoupling forward and backward propagation in historical embedding methods and adjusting their execution frequency.
Heads up! Large Language Models Can Perform Tasks Without Your Instruction via Selective Attention Head Masking
Senyu Han (Shanghai Jiao Tong University), Lu Chen (Shanghai Jiao Tong University)
GenerationTransformerLarge Language ModelText
🎯 What it does: The study found that by selecting or masking specific attention heads during the inference phase, LLMs can perform specific tasks without prompts, and this is achieved by training a binarized head mask.
HealthGPT: A Medical Large Vision-Language Model for Unifying Comprehension and Generation via Heterogeneous Knowledge Adaptation
Tianwei Lin (Zhejiang University), Beng Chin Ooi (National University of Singapore)
Image TranslationGenerationSuper ResolutionTransformerLarge Language ModelMixture of ExpertsVision Language ModelImageTextMultimodalityMagnetic Resonance ImagingComputed Tomography
🎯 What it does: This paper presents HealthGPT, a unified medical large vision language model that integrates visual understanding and generation capabilities of medical images.
HEAP: Hyper Extended A-PDHG Operator for Constrained High-dim PDEs
Mingquan Feng (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationTabularFinance RelatedPhysics Related
🎯 What it does: This paper proposes a new neural operator, HEAP, for solving high-dimensional constrained partial differential equations (PDEs);
Heavy-Tailed Linear Bandits: Huber Regression with One-Pass Update
Jing Wang (Nanjing University), Zhi-Hua Zhou (Nanjing University)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes a one-pass updated heavy-tailed linear contextual bandit algorithm Hvt-UCB, which achieves low computational cost and optimal mutation-informed convergence rate by utilizing the online mirror descent framework and adaptive Huber regression.
Hessian Geometry of Latent Space in Generative Models
Alexander Lobashev (Glam AI), Mikhail Tamm (Tallinn University)
GenerationData SynthesisOptimizationDiffusion modelImage
🎯 What it does: This study investigates the Fisher metric of the latent space of generative models, proposing a reconstruction of Hessian geometry through posterior approximation and learning of the log partition function, validated on Ising, TASEP, and diffusion models.
Heterogeneous Data Game: Characterizing the Model Competition Across Multiple Data Sources
Renzhe Xu (Shanghai University of Finance and Economics), Bo Li (Tsinghua University)
Tabular
🎯 What it does: Proposed the 'Heterogeneous Data Game' framework to study the competition among multi-model suppliers across different data sources;
Heterogeneous Label Shift: Theory and Algorithm
Chao Xu (National University of Defense Technology), Chenping Hou (National University of Defense Technology)
Domain AdaptationGenerative Adversarial NetworkTextMultimodality
🎯 What it does: This paper proposes a theoretical framework for Heterogeneous Label Shift (HLS) and designs the HLSAN network to achieve cross-modal knowledge transfer.
Heterogeneous Sufficient Dimension Reduction and Subspace Clustering
Lei Yan (Florida State University), Qing Mai (Florida State University)
OptimizationTabular
🎯 What it does: A mixed principal component fitting model (mixPFC) is proposed to achieve clustering, low-dimensional subspace estimation, and variable selection.
Heterogeneous Treatment Effect in Time-to-Event Outcomes: Harnessing Censored Data with Recursively Imputed Trees
Tomer Meir (Technion), Malka Gorfine (Tel Aviv University)
Biomedical DataElectronic Health Records
🎯 What it does: This paper proposes a method for estimating heterogeneous treatment effects on survival time based on multiple imputation, called MISTR, which can directly utilize all observations in the presence of right-censored data.
HetSSNet: Spatial-Spectral Heterogeneous Graph Learning Network for Panchromatic and Multispectral Images Fusion
Mengting Ma (Zhejiang University), Wei Zhang (Zhejiang University)
Image TranslationRestorationGraph Neural NetworkContrastive LearningImage
🎯 What it does: A spatial-spectral heterogeneous graph learning network (HetSSNet) is proposed for the fusion of full-resolution multispectral images and high-resolution single-spectral images in remote sensing.
Hgformer: Hyperbolic Graph Transformer for Collaborative Filtering
Xin Yang (University of Tsukuba), Erxue Min (Baidu Inc.)
Recommendation SystemGraph Neural NetworkTransformerGraph
🎯 What it does: A collaborative filtering model called Hgformer based on hypercurvature graph transformers has been developed to address the issues of local structure modeling and embedding distortion in long-tail product recommendations.
HGOT: Self-supervised Heterogeneous Graph Neural Network with Optimal Transport
Yanbei Liu (Tiangong University), Xiao Wang (Beihang University)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes a self-supervised heterogeneous graph neural network (HGOT) that aligns meta-path views and aggregation views through optimal transport, eliminating the need for graph augmentation and positive-negative sample selection, and directly learning high-quality node representations.
Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models
Lucy Xiaoyang Shi (Physical Intelligence), Chelsea Finn (Physical Intelligence)
Robotic IntelligenceTransformerVision Language ModelVision-Language-Action ModelMultimodality
🎯 What it does: A hierarchical visual-language-action model (Hi Robot) is proposed, capable of understanding and responding to complex natural language instructions, corrections, and feedback in real-time during robot execution, completing multi-stage tasks.
Hi-Patch: Hierarchical Patch GNN for Irregular Multivariate Time Series
Yicheng Luo (South China University of Technology), Qianli Ma (South China University of Technology)
ClassificationAnomaly DetectionGraph Neural NetworkTime SeriesBiomedical Data
🎯 What it does: This paper presents Hi-Patch, a hierarchical patch graph neural network designed for modeling irregular multivariate time series (IMTS) and performing prediction and classification tasks.
Hidden No More: Attacking and Defending Private Third-Party LLM Inference
Rahul Krishna Thomas (Ritual AI), Arka Pal (Ritual AI)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: The paper proposes a vocabulary-matching-based attack method that can almost perfectly recover the original user prompt text from the intermediate hidden states of LLMs (even when these hidden states are replaced or noised). It demonstrates that existing privacy inference schemes based on replacement are insecure in open-source weight environments. Additionally, it designs a Cascade multi-party inference framework that utilizes token dimension sharding to defend against this attack while maintaining low communication and computational overhead.
Hide & Seek: Transformer Symmetries Obscure Sharpness & Riemannian Geometry Finds It
Marvin F. da Silva (Dalhousie University), Sageev Oore (Vector Institute for Artificial Intelligence)
TransformerImageText
🎯 What it does: This paper studies the relationship between 'sharpness' in Transformer models and generalization performance, proposing a geodesic sharpness defined on Riemannian quotient manifolds, and proving that this metric can better predict the generalization of Transformers.
Hierarchical Equivariant Policy via Frame Transfer
Haibo Zhao (Northeastern University), Robert Platt (Northeastern University)
Robotic IntelligenceReinforcement LearningDiffusion modelPoint CloudBenchmark
🎯 What it does: A Hierarchical Equivariant Policy (HEP) is proposed, which predicts key poses at a high level and uses a 'Frame Transfer' interface to serve as a coordinate frame for the low level, thereby achieving trajectory generation from coarse to fine.
Hierarchical Graph Tokenization for Molecule-Language Alignment
Yongqiang Chen (Carnegie Mellon University), Yatao Bian (National University of Singapore)
Drug DiscoveryGraph Neural NetworkSupervised Fine-TuningGraph
🎯 What it does: This paper presents HIGHT, which utilizes a hierarchical graph tokenizer and hierarchical instruction tuning data to achieve better alignment between molecular graphs and language.
Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots
Guangting Zheng (University of Science and Technology of China), Tao Mei (HiDream.ai Inc)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: A hierarchical masked autoregressive model Hi-MAR is designed and implemented, using low-resolution image tokens as pivots to first generate global structures and then refine details.
Hierarchical Overlapping Clustering on Graphs: Cost Function, Algorithm and Scalability
Yicheng Pan (Beihang University), Bingchen Fan (Beihang University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A new Hierarchical Overlapping Clustering (HOC) cost function is proposed, along with corresponding approximate algorithms and a scalable accelerated version.
Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification
Flavio Petruzzellis (University of Padova), Pietro Lio
Robotic IntelligenceTransformerLarge Language ModelGraphRetrieval-Augmented Generation
🎯 What it does: This paper proposes a neural-symbolic framework called HVR, which combines large language models (LLMs) with knowledge graph retrieval-augmented generation (KG-RAG), hierarchical planning, and symbolic verification for robot planning and execution of complex tasks.
Hierarchical Refinement: Optimal Transport to Infinity and Beyond
Peter Halmos (Princeton University), Benjamin Raphael
OptimizationImageTabular
🎯 What it does: A multi-scale method called Hierarchical Refinement (HiRef) has been developed, which recursively refines dataset partitions using low-rank optimal transport subproblems, ultimately achieving a complete injective Monge mapping while ensuring linear space complexity.
Hierarchical Reinforcement Learning with Targeted Causal Interventions
Mohammadsadegh Khorasani, Matthias Grossglauser (Ecole Polytechnique Fédérale de Lausanne)
Reinforcement LearningGraphSequential
🎯 What it does: This paper proposes a hierarchical reinforcement learning framework HRC based on causal intervention, which can accelerate the learning of the final goal by discovering the causal structure of sub-goals and utilizing targeted interventions.
