NeurIPS 2025 Papers — Page 21
Conference on Neural Information Processing Systems · 5275 papers
GUARDIAN: Safeguarding LLM Multi-Agent Collaborations with Temporal Graph Modeling
Jialong Zhou (King's College London), Xiao Yang (Tsinghua University)
Anomaly DetectionGraph Neural NetworkTransformerLarge Language ModelTextGraph
🎯 What it does: This paper proposes a framework called GUARDIAN, designed to detect and mitigate the issues of hallucination amplification and error injection and propagation in multi-agent collaboration of large language models.
GuardReasoner-VL: Safeguarding VLMs via Reinforced Reasoning
Yue Liu (National University of Singapore), Bryan Hooi (National University of Singapore)
Safty and PrivacyTransformerReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: A multimodal reasoning safety guardian model, GuardReasoner-VL, has been constructed, which can perform reasoning before determining whether the input and output are harmful.
GUI Exploration Lab: Enhancing Screen Navigation in Agents via Multi-Turn Reinforcement Learning
Haolong Yan (Beijing University of Posts and Telecommunications), Daxin Jiang (StepFun)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision Language ModelImage
🎯 What it does: The GE-Lab simulation environment is proposed for training and evaluating GUI agents' screen navigation, systematically exploring three training strategies: supervised fine-tuning, single-turn reinforcement learning (ST-RL), and multi-turn reinforcement learning (MT-RL).
GUI-Actor: Coordinate-Free Visual Grounding for GUI Agents
Qianhui Wu (Microsoft), Jianfeng Gao (Microsoft)
TransformerVision Language ModelImage
🎯 What it does: A coordinate-free visual grounding framework called GUI-Actor is proposed, which utilizes a dedicated <ACTOR> token to directly locate GUI elements through an attention mechanism.
GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents
Yuqi Zhou (Renmin University of China), Jun Xu (Huawei)
Object DetectionTransformerLarge Language ModelReinforcement LearningMultimodality
🎯 What it does: This paper studies and improves the R1-Zero-like training framework for graphical user interface (GUI) visual localization tasks. By systematically analyzing the three core components: input templates, reward functions, and policy updates, we propose the Fast Thinking template, Box Size constraint reward, and the GRPO improvement method that removes length bias and incorporates difficulty weighting. Ultimately, we train GUI-G1-3B on 17K public samples.
GUI-Reflection: Empowering Multimodal GUI Models with Self-Reflection Behavior
Penghao Wu, Ziwei Liu (Nanyang Technological University)
Large Language ModelSupervised Fine-TuningReinforcement LearningMultimodality
🎯 What it does: A GUI-Reflection framework is proposed, endowing end-to-end multimodal GUI models with self-reflection and error correction capabilities, covering three stages: pre-training, offline SFT, and online fine-tuning.
GUI-Rise: Structured Reasoning and History Summarization for GUI Navigation
Tao Liu (ShanghaiTech University), Song Bai
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: A GUI navigation framework (GUI-Rise) that combines structured reasoning, action prediction, and historical summarization has been developed, enhancing multi-step interaction capabilities based on two-stage training (first SFT then RL).
Guided Diffusion Sampling on Function Spaces with Applications to PDEs
Jiachen Yao (California Institute of Technology), Anima Anandkumar (California Institute of Technology)
RestorationGenerationDiffusion modelTime SeriesPhysics Related
🎯 What it does: A discretization-invariant function space diffusion model called FunDPS is proposed to recover the posterior distribution of PDE solutions from extremely sparse or noisy measurements.
GUIDED: Granular Understanding via Identification, Detection, and Discrimination for Fine-Grained Open-Vocabulary Object Detection
Jiaming Li (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)
Object DetectionTransformerLarge Language ModelVision Language ModelImage
🎯 What it does: Proposed the GUIDED framework, which splits fine-grained open vocabulary object detection into coarse-grained localization and attribute discrimination, and utilizes LLM to parse labels, attention fusion of attributes, and projection correction of VLM;
GuideFlow3D: Optimization-Guided Rectified Flow For Appearance Transfer
Sayan Deb Sarkar (Stanford University), Iro Armeni (Stanford University)
Image TranslationOptimizationRectified FlowMesh
🎯 What it does: A training-independent 3D visual style transfer framework called GuideFlow3D is proposed, which achieves the transfer of texture and detail from an input 3D mesh to a target appearance by applying differentiable guidance on a pre-trained Rectified Flow model.
Guiding Cross-Modal Representations with MLLM Priors via Preference Alignment
Pengfei Zhao (Apple), Sifeng He (Apple)
RetrievalTransformerLarge Language ModelContrastive LearningImageMultimodality
🎯 What it does: Proposes the MAPLE framework, which automatically constructs preference data using the preference alignment prior of MLLM and trains cross-modal embeddings through RPA loss to address the modality gap issue in retrieval tasks.
Guiding LLM Decision-Making with Fairness Reward Models
Zara Hall (Columbia University), Richard Zemel (Columbia University)
ClassificationRecommendation SystemTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: A general Fairness Reward Model (FRM) is constructed and trained to score the fairness of each step in the chain of thought (CoT) of large language models (LLMs) during the inference phase, thereby emphasizing fair reasoning paths in final decisions and improving the fairness of high-risk decisions (such as judicial risk assessment, social media content moderation, and job screening) without compromising accuracy.
GVPO: Group Variance Policy Optimization for Large Language Model Post-Training
Kaichen Zhang (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
OptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes the Group Variance Policy Optimization (GVPO) method for post-training of large language models;
GyroSwin: 5D Surrogates for Gyrokinetic Plasma Turbulence Simulations
Fabian Paischer (Johannes Kepler University Linz), Johannes Brandstetter (Johannes Kepler University Linz)
TransformerTime SeriesPhysics Related
🎯 What it does: This paper proposes GyroSwin, a neural surrogate model based on the 5D Swin Transformer, for efficiently simulating nonlinear 5D turbulence in gyrokinetics, directly predicting distribution functions, potential fields, and heat flux.
H-SPLID: HSIC-based Saliency Preserving Latent Information Decomposition
Lukas Miklautz (Max Planck Institute of Biochemistry), Stratis Ioannidis (Northeastern University)
ClassificationRecognitionSegmentationAdversarial AttackContrastive LearningImage
🎯 What it does: The H-SPLID algorithm is proposed, explicitly dividing the latent space into significant and non-significant subspaces to learn task-relevant features and compress redundant information.
H3D-DGS: Exploring Heterogeneous 3D Motion Representation for Deformable 3D Gaussian Splatting
Bing He (Shanghai Jiao Tong University), Wenjun Zhang (Shanghai Jiao Tong University)
RestorationSegmentationOptimizationComputational EfficiencyGaussian SplattingOptical FlowVideo
🎯 What it does: The H3D-DGS method is proposed, which decomposes and models scene motion using heterogeneous 3D control points, achieving efficient real-time dynamic 3D reconstruction.
Hadamard Test is Sufficient for Efficient Quantum Gradient Estimation with Lie Algebraic Symmetries
Mohsen Heidari (Indiana University), Wojciech Szpankowski (Purdue University)
OptimizationComputational Efficiency
🎯 What it does: A gradient estimation framework based on Hadamard testing and Lie algebra structure is proposed for the gradient estimation problem of parameterized quantum circuits (PQC).
Hadamax Encoding: Elevating Performance in Model-Free Atari
Jacob Eeuwe Kooi, Vincent Francois-Lavet
Convolutional Neural NetworkReinforcement LearningImage
🎯 What it does: A new pixel input encoder called Hadamax encoder is proposed to enhance the performance of model-free reinforcement learning.
HAIF-GS: Hierarchical and Induced Flow-Guided Gaussian Splatting for Dynamic Scene
Jianing Chen (Institute of Computing Technology Chinese Academy of Sciences), Yucheng Zhang (Institute of Computing Technology Chinese Academy of Sciences)
RestorationGenerationGaussian SplattingOptical FlowVideo
🎯 What it does: This paper proposes a Hierarchical Anchor-based Flow-guided Gaussian Profile (HAIF-GS) framework for real-time and consistent dynamic 3D reconstruction from monocular videos.
HairFree: Compositional 2D Head Prior for Text-Driven 360° Bald Texture Synthesis
Mirela Ostrek (Max Planck Institute for Intelligent Systems), Justus Thies (Max Planck Institute for Intelligent Systems)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Generate high-quality 360° bald head textures through unsupervised text-driven 2D diffusion priors and FLAME geometry.
Hallucination at a Glance: Controlled Visual Edits and Fine-Grained Multimodal Learning
Tianyi Bai (Hong Kong University of Science and Technology), Binhang Yuan (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerSupervised Fine-TuningContrastive LearningImageMultimodalityBenchmark
🎯 What it does: This paper proposes a generation and annotation pipeline for fine-grained visual differences, and constructs the Micro Edit Dataset (MED) along with corresponding evaluation benchmarks.
HALO: Hadamard-Assisted Lower-Precision Optimization for LLMs
Saleh Ashkboos (ETH Zurich), Dan Alistarh (ISTAustria)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A low-precision quantization training method named HALO has been developed, utilizing Hadamard rotations to eliminate outliers during forward and backward propagation, achieving low-precision fine-tuning of LLMs.
Hamiltonian Descent Algorithms for Optimization: Accelerated Rates via Randomized Integration Time
Qiang Fu (Yale University), Andre Wibisono (Yale University)
OptimizationTabularStochastic Differential Equation
🎯 What it does: This paper introduces the Random Hamiltonian Flow (RHF) and its discretized version, the Random Hamiltonian Gradient Descent (RHGD), and proves that it can achieve accelerated convergence under both strongly convex and weakly convex objectives.
Hamiltonian Neural PDE Solvers through Functional Approximation
Anthony Zhou (Carnegie Mellon University), Amir Barati Farimani (Carnegie Mellon University)
Time SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A PDE solver based on the Hamiltonian framework is proposed—Hamiltonian Neural Solver (HNS), which approximates the Hamiltonian functional through a learnable Integral Kernel Functional (IKF) and uses automatic differentiation to obtain functional derivatives for predicting the time evolution of infinite-dimensional systems.
Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization
Kuan Zhang (Beijing Institute of Technology), Lei Cao (University of Arizona)
ClassificationOptimizationConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: The IDO framework is proposed, which achieves instance-level difficulty modeling and optimization for noisy label learning through two-stage training and dynamic weighted loss.
Handling Missing Responses under Cluster Dependence with Applications to Language Model Evaluation
Zhenghao Zeng (Stanford University), Edward Kennedy
Text
🎯 What it does: This paper studies how to use the Doubly Robust Estimator to estimate the overall mean in clustered data with missing responses, and provides its asymptotic normality and convergence rate under cluster dependence.