Hierarchical Reinforcement Learning with Uncertainty-Guided Diffusional Subgoals
Vivienne Huiling Wang (Aalto University), Joni Pajarinen (Aalto University)
Reinforcement LearningDiffusion modelSequential
🎯 What it does: This paper proposes a subgoal generation framework based on a conditional diffusion model, utilizing Gaussian process priors for regularization and designing a mixed subgoal selection strategy to address the issue of subgoal failure caused by low-level policy drift in hierarchical reinforcement learning.
High Dynamic Range Novel View Synthesis with Single Exposure
Kaixuan Zhang (Nanjing University of Science and Technology), Xiatian Zhu (University of Surrey)
GenerationData SynthesisNeural Radiance FieldGaussian SplattingImage
🎯 What it does: A new single-exposure HDR view synthesis method called Mono-HDR-3D is proposed, which utilizes single-exposure LDR images for HDR 3D scene modeling.
High Probability Bound for Cross-Learning Contextual Bandits with Unknown Context Distributions
Ruiyuan Huang (Fudan University), Zengfeng Huang (Fudan University)
Reinforcement LearningTabular
🎯 What it does: This study investigates the multi-armed bandit problem in cross-learning scenarios with unknown contextual distributions and proves that the original algorithm can achieve approximately optimal returns with high probability.
High-Dimensional Prediction for Sequential Decision Making
Georgy Noarov (University of Pennsylvania), Stephan Xie (Carnegie Mellon University)
OptimizationReinforcement Learning
🎯 What it does: A high-dimensional prediction framework is proposed, which can make unbiased predictions of multi-dimensional states in an online adversarial environment, and this framework is applied to tasks such as multi-agent decision-making, combinatorial optimization, and multi-class calibration.
High-Dimensional Tensor Regression With Oracle Properties
Wenbin Wang (ShanghaiTech University), Ziping Zhao (ShanghaiTech University)
RestorationOptimizationImage
🎯 What it does: A high-dimensional tensor regression model is proposed, utilizing non-convex regularization to achieve sparse/low-rank structure estimation, and theoretical risk bounds and a proximal gradient acceleration algorithm are provided;
High-Fidelity Simultaneous Speech-To-Speech Translation
Tom Labiausse (Kyutai), Neil Zeghidour (Kyutai)
GenerationData SynthesisTransformerTextAudio
🎯 What it does: A decoder-only multi-stream model named Hibiki is proposed, capable of real-time synchronous processing of source language speech and generating target language text and speech (S2TT + S2ST).
Highly Compressed Tokenizer Can Generate Without Training
Lukas Lao Beyer (Massachusetts Institute of Technology), Kaiming He (Meta)
GenerationData SynthesisOptimizationDiffusion modelImage
🎯 What it does: Utilizing a pre-trained one-dimensional high-compression image tokenizer (TiTok) to achieve image editing, inpainting, and zero-shot text generation through gradient optimization of discrete/continuous tokens.
HiRemate: Hierarchical Approach for Efficient Re-materialization of Neural Networks
Julia Gusak (Inria), Olivier Beaumont (Inria)
OptimizationComputational EfficiencyTransformer
🎯 What it does: This paper proposes HIREMATE, a hierarchical rematerialization framework for efficiently training deep networks under GPU memory constraints.
History-Guided Video Diffusion
Kiwhan Song (Massachusetts Institute of Technology), Vincent Sitzmann (Massachusetts Institute of Technology)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: This paper proposes the Diffusion Forcing Transformer (DFoT) and the History Guidance method, enabling video diffusion models to generate based on historical frames of arbitrary length.
Holistic Physics Solver: Learning PDEs in a Unified Spectral-Physical Space
Xihang Yue (Zhejiang University), Linchao Zhu (Zhejiang University)
OptimizationComputational EfficiencyTabularPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes the Holistic Physics Mixer (HPM), which constructs a new neural operator for solving partial differential equations by simultaneously utilizing spectral structure and point state information in a unified spectral-physical space.
Homophily Enhanced Graph Domain Adaptation
Ruiyi Fang (Western University), Boyu Wang (Western University)
Domain AdaptationGraph Neural NetworkGraph
🎯 What it does: A graph domain adaptation framework HGDA based on homogeneity alignment is proposed to address the performance decline caused by homogeneous distribution shifts in cross-network node classification.
How Compositional Generalization and Creativity Improve as Diffusion Models are Trained
Alessandro Favero (École Polytechnique Fédérale de Lausanne), Matthieu Wyart (Johns Hopkins University)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: This study investigates how diffusion models learn and utilize hierarchical combination rules to achieve synthetic generalization and creativity, demonstrating through theory and experiments that their sample complexity grows polynomially with hierarchy.
How Contaminated Is Your Benchmark? Measuring Dataset Leakage in Large Language Models with Kernel Divergence
Hyeong Kyu Choi (University of Wisconsin Madison), Yixuan Li (University of Wisconsin Madison)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: A kernel divergence score (KDS) is proposed to measure the extent of data leakage in large language models during benchmark testing.
How Distributed Collaboration Influences the Diffusion Model Training? A Theoretical Perspective
Jing Qiao (Shandong University), Dongxiao Yu (Shandong University)
GenerationOptimizationFederated LearningSafty and PrivacyDiffusion modelScore-based ModelStochastic Differential Equation
🎯 What it does: This paper theoretically studies the training performance of diffusion models in distributed environments with significant differences in computational resources and data availability, and provides an upper bound on the generative error of distributed training.
How Do Images Align and Complement LiDAR? Towards a Harmonized Multi-modal 3D Panoptic Segmentation
Yining Pan (Singapore University of Technology and Design), Na Zhao (Singapore University of Technology and Design)
SegmentationAutonomous DrivingTransformerImageMultimodalityPoint Cloud
🎯 What it does: A multi-modal LiDAR-image collaborative 3D panoramic segmentation framework IAL is proposed;
How Do Large Language Monkeys Get Their Power (Laws)?
Rylan Schaeffer (Stanford), Sanmi Koyejo (Stanford)
Large Language ModelText
🎯 What it does: This study explores the power-law scaling relationship of the success rate of large language models on different tasks as the reasoning computation improves, through multiple independent reasoning attempts (Repeat Sampling);
How Do Transformers Learn Variable Binding in Symbolic Programs?
Yiwei Wu, Raphaël Millière (Macquarie University)
TransformerLarge Language ModelText
🎯 What it does: Train a Transformer to handle simple programs with variable assignments, learning variable binding and reflection.
How does Labeling Error Impact Contrastive Learning? A Perspective from Data Dimensionality Reduction
Jun Chen (Huazhong Agricultural University), Yiming Ying (University of Sydney)
ClassificationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper theoretically analyzes the impact of label noise on the downstream classification performance of contrastive learning and proposes reducing label noise through singular value decomposition (SVD) dimensionality reduction of the original data, thereby enhancing the effectiveness of contrastive learning.
How Effective Can Dropout Be in Multiple Instance Learning ?
Wenhui Zhu (Arizona State University), Yalin Wang
ClassificationOptimizationImageBiomedical Data
🎯 What it does: This study investigates the role of dropout in multi-instance learning and proposes the MIL-Dropout method, which specifically drops the most important instances.
How Expressive are Knowledge Graph Foundation Models?
Xingyue Huang (University of Oxford), Miguel Romero Orth (Universidad Catolica de Chile)
Knowledge DistillationRepresentation LearningGraph Neural NetworkSupervised Fine-TuningGraph
🎯 What it does: A general framework called MOTIF is proposed to evaluate and enhance the expressiveness of Knowledge Graph Foundation Models (KGFM), clarifying the impact of different motifs on relation invariance learning.
How Far Is Video Generation from World Model: A Physical Law Perspective
Bingyi Kang (ByteDance Research), Jiashi Feng (ByteDance Research)
GenerationTransformerDiffusion modelWorld ModelVideoPhysics Related
🎯 What it does: Evaluate the generalization ability of video generation models in discovering physical laws, construct 2D geometric simulation data, and conduct ID, OOD, and compositional generalization experiments.
How Much Can Transfer? BRIDGE: Bounded Multi-Domain Graph Foundation Model with Generalization Guarantees
Haonan Yuan (Beihang University), Philip S. Yu (University of Illinois)
ClassificationDomain AdaptationGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: BRIDGE is proposed, a multi-domain graph-based model that employs domain-invariant aligners, lightweight Mixture of Experts, and spectral regularization for pre-training and prompt fine-tuning to achieve cross-domain graph learning.
How Much Can We Forget about Data Contamination?
Sebastian Bordt (University of Tübingen), Ulrike von Luxburg (University of Tübingen)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This study investigates the impact of benchmark data leakage (data contamination) on the evaluation of large language models, quantifying the degrees of overfitting and forgetting, and exploring the interactions between model size, training data volume, and contamination frequency.
How to Evaluate and Mitigate IP Infringement in Visual Generative AI?
Zhenting Wang (Rutgers University), Lingjuan Lyu (Sony AI)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextBenchmark
🎯 What it does: In response to the potential infringement of intellectual property rights by visual generative AI, this paper constructs an IP infringement assessment benchmark and proposes a defense framework called TRIM, based on large visual language models and classifier-free guidance, to detect and suppress IP infringement in generated content.
How to Move Your Dragon: Text-to-Motion Synthesis for Large-Vocabulary Objects
Wonkwang Lee (Seoul National University), Byeong-Uk Lee (KRAFTON)
GenerationData SynthesisTransformerDiffusion modelText
🎯 What it does: Implement text-driven motion synthesis on large vocabulary objects.