Hankel Singular Value Regularization for Highly Compressible State Space Models
Paul Schwerdtner (Courant Institute of Mathematical Sciences New York University), Benjamin Peherstorfer (Courant Institute of Mathematical Sciences New York University)
CompressionSequentialBenchmark
🎯 What it does: This paper proposes a method for regularizing state space models (Hankel structure) during the training process, allowing the model to be efficiently compressed while maintaining high accuracy.
HAODiff: Human-Aware One-Step Diffusion via Dual-Prompt Guidance
Jue Gong (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
RestorationGenerationTransformerPrompt EngineeringDiffusion modelImage
🎯 What it does: This paper presents HAODiff, a single-step diffusion model for portrait images that can achieve high-quality recovery in the presence of both global noise and human motion blur.
Hardware-aligned Hierarchical Sparse Attention for Efficient Long-term Memory Access
Xiang Hu (Ant Group), Wei Wu (Ant Group)
RetrievalComputational EfficiencyRecurrent Neural NetworkText
🎯 What it does: The Hierarchical Sparse Attention (HSA) mechanism is proposed, and based on this, the RAMba model is constructed, integrating RNN backbone, sparse attention, and memory reset mechanism to achieve efficient random access and length generalization for long contexts.
Harmony in Divergence: Towards Fast, Accurate, and Memory-efficient Zeroth-order LLM Fine-tuning
Qitao Tan (University of Georgia), Geng Yuan (University of Georgia)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: An efficient LLM fine-tuning method based on zero-order optimization, DiZO, is proposed. By comparing the hierarchical update differences between FO and ZO, a hierarchical diversification projection mechanism is designed to achieve learning effects similar to FO while significantly reducing memory usage.
Harnessing Feature Resonance under Arbitrary Target Alignment for Out-of-Distribution Node Detection
Shenzhi Yang (Zhejiang University), Haobo Wang (Zhejiang University)
Anomaly DetectionGraph Neural NetworkGraph
🎯 What it does: This paper proposes an unsupervised, label-free, and pre-training-free method for out-of-distribution (OOD) detection of graph nodes, called RSL. It aligns the features of known in-distribution (ID) nodes to random targets and utilizes the differences in 'feature resonance' between ID nodes and unknown ID/OOD nodes in the single-step gradient direction to filter reliable OOD candidate nodes. It also uses Stochastic Gradient Langevin Dynamics (SGLD) to synthesize more realistic out-of-vocabulary (OOV) samples for training a binary classifier.
Harnessing the Computation Redundancy in ViTs to Boost Adversarial Transferability
Jiani Liu (University of Rochester), Chenliang Xu (University of Rochester)
Adversarial AttackTransformerMixture of ExpertsVision Language ModelImage
🎯 What it does: This paper conducts an in-depth study of the computational redundancy in the data layer and model layer of Vision Transformer, proposing various technical means to significantly enhance the transferability of adversarial attacks, and constructs a complete redundancy-driven attack framework.
Harnessing the Universal Geometry of Embeddings
Rishi Dev Jha (Cornell University), John Xavier Morris (Cornell University)
Domain AdaptationRepresentation LearningGenerative Adversarial NetworkContrastive LearningTextMultimodality
🎯 What it does: An unsupervised method named vec2vec is proposed, which can map vectors from one text embedding space to another without paired samples, encoders, or predefined matching sets, while preserving the semantic geometric structure.
Hawaii: Hierarchical Visual Knowledge Transfer for Efficient Vision-Language Models
Yimu Wang (University of Waterloo), Krzysztof Czarnecki (University of Waterloo)
Knowledge DistillationRepresentation LearningTransformerMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: Proposes the HAWAII framework, which distills the knowledge of multiple visual experts into a single visual encoder to enhance the visual understanding capabilities of VLM;
Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation
Zhi-Kai Chen (Nanjing University), De-Chuan Zhan (Nanjing University)
GenerationComputational EfficiencyTransformerSupervised Fine-TuningImage
🎯 What it does: A framework named Hawk is proposed to enhance inference speed by using spatial speculative decoding during the autoregressive image generation process.
HBLLM: Wavelet-Enhanced High-Fidelity 1-Bit Quantization for LLMs
Ningning CHEN, Ying Jiang (Sun Yat-sen University)
CompressionTransformerLarge Language ModelText
🎯 What it does: This paper proposes HBLLM, a 1-bit post-training quantization framework based on Haar wavelet transform, aimed at compressing large language models (LLMs) while maintaining high inference accuracy.
HCRMP: An LLM-Hinted Contextual Reinforcement Learning Framework for Autonomous Driving
Zhiwen Chen (Tongji University), Bo Leng (Tongji University)
Autonomous DrivingLarge Language ModelReinforcement LearningRetrieval-Augmented Generation
🎯 What it does: A prompt-based reinforcement learning framework called HCRMP is proposed, which uses LLM to provide semantic prompts to assist RL in achieving autonomous driving motion planning.
Head Pursuit: Probing Attention Specialization in Multimodal Transformers
Lorenzo Basile (Area Science Park), Alberto Cazzaniga (Area Science Park)
ClassificationGenerationTransformerImageTextMultimodality
🎯 What it does: This paper studies the specialization of attention heads in multimodal Transformers, exploring the contribution of heads to specific semantic attributes through sparse decoding, and utilizing this information for fine control over model generation.