How to set AdamW's weight decay as you scale model and dataset size
Xi Wang (Johns Hopkins University), Laurence Aitchison (University of Bristol)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper views the weight update of AdamW as an Exponential Moving Average (EMA) and studies how weight decay adjusts with changes in model size and dataset size from this perspective.
How to Synthesize Text Data without Model Collapse?
Xuekai Zhu (Shanghai Jiao Tong University), Bowen Zhou (Tsinghua University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextFinance Related
🎯 What it does: This study investigates the impact of synthetic text data on language model training and proposes a semi-synthetic data method based on token-level editing to avoid model collapse.
How to Train Your Multi-Exit Model? Analyzing the Impact of Training Strategies
Piotr Kubaty (Jagiellonian University), Kamil Adamczewski (Wroclaw University of Science and Technology)
ClassificationComputational EfficiencyConvolutional Neural NetworkTransformerImageText
🎯 What it does: This paper systematically evaluates three training strategies for multi-exit models (disjoint, joint, mixed) and proposes a hybrid strategy that pre-trains the backbone network followed by joint training, demonstrating significant improvements in accuracy and efficiency across multiple budgets, different networks, and datasets.
How Transformers Learn Regular Language Recognition: A Theoretical Study on Training Dynamics and Implicit Bias
Ruiquan Huang (Pennsylvania State University), Jing Yang (University of Virginia)
RecognitionTransformerSequentialChain-of-Thought
🎯 What it does: This study investigates how a single layer Transformer learns to discriminate regular language tasks (even pairs and parity) through attention and linear layers during training.
How Transformers Learn Structured Data: Insights From Hierarchical Filtering
Jerome Garnier-Brun (Bocconi University), Luca Saglietti (Bocconi University)
GenerationData SynthesisTransformerSequential
🎯 What it does: This paper investigates how the standard Transformer encoder approximates the exact inference (Belief Propagation) algorithm when learning hierarchical structured data by constructing an adjustable tree sequence generation model, revealing its learning dynamics and attention distribution.
HPS: Hard Preference Sampling for Human Preference Alignment
Xiandong Zou (Singapore Management University), Pan Zhou (Singapore Management University)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes the Hard Preference Sampling (HPS) framework to align the outputs of large language models (LLMs) with human preferences, with a particular focus on the strong rejection of harmful content.
Human Body Restoration with One-Step Diffusion Model and A New Benchmark
Jue Gong (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
RestorationDiffusion modelImageBenchmark
🎯 What it does: This paper proposes an end-to-end process for portrait restoration, including dataset construction, model design, and training, and is validated on the newly constructed PERSONA dataset.
Human Cognition-Inspired Hierarchical Fuzzy Learning Machine
Junbiao Cui (Shanxi University), Jiye Liang (Shanxi University)
ClassificationExplainability and InterpretabilityImage
🎯 What it does: A Hierarchical Cognitive-based Fuzzy Learning Machine (HC-HFLM) is proposed, which guides model learning by integrating fuzzy similarity relationships in the knowledge system.
Human-Aligned Image Models Improve Visual Decoding from the Brain
Nona Rajabi (KTH Royal Institute of Technology), Danica Kragic (KTH Royal Institute of Technology)
RetrievalExplainability and InterpretabilityRepresentation LearningTransformerContrastive LearningImageMultimodalityMagnetic Resonance Imaging
🎯 What it does: Using an image encoder consistent with human perception, brain EEG/MEG signals are mapped to image space for image retrieval.
Hybrid Batch Normalisation: Resolving the Dilemma of Batch Normalisation in Federated Learning
Hongyao Chen (Jiangnan University), Josef Kittler (University of Surrey)
Federated LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes Hybrid Batch Normalization (HBN), which achieves adaptive fusion of local batch normalization and global normalization in federated learning by separating statistical parameters from learnable parameters, using global unbiased statistical estimates, and introducing a learnable mixing factor α, thereby addressing the performance degradation issue of traditional BN in non-IID clients.
Hybrid Quantum-Classical Multi-Agent Pathfinding
Thore Gerlach (University of Bonn), Nico Piatkowski (Fraunhofer IAIS)
OptimizationBenchmark
🎯 What it does: Two multi-agent pathfinding (MAPF) algorithms based on quantum-classical hybrid approaches are proposed, namely QUBO-and-Price (QP) and QUBO-and-Cut-and-Price (QCP).
Hybrid Spiking Vision Transformer for Object Detection with Event Cameras
Qi Xu (Dalian University of Technology), Gang Pan (Zhejiang University)
Object DetectionSpiking Neural NetworkTransformerTime Series
🎯 What it does: A hybrid spiking visual Transformer (HsVT) that integrates ANN and SNN is proposed for object detection using event cameras.