Heavy-Ball Momentum Method in Continuous Time and Discretization Error Analysis
Bochen Lyu (University of Southampton), Zhanxing Zhu (University of Southampton)
OptimizationImageOrdinary Differential Equation
🎯 What it does: This paper studies the approximation model of the Heavy-Ball momentum method in continuous time and provides a continuous dynamics with controllable error order.
HeavyWater and SimplexWater: Distortion-free LLM Watermarks for Low-Entropy Distributions
Dor Tsur, Flavio Calmon
OptimizationLarge Language ModelTextSequentialFinance Related
🎯 What it does: Two watermarking schemes for low-entropy text generation are designed: SimplexWater (based on binary scores and triangular codes) and HeavyWater (based on heavy-tailed distribution scores).
HEIR: Learning Graph-Based Motion Hierarchies
Cheng Zheng (Princeton University), Felix Heide (Princeton University)
Graph Neural NetworkTime Series
🎯 What it does: This study proposes a graph-based motion hierarchy learning framework that can automatically learn motion hierarchical structures from observational data, decomposing global motion into parent node inheritance patterns and local residuals, and applying this method to dynamic scenes in 1D, 2D, and 3D high-dimensional Gaussian fields.
HELM: Hyperbolic Large Language Models via Mixture-of-Curvature Experts
Neil He (Yale University), Rex Ying (Yale University)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: A series of large language models (HELM) trained entirely in hyperbolic space is proposed, achieving this for the first time at a scale of one billion parameters;
Hephaestus: Mixture Generative Modeling with Energy Guidance for Large-scale QoS Degradation
Nguyen Hoang Khoi Do (University of Florida), My T. Thai (University of Florida)
OptimizationGraph Neural NetworkReinforcement LearningGenerative Adversarial NetworkGraph
🎯 What it does: Proposes the Hephaestus framework to address the issue of network QoS degradation (QoSD), capable of finding the minimum budget perturbation scheme on large-scale graphs under nonlinear edge weight functions.
HermesFlow: Seamlessly Closing the Gap in Multimodal Understanding and Generation
Ling Yang (Princeton University), Bin CUI
GenerationOptimizationTransformerLarge Language ModelReinforcement LearningImageTextMultimodality
🎯 What it does: This paper proposes the HermesFlow framework, which optimizes data through self-generated comparative advantages and disadvantages using Pair-DPO, achieving simultaneous improvements in understanding and generation capabilities in multimodal large language models (MLLMs) while narrowing the performance gap between the two.
HeroFilter: Adaptive Spectral Graph Filter for Varying Heterophilic Relations
Shuaicheng Zhang, Dawei Zhou
ClassificationGraph Neural NetworkGraph
🎯 What it does: This paper proposes HEROFILTER, an adaptive spectral graph filter for node classification on different heterogeneous graphs.
Hessian-guided Perturbed Wasserstein Gradient Flows for Escaping Saddle Points
Naoya Yamamoto (University of Tokyo), Taiji Suzuki (University of Tokyo)
OptimizationStochastic Differential Equation
🎯 What it does: A new algorithm for non-convex optimization in probability measure spaces is proposed—Perturbed Wasserstein Gradient Flow (PWGF) to escape saddle points and achieve second-order optimality;
Heterogeneous Adversarial Play in Interactive Environments
Manjie Xu (Peking University), Yixin Zhu (Peking University)
OptimizationReinforcement LearningGenerative Adversarial NetworkImageText
🎯 What it does: The Heterogeneous Adversarial Play (HAP) framework is proposed, which automatically generates dynamic curricula through adversarial learning in a zero-sum game format between the teacher and the student, thereby enhancing learning efficiency in multi-task environments.
Heterogeneous Diffusion Structure Inference for Network Cascade
Siyu Huang (Pennsylvania State University), Yubai Yuan (Pennsylvania State University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A dual-network hybrid model is proposed to infer hidden diffusion networks, addressing the network identification problem in heterogeneous cascade processes.
Heterogeneous Graph Transformers for Simultaneous Mobile Multi-Robot Task Allocation and Scheduling under Temporal Constraints
Batuhan Altundas (Georgia Institute of Technology), Matthew Craig Gombolay
OptimizationRobotic IntelligenceGraph Neural NetworkTransformerReinforcement LearningGraph
🎯 What it does: This paper studies a multi-agent task allocation and scheduling framework called TARGETNET based on heterogeneous graph Transformer, which can generate task allocation and scheduling plans in one go.
Heterogeneous Swarms: Jointly Optimizing Model Roles and Weights for Multi-LLM Systems
Shangbin Feng (University of Washington), Tomas Pfister (Google)
OptimizationLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Designed the HETEROGENEOUS SWARMS algorithm to jointly optimize the model roles (DAG structure) and weights of multiple LLM systems using particle swarm optimization;
HetSyn: Versatile Timescale Integration in Spiking Neural Networks via Heterogeneous Synapses
Zhichao Deng (Tianjin University), Qiang Yu (Tianjin University)
Spiking Neural NetworkTime SeriesAudio
🎯 What it does: Designed and implemented the HetSyn framework, introducing adjustable time constants at the synaptic level to achieve multi-time scale integration, and validated its effectiveness on multiple tasks using the HetSynLIF model.
HIDISC: A Hyperbolic Framework for Domain Generalization with Generalized Category Discovery
Vaibhav Rathore (Indian Institute of Technology Bombay), Biplab Banerjee (Indian Institute of Technology Bombay)
Domain AdaptationDiffusion modelContrastive LearningImage
🎯 What it does: A domain and category generalization framework called HIDISC based on hyperbolic space is proposed, achieving the DG-GCD task without accessing target domain data.
Hierachical Balance Packing: Towards Efficient Supervised Fine-tuning for Long-Context LLM
Yongqiang Yao (Shanghai Jiao Tong University), Ningyi Xu (Shanghai Jiao Tong University)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes Hierarchical Balance Packing (HBP), which addresses the workload imbalance issue in long-context LLM training through multi-layer data packing and dynamic training pipelines.
Hierarchical Demonstration Order Optimization for Many-shot In-Context Learning
Yinhan He (University of Virginia), Jundong Li (University of Virginia)
OptimizationLarge Language ModelText
🎯 What it does: This paper studies the issue of demonstration order instability in many-shot in-context learning (ICL), proposing an information-theoretic measure called ICD-OVI and a hierarchical optimization framework (HIDO) that can efficiently search within a large-scale demonstration arrangement space.
Hierarchical Fine-grained Preference Optimization for Physically Plausible Video Generation
Harold Haodong Chen (Hong Kong University of Science and Technology), Ser-Nam Lim (Everlyn AI)
GenerationOptimizationLarge Language ModelDiffusion modelOptical FlowVideoText
🎯 What it does: The PhysHPO framework is proposed to achieve multi-level (instance, state, motion, semantic) video preference optimization to enhance the physical feasibility of video generation.
Hierarchical Frequency Tagging Probe (HFTP): A Unified Approach to Investigate Syntactic Structure Representations in Large Language Models and the Human Brain
Jingmin An (Peking University), Fang Fang (Peking University)
Large Language ModelTextAudio
🎯 What it does: Proposed and implemented the Hierarchical Frequency Tagging Probe (HFTP) to detect the hierarchical structural representations of sentences and phrases in large language models (LLMs) and the human brain, and aligned them across various LLMs and human brain data.
Hierarchical Implicit Neural Emulators
Ruoxi Jiang (Fudan University), Rebecca Willett (University of Chicago)
Time SeriesPhysics Related
🎯 What it does: A multi-scale implicit neural simulator is proposed, which significantly improves the stability and accuracy of long-term predictions by using multi-layer low-dimensional future state representations during prediction.
Hierarchical Information Aggregation for Incomplete Multimodal Alzheimer's Disease Diagnosis
Chengliang Liu (Shenzhen University), Xiaoling Luo (Shenzhen University)
ClassificationRecognitionMixture of ExpertsMultimodalityBiomedical DataMagnetic Resonance ImagingPositron Emission TomographyAlzheimer's Disease
🎯 What it does: The HAD framework is proposed, which can perform multimodal diagnosis of Alzheimer's disease under any missing modality and efficiently captures long-range dependencies in three-dimensional images through multi-view Hilbert curve Mamba blocks and hierarchical spatial feature extraction modules.
Hierarchical Koopman Diffusion: Fast Generation with Interpretable Diffusion Trajectory
Hanru Bai (Fudan University), Difan Zou (The University of Hong Kong)
GenerationExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationDiffusion modelImage
🎯 What it does: This paper proposes a hierarchical one-shot image generation framework HKD based on Koopman operator theory, which maintains first-order sampling speed while preserving the interpretability and controllability of the generated trajectories.
Hierarchical Optimization via LLM-Guided Objective Evolution for Mobility-on-Demand Systems
Yi Zhang (Agency for Science Technology and Research), Jun Liu (Lancaster University)
OptimizationTransformerLarge Language ModelPrompt EngineeringTabular
🎯 What it does: A hierarchical framework that combines large language models (LLM) with mathematical optimizers is proposed for real-time dynamic ride-hailing scheduling, where the LLM is responsible for generating high-level objectives and the optimizer is responsible for low-level route planning.
Hierarchical Retrieval: The Geometry and a Pretrain-Finetune Recipe
Chong You (Google), Sanjiv Kumar (Google)
RetrievalSupervised Fine-TuningText
🎯 What it does: This paper studies and verifies the feasibility of Dual Encoder in the Hierarchical Retrieval task, demonstrating the existence of embeddings that meet hierarchical retrieval requirements in Euclidean space and achieving this goal through learning.
Hierarchical Self-Attention: Generalizing Neural Attention Mechanics to Multi-Scale Problems
Saeed Amizadeh (Microsoft), Kazuhito Koishida (Microsoft)
ClassificationOptimizationTransformerSupervised Fine-TuningTextMultimodality
🎯 What it does: A hierarchical self-attention (HSA) is proposed, extending the Softmax attention of the Transformer to multi-scale, multi-modal, and hierarchical data.
Hierarchical Semantic-Augmented Navigation: Optimal Transport and Graph-Driven Reasoning for Vision-Language Navigation
Xiang Fang (Nanyang Technological University), Changshuo Wang (University College London)
Robotic IntelligenceGraph Neural NetworkReinforcement LearningVision Language ModelMultimodalityGraph
🎯 What it does: A Hierarchical Semantic-Augmented Navigation (HSAN) framework is proposed for performing visual language navigation tasks in continuous environments.