HybridGS: High-Efficiency Gaussian Splatting Data Compression using Dual-Channel Sparse Representation and Point Cloud Encoder
Qi Yang (University of Missouri Kansas City), Zhu Li (University of Missouri Kansas City)
CompressionGaussian SplattingPoint Cloud
🎯 What it does: Proposes the HybridGS framework, which first generates explicit compact 3D Gaussian Splatting (3DGS) data and then uses a standard point cloud encoder for efficient compression;
HYGMA: Hypergraph Coordination Networks with Dynamic Grouping for Multi-Agent Reinforcement Learning
Chiqiang Liu (Beijing University of Chemical Technology), Dazi Li (Beijing University of Chemical Technology)
Graph Neural NetworkReinforcement LearningGraph
🎯 What it does: The HYGMA framework is proposed, utilizing dynamic spectral clustering and hypergraph neural networks to achieve adaptive grouping and information transmission for multi-agent coordination.
Hyper-Transforming Latent Diffusion Models
Ignacio Peis (Technical University of Denmark), Jes Frellsen (Technical University of Denmark)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImageTime Series
🎯 What it does: A generative framework LDMI is proposed, which combines implicit neural representations (INR) and latent diffusion models (LDM), and a Transformer-based probabilistic hypernetwork HD is designed for scalable generation and reconstruction of continuous functions.
Hyper: Hyperparameter Robust Efficient Exploration in Reinforcement Learning
Yiran Wang (University of California), Lin Yang
Reinforcement Learning
🎯 What it does: A reinforcement learning algorithm named Hyper is proposed, which introduces a relocation phase between exploration and exploitation and separates the two strategies to achieve robustness against the curiosity reward coefficient β.
Hyperband-based Bayesian Optimization for Black-box Prompt Selection
Lennart Schneider (Amazon Web Services), Felice Antonio Merra (Cognism)
OptimizationLarge Language ModelPrompt EngineeringText
🎯 What it does: This study proposes the HbBoPs method, which utilizes a structure-aware deep kernel Gaussian process and the Hyperband multi-precision scheduler to achieve optimization that balances sample and query efficiency in black-box prompt selection tasks.
Hyperbolic-PDE GNN: Spectral Graph Neural Networks in the Perspective of A System of Hyperbolic Partial Differential Equations
Juwei Yue (Institute of Information Engineering, Chinese Academy of Sciences), Li Guo (Institute of Information Engineering, Chinese Academy of Sciences)
Explainability and InterpretabilityRepresentation LearningGraph Neural NetworkGraphStochastic Differential Equation
🎯 What it does: This paper proposes Hyperbolic-PDE GNN by modeling the message passing of graph neural networks as a set of hyperbolic partial differential equations, allowing node representations to be directly mapped to the solution space spanned by the Laplacian eigenvectors, thereby enhancing interpretability and expressiveness.
HyperIMTS: Hypergraph Neural Network for Irregular Multivariate Time Series Forecasting
Boyuan Li (South China University of Technology), Qianli Ma (South China University of Technology)
Graph Neural NetworkTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: Designed HyperIMTS, a prediction model for IMTS based on hypergraph neural networks, which directly constructs nodes from observations and connects them with hyperedges of time and variables without the need for padding;
HyperIV: Real-time Implied Volatility Smoothing
Yongxin Yang (Queen Mary University of London), Timothy Hospedales (University of Edinburgh)
GenerationOptimizationComputational EfficiencyTransformerTime SeriesFinance Related
🎯 What it does: We propose HyperIV, a method for generating arbitrage-free implied volatility surfaces under real-time sparse observations using hypernetworks, with a generation speed of only 2 milliseconds;
HyperNear: Unnoticeable Node Injection Attacks on Hypergraph Neural Networks
Tingyi Cai (Zhejiang Normal University), Yi Wang (Zhejiang Normal University)
Adversarial AttackGraph Neural NetworkGraph
🎯 What it does: This paper proposes HyperNear—a stealthy node injection attack framework for Hypergraph Neural Networks (HNN), achieving attack concealment through the preservation of homophily.
Hyperspherical Normalization for Scalable Deep Reinforcement Learning
Hojoon Lee (KAIST), Jaegul Choo (KAIST)
Reinforcement Learning
🎯 What it does: A reinforcement learning architecture named SimbaV2 is proposed, which utilizes hyperspherical normalization to simultaneously constrain the norms of weights, features, and gradients, addressing the issues of overfitting and instability during the scaling of RL training.
HyperTree Planning: Enhancing LLM Reasoning via Hierarchical Thinking
Runquan Gui (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Proposes the HyperTree Planning (HTP) framework, which utilizes a hyper-tree structure to achieve multi-layered divide-and-conquer hierarchical thinking, enhancing the reasoning and generation capabilities of LLMs in complex planning tasks.