Hierarchical Shortest-Path Graph Kernel Network
Jiaxin Wang (Hainan University), Jieren Cheng (Hainan University)
OptimizationRepresentation LearningHyperparameter SearchGraph Neural NetworkGraph
🎯 What it does: An end-to-end graph kernel network based on hierarchical shortest path graph kernels (HSP-GKN) is proposed, combining graph kernels with neural networks to achieve task-related graph representation learning.
HiFC: High-efficiency Flash-based KV Cache Swapping for Scaling LLM Inference
Inho Jeong (Seoul National University), Dongsuk Jeon (Seoul National University)
TransformerLarge Language ModelText
🎯 What it does: A DRAM-free KV cache swapping framework named HiFC has been designed and implemented, which achieves direct, low-latency KV cache read and write between GPU and flash memory by utilizing the pseudo-SLC (pSLC) area of NVMe SSDs and GPU Direct Storage (GDS), supporting memory expansion for long-context LLM inference.
HiFlow: Training-free High-Resolution Image Generation with Flow-Aligned Guidance
Jiazi Bu (Shanghai Jiao Tong University), Jiaqi Wang (Shanghai AI Laboratory)
GenerationData SynthesisFlow-based ModelRectified FlowImage
🎯 What it does: HiFlow is proposed, a training-independent and model-independent framework that constructs a virtual reference flow using low-resolution sampled trajectories, and achieves high-resolution image generation through initialization, direction, and acceleration alignment.
High Dynamic Range Imaging with Time-Encoding Spike Camera
Zhenkun Zhu (Peking University), Tiejun Huang (Peking University)
RestorationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a Time Encoding (TE) burst camera that utilizes a clock cycle counter to record super-threshold moments, significantly enhancing the dynamic range of the burst camera, and designs a complete image reconstruction network for TE burst streams.
High Resolution UDF Meshing via Iterative Networks
Federico Stella (École Polytechnique Fédérale de Lausanne), Pascal Fua (École Polytechnique Fédérale de Lausanne)
GenerationOptimizationConvolutional Neural NetworkAuto EncoderPoint CloudMesh
🎯 What it does: An iterative network is proposed to grid the neural UDF to address the surface missing problem caused by noise at high resolutions.
High-Dimensional Calibration from Swap Regret
Maxwell Fishelson (Massachusetts Institute of Technology), Jon Schneider (Google Research)
Optimization
🎯 What it does: A general online multidimensional prediction calibration algorithm called TreeCal is proposed, along with its theoretical upper bound on calibration error.
High-dimensional neuronal activity from low-dimensional latent dynamics: a solvable model
Valentin Schmutz (University College London), Kenneth D. Harris (University College London)
Recurrent Neural NetworkTime Series
🎯 What it does: A parsable low-rank RNN model was constructed, proving that low-dimensional pre-activation dynamics can generate high-dimensional post-activation activities. The three-way relationship between post-activation spectrum, pre-activation dimensions, and activation functions was derived using random feature kernel theory. Subsequently, the Neural Cross-Encoder (NCE) model was proposed and applied to estimate the low-dimensional pre-activation dimensions from two-photon calcium imaging data of the mouse visual cortex.
High-order Equivariant Flow Matching for Density Functional Theory Hamiltonian Prediction
Seongsu Kim (KAIST), Sungsoo Ahn (KAIST)
GenerationComputational EfficiencyGraph Neural NetworkFlow-based ModelGraphPhysics RelatedOrdinary Differential Equation
🎯 What it does: We propose a high-order SE(3) symmetric flow matching framework called QHFLOW, which is used to predict the Kohn-Sham Hamiltonian matrix in density functional theory (DFT), generating it directly rather than through regression, significantly reducing the number of iterations required in the SCF cycle.
High-Order Flow Matching: Unified Framework and Sharp Statistical Rates
Maojiang Su (Northwestern University), Han Liu (Northwestern University)
GenerationTransformerImageVideoAudio
🎯 What it does: A unified theoretical framework for high-order flow matching is proposed, and its statistical rate is established.
High-order Interactions Modeling for Interpretable Multi-Agent Q-Learning
Qinyu Xu (Nanjing University), Chunlin Chen (Nanjing University)
Explainability and InterpretabilityReinforcement Learning
🎯 What it does: An interpretable high-order interaction modeling framework QCoFr is proposed, which implements value decomposition in multi-agent Q-learning using Continuous Fraction Networks (CFN) and enhances credit allocation by combining Variational Information Bottleneck (VIB);
High-Performance Arithmetic Circuit Optimization via Differentiable Architecture Search
Xilin Xia (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A differentiable architecture search framework ARITH-DAS is proposed, which directly performs fine-grained optimization of the interconnections of arithmetic circuits on multi-relation directed acyclic graphs.
Higher-Order Learning with Graph Neural Networks via Hypergraph Encodings
Raphaël Pellegrin (Independent Researcher), Melanie Weber (Harvard University)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: A hierarchical encoding method based on hypergraphs (such as Hodge-Laplacian, random walk, discrete curvature, and local degree) is proposed to inject high-order structural information into traditional graph neural networks, enhancing the performance of multi-relational learning.
Highlighting What Matters: Promptable Embeddings for Attribute-Focused Image Retrieval
Siting Li (University of Washington), Simon Shaolei Du (University of Washington)
RetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Construct the COCO-FACET benchmark dataset to evaluate attribute-focused text-image retrieval, and propose the use of promptable embeddings generated by multimodal large language models to enhance retrieval performance.
HiMoLE: Towards OOD-Robust LoRA via Hierarchical Mixture of Experts
Yinuo Jiang (Zhejiang University), Huajun Chen (Zhejiang University)
Domain AdaptationOptimizationTransformerSupervised Fine-TuningMixture of ExpertsTextBiomedical Data
🎯 What it does: In this paper, the authors propose the HiMoLE (Hierarchical Mixture of LoRA Experts) framework, which enhances the robustness of parameter-efficient fine-tuning methods in out-of-distribution (OOD) scenarios by combining LoRA with hierarchical expert modules and a hierarchical routing strategy.
HiPoNet: A Multi-View Simplicial Complex Network for High Dimensional Point-Cloud and Single-Cell data
Siddharth Viswanath (Yale University), Smita Krishnaswamy (Yale University)
ClassificationRepresentation LearningGraph Neural NetworkPoint CloudBiomedical Data
🎯 What it does: Designed and implemented HiPoNet, an end-to-end differentiable high-dimensional point cloud network that utilizes multi-view reweighted features, Vietoris-Rips simplicial complex construction, and simplicial wave-particle transforms for multi-scale feature extraction, applied to regression, classification, and representation learning.
Hippocampal-like Sequential Editing for Continual Knowledge Updates in Large Language Models
Quntian Fang (National University of Defense Technology), Guotong Geng
Large Language ModelSupervised Fine-TuningText
🎯 What it does: A hippocampus-inspired sequential model editing framework (HSE) has been designed and implemented to continuously update knowledge without retraining large language models (LLMs), addressing the issues of parameter drift and catastrophic forgetting.
HM3: Hierarchical Multi-Objective Model Merging for Pretrained Models
Yu Zhou (Hong Kong Polytechnic University), KC Tan
OptimizationTransformerLarge Language ModelReinforcement LearningImageText
🎯 What it does: This paper proposes and implements HM3—a hierarchical multi-objective model merging framework that can simultaneously search in the parameter space and architecture space to generate customizable high-performance merged models.
HMARL-CBF – Hierarchical Multi-Agent Reinforcement Learning with Control Barrier Functions for Safety-Critical Autonomous Systems
H M Sabbir Ahmad, Wenchao Li (Massachusetts Institute of Technology)
Autonomous DrivingSafty and PrivacyReinforcement LearningAgentic AI
🎯 What it does: A hierarchical reinforcement learning framework HMARL-CBF is proposed for multi-agent safety-critical systems, utilizing Control Barrier Functions (CBF) to achieve state-based safety constraints, and employing high-level policies to select shared skills while low-level policies safely execute these skills.
HMVLM:Human Motion-Vision-Language Model via MoE LoRA
Lei Hu (Institute of Computing Technology Chinese Academy of Sciences University of Chinese Academy of Sciences), Shihong Xia (Institute of Computing Technology Chinese Academy of Sciences University of Chinese Academy of Sciences)
GenerationPose EstimationTransformerLarge Language ModelMixture of ExpertsVision-Language-Action ModelVideoTextMultimodality
🎯 What it does: A unified framework HMVLM based on Mixture of Experts LoRA is proposed, which can support various human motion-related tasks such as text-to-action generation, human pose estimation, and action video understanding while retaining the knowledge of the base language model.
HOComp: Interaction-Aware Human-Object Composition
Dong Liang (Tongji University), Rynson W. H. Lau
GenerationData SynthesisPose EstimationTransformerLarge Language ModelDiffusion modelImageMultimodality
🎯 What it does: This paper proposes the HOComp framework, achieving natural interactive synthesis of human images and foreground objects.
Hogwild! Inference: Parallel LLM Generation via Concurrent Attention
Gleb Rodionov, Dan Alistarh
GenerationOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a parallel inference framework called Hogwild! Inference, which allows multiple LLM instances to synchronize in real-time and collaborate spontaneously to solve complex reasoning tasks through shared Key-Value caches.
HOI-Dyn: Learning Interaction Dynamics for Human-Object Motion Diffusion
Lin Wu (University of Glasgow), Jianglin Lan (University of Glasgow)
GenerationData SynthesisPose EstimationTransformerDiffusion modelVideo
🎯 What it does: The HOI-Dyn framework is proposed, which achieves realistic human-object interaction generation through a drive-response system and a lightweight Transformer interaction dynamics model.
HoliGS: Holistic Gaussian Splatting for Embodied View Synthesis
Xiaoyuan Wang (Carnegie Mellon University), Ming-Hsuan Yang (Google DeepMind)
GenerationData SynthesisPose EstimationGaussian SplattingVideo
🎯 What it does: We propose HoliGS, a deformable Gaussian splatting framework for Embodied View Synthesis (EVS) in monocular video;
Holistic Large-Scale Scene Reconstruction via Mixed Gaussian Splatting
Chuandong Liu (Wuhan University), Gui-Song Xia (Wuhan University)
GenerationOptimizationNeural Radiance FieldGaussian SplattingPoint Cloud
🎯 What it does: This paper presents MixGS, a global optimization framework for large-scale scene reconstruction that achieves high-quality novel view synthesis by mixing high-resolution and original 3D Gaussians.
Holistic Order Prediction in Natural Scenes
Pierre Musacchio (Seoul National University), Jaesik Park (Seoul National University)
Object DetectionSegmentationDepth EstimationTransformerImage
🎯 What it does: The paper proposes a network called InstaFormer, which can directly predict the occlusion and depth order matrix of all instances in a scene from RGB images in a single forward pass.