Hypo3D: Exploring Hypothetical Reasoning in 3D
Ye Mao (Imperial College London), Krystian Mikolajczyk (Imperial College London)
RecognitionGenerationData SynthesisTransformerLarge Language ModelVision Language ModelVision-Language-Action ModelImageMultimodalityPoint Cloud
🎯 What it does: Proposes the Hypo3D task and dataset to evaluate the model's 3D hypothesis reasoning ability in the absence of real-time scene information.
Hypothesis Testing for Generalized Thurstone Models
Anuran Makur (Purdue University), Japneet Singh (Purdue University)
GraphTabular
🎯 What it does: A hypothesis testing framework for minimizing risk is proposed to determine whether a set of paired comparison data conforms to the generalized Thurstone model (GTM) given a selection function F.
I Think, Therefore I Diffuse: Enabling Multimodal In-Context Reasoning in Diffusion Models
Zhenxing Mi (Hong Kong University of Science and Technology), Dan Xu (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: This paper presents ThinkDiff, a framework that aligns visual language models (VLM) to diffusion model decoders through visual-language training, enabling the diffusion model to possess multimodal contextual reasoning capabilities.
IBCircuit: Towards Holistic Circuit Discovery with Information Bottleneck
Tian Bian (Chinese University of Hong Kong), Jia Li (Hong Kong University of Science and Technology)
OptimizationTransformerText
🎯 What it does: An end-to-end information bottleneck-based circuit discovery framework, IBCircuit, is proposed for the automatic identification of the minimal and most important computational subgraphs in Transformer models.
ICLShield: Exploring and Mitigating In-Context Learning Backdoor Attacks
Zhiyao Ren (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: Proposed a dual learning hypothesis for ICL backdoor attacks and designed the ICLShield defense method.
Identifiable Object Representations under Spatial Ambiguities
Avinash Kori (Imperial College London), Ben Glocker (Imperial College London)
Object DetectionRepresentation LearningConvolutional Neural NetworkTransformerAuto EncoderImage
🎯 What it does: A multi-view probabilistic slot attention mechanism (VISA) is proposed, which can learn recognizable object center representations in the presence of occlusion and viewpoint ambiguity.
Identification of Latent Confounders via Investigating the Tensor Ranks of the Nonlinear Observations
Zhengming Chen (Shantou University), Kun Zhang (Carnegie Mellon University)
Tabular
🎯 What it does: This study investigates how to identify discrete latent variables and their causal structure in mixed (discrete/continuous) observational data through tensor rank conditions in nonlinear causal models.
Identifying and Understanding Cross-Class Features in Adversarial Training
Zeming Wei (Peking University), Yisen Wang (Peking University)
Knowledge DistillationAdversarial AttackConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: This study investigates the learning dynamics of cross-class features (shared across multiple classes) during adversarial training and reveals their role in enhancing robustness and robust overfitting.
Identifying biological perturbation targets through causal differential networks
Menghua Wu (Massachusetts Institute of Technology), Tommi Jaakkola (Massachusetts Institute of Technology)
Drug DiscoveryGraph Neural NetworkBiomedical Data
🎯 What it does: This paper proposes the Causal Differential Networks (CDN) model for identifying intervention targets in gene regulatory networks from observational and interventional data.
Identifying Causal Direction via Variational Bayesian Compression
Quang-Duy Tran (Deakin University), Thin Nguyen (Deakin University)
CompressionOptimization
🎯 What it does: This study proposes a causal direction identification method based on variational Bayesian compression (COMIC), which utilizes a neural network model to approximate conditional distributions and estimates the model's Kolmogorov complexity through variational Bayesian coding, thereby determining causal direction from observational data without experimental intervention.
Identifying Metric Structures of Deep Latent Variable Models
Stas Syrota (Technical University of Denmark), Søren Hauberg (Technical University of Denmark)
GenerationData SynthesisImage
🎯 What it does: This paper proposes using Riemannian geometry to identify the metric structure (such as distance, angle, volume) between latent variables in deep latent variable models, rather than the coordinates of individual latent variables, and proves that these metrics can be statistically identified under mild assumptions.
Identifying Neural Dynamics Using Interventional State Space Models
Amin Nejatbakhsh (Flatiron Institute), Yixin Wang (University of Michigan)
Recurrent Neural NetworkTime SeriesSequential
🎯 What it does: This paper proposes a causal intervention-based state space model (iSSM) for inferring neural dynamics and making predictions under intervention conditions.
Idiosyncrasies in Large Language Models
Mingjie Sun (Carnegie Mellon University), Zhuang Liu (Princeton University)
ClassificationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates and quantifies the unique features (self-referentiality) in the outputs of large language models (LLMs) and verifies their distinguishability by classifying texts generated by different LLMs.
iDPA: Instance Decoupled Prompt Attention for Incremental Medical Object Detection
Huahui Yi (West China Hospital Sichuan University), Qicheng Lao (Beijing University of Posts and Telecommunications)
Object DetectionTransformerPrompt EngineeringVision Language ModelMultimodalityBiomedical Data
🎯 What it does: This paper proposes an incremental medical object detection framework named iDPA, aimed at addressing the difficulties in localization and recognition caused by the coupling of foreground and background information as well as prompt-attention coupling.