HoliTom: Holistic Token Merging for Fast Video Large Language Models
Kele Shao (Zhejiang University), Huan Wang (Westlake University)
CompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideo
🎯 What it does: A training-independent HoliTom framework is proposed, which significantly compresses the visual tokens of video LLMs through global spatiotemporal segmentation and dual token merging, enhancing inference efficiency.
HollowFlow: Efficient Sample Likelihood Evaluation using Hollow Message Passing
Johann Flemming Gloy (Chalmers University of Technology and University of Gothenburg), Simon Olsson (Chalmers University of Technology and University of Gothenburg)
GenerationComputational EfficiencyGraph Neural NetworkFlow-based ModelGraph
🎯 What it does: A continuous normalizing flow (CNF) model called HollowFlow is constructed using a new non-backtracking graph neural network (NoBGNN) and Hollow Message Passing (HoMP) for efficient evaluation of sample likelihood in high-dimensional Boltzmann generators;
HoloLLM: Multisensory Foundation Model for Language-Grounded Human Sensing and Reasoning
Chuhao Zhou (Nanyang Technological University), Jianfei Yang (Nanyang Technological University)
RecognitionTransformerLarge Language ModelVision Language ModelVideoMultimodality
🎯 What it does: This paper proposes HoloLLM, a model that integrates scarce multimodal perceptions (LiDAR, infrared, millimeter-wave radar, WiFi) into a multimodal large language model to achieve language-based intelligent perception and reasoning.
HoloScene: Simulation‑Ready Interactive 3D Worlds from a Single Video
Hongchi Xia (University of Illinois Urbana-Champaign), Shenlong Wang (University of Illinois Urbana-Champaign)
GenerationOptimizationGaussian SplattingVideo
🎯 What it does: Reconstructing complete, interactive, and simulative 3D digital twin scenes from a single video
Homogeneous Algorithms Can Reduce Competition in Personalized Pricing
Nathanael Jo, Manish Raghavan
TabularFinance Related
🎯 What it does: This study investigates the price coordination effect caused by algorithm homogeneity, constructs a game theory model to analyze the impact of algorithm correlation on consumer welfare and corporate profits, and validates it through simulation experiments.
Homogeneous Keys, Heterogeneous Values: Exploiting Local KV Cache Asymmetry for Long-Context LLMs
Wanyun Cui (Shanghai University of Finance and Economics), Mingwei Xu (Shanghai University of Finance and Economics)
CompressionOptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Designed and implemented a training-free KV cache compression framework AsymKV, which merges based on the local homogeneity of keys and maintains the attention output unchanged through lossless value compression.
HopaDIFF: Holistic-Partial Aware Fourier Conditioned Diffusion for Referring Human Action Segmentation in Multi-Person Scenarios
Kunyu Peng (Karlsruhe Institute of Technology), Rainer Stiefelhagen (Karlsruhe Institute of Technology)
Object DetectionSegmentationRecurrent Neural NetworkTransformerVision Language ModelDiffusion modelVideoTextMultimodality
🎯 What it does: Proposed and implemented the Referring Human Action Segmentation (RHAS) task, which can segment the actions of target individuals in multi-person videos based on textual descriptions.
HoPE: Hybrid of Position Embedding for Long Context Vision-Language Models
Haoran Li (Carnegie Mellon University), Ruiwen Xu (Xiaohongshu Inc.)
RetrievalTransformerVision Language ModelVideoMultimodality
🎯 What it does: A hybrid position encoding (HoPE) suitable for long-context visual-language models is proposed to enhance long video understanding and retrieval performance.
Horizon Reduction Makes RL Scalable
Seohong Park (University of California), Sergey Levine (University of California)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: In this paper, the authors conduct experiments on complex, long-horizon robotic tasks by constructing a large-scale (up to 1 billion transitions) offline reinforcement learning dataset. They systematically evaluate the scalability of existing offline RL algorithms and find that the 'horizon' is the main reason for the limited scalability of the algorithms. They then propose to enhance scalability by reducing the value and policy horizon (e.g., n-step returns, hierarchical policies) and design a simple scalable offline RL method called SHARSA based on this.
HoT-VI: Reparameterizable Variational Inference for Capturing Instance-Level High-Order Correlations
Junxi Xiao (Sun Yat-sen University), Zexin Yuan (Sun Yat-sen University)
Anomaly DetectionGraph Neural NetworkAuto EncoderGraphTime Series
🎯 What it does: This paper proposes HoT-VI, a reparameterized variational inference framework that utilizes k-order associations to model instance-level high-order correlations.
How Benchmark Prediction from Fewer Data Misses the Mark
Guanhua Zhang (Max Planck Institute for Intelligent Systems), Moritz Hardt (Max Planck Institute for Intelligent Systems)
Benchmark
🎯 What it does: This paper conducts large-scale experiments on 11 existing and newly proposed benchmark prediction methods across 19 different benchmarks, systematically evaluating their effectiveness in predicting complete benchmark performance when only a small number of samples are assessed.
How Classifier Features Transfer to Downstream: An Asymptotic Analysis in a Two-Layer Model
HEE BIN YOO (Seoul National University), Byoung-Tak Zhang (Seoul National University)
ClassificationRepresentation LearningImage
🎯 What it does: This study investigates the feature learning dynamics of a two-layer proportional scale network after single-step gradient descent and analyzes the clustering effects of these features in unseen classes.