IL-SOAR : Imitation Learning with Soft Optimistic Actor cRitic
Stefano Viel (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
OptimizationRobotic IntelligenceReinforcement LearningTabular
🎯 What it does: The SOAR framework is proposed, which generates optimistic value estimates through a multi-Critic ensemble in simulation learning, combined with SAC for policy and cost updates, thereby achieving more efficient exploration and learning.
Imagine While Reasoning in Space: Multimodal Visualization-of-Thought
Chengzu Li (University of Cambridge), Furu Wei (Microsoft Research)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: Proposes a multi-modal visual thinking (MVoT) reasoning paradigm that enables large multi-modal language models to generate both text and image thinking trajectories during the reasoning process;
Imitation Learning from a Single Temporally Misaligned Video
William Huey (Cornell University), Sanjiban Choudhury (Cornell University)
Robotic IntelligenceReinforcement LearningVideo
🎯 What it does: A framework for imitation learning named ORCA is proposed, which generates dense rewards using a single temporally mismatched video, encouraging the agent to complete tasks in the order of sub-goals.
IMPACT: Iterative Mask-based Parallel Decoding for Text-to-Audio Generation with Diffusion Modeling
Kuan-Po Huang (National Taiwan University), Chao Wang (Amazon AGI)
GenerationData SynthesisTransformerDiffusion modelTextAudio
🎯 What it does: An iterative mask parallel decoding framework (IMPACT) based on continuous latent space has been developed for text-to-audio generation.
Implicit Bias of Gradient Descent for Non-Homogeneous Deep Networks
Yuhang Cai (University of California), Peter Bartlett
Optimization
🎯 What it does: This paper studies the implicit bias of gradient descent (GD) on non-uniform deep networks under exponential loss, particularly by establishing its asymptotic implicit bias through the analysis of three key characteristics of GD iterations.
Implicit degree bias in the link prediction task
Rachith Aiyappa (Indiana University), Sadamori Kojaku (Binghamton University)
Recommendation SystemGraph Neural NetworkGraphBenchmark
🎯 What it does: This paper studies the 'implicit degree bias' present in traditional link prediction benchmarks and demonstrates that this bias leads to evaluation results favoring methods that only utilize node degrees. To eliminate this issue, the authors propose a degree-corrected link prediction benchmark and show through extensive experiments that this benchmark better reflects the performance of recommendation tasks and effectively reduces overfitting to degree when training graph neural networks, thereby improving the performance of downstream tasks such as community detection.
Implicit Language Models are RNNs: Balancing Parallelization and Expressivity
Mark Schöne, Jannes Gladrow (Microsoft Research)
Recurrent Neural NetworkTransformerLarge Language ModelTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes an implicit state space model (implicit SSM) and an implicit transformer, achieving comparable expressiveness to RNNs through fixed-point iteration while maintaining parallelization during training and providing adaptive depth during inference; its superiority is validated in large-scale language modeling tasks.
Implicit Regularization for Tubal Tensor Factorizations via Gradient Descent
Santhosh Karnik (Northeastern University), Felix Krahmer (Technical University of Munich)
Optimization
🎯 What it does: This study investigates the implicit regularization phenomenon of gradient descent in over-parameterized tubal tensor factorization and proves that small random initialization can lead to convergence to low tubal rank solutions.
Implicit Riemannian Optimism with Applications to Min-Max Problems
Christophe Roux (Zuse Institute Berlin), Sebastian Pokutta (Zuse Institute Berlin)
Optimization
🎯 What it does: This paper proposes an implicit optimistic online learning algorithm RIOD implemented on Hadamard manifolds, and further extends it to RIODA for solving g-convex g-concave smooth min-max problems. Both algorithms achieve regret bounds and gradient complexities comparable to those in Euclidean space while satisfying manifold constraints.
Implicit Subgraph Neural Network
Yongjian Zhong (University of Iowa), Bijaya Adhikari (University of Iowa)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: This paper studies subgraph representation learning and proposes the Implicit Subgraph Neural Network (ISNN) for subgraph classification.
Importance Corrected Neural JKO Sampling
Johannes Hertrich (University of Paris Dauphine), Robert Gruhlke (Free University of Berlin)
Flow-based ModelMultimodalityStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A novel sampling method combining Continuous Normalizing Flows (CNFs) with importance-weighted rejection resampling—Importance Corrected Neural JKO Sampling—is proposed to generate independent samples from target distributions with unknown normalization constants and to evaluate the density of the generated distributions.