ICLR 2025 Papers — Page 15
International Conference on Learning Representations · 3704 papers
Graph Neural Ricci Flow: Evolving Feature from a Curvature Perspective
Jialong Chen (Sun Yat-sen University), Zibin Zheng (Sun Yat-sen University)
Graph Neural NetworkGraphOrdinary Differential Equation
🎯 What it does: This paper studies the extension of discrete Riemann flow on attribute graphs (Attri-DRF) and proposes a novel continuous deep graph neural network - Graph Neural Ricci Flow (GNRF).
Graph Sparsification via Mixture of Graphs
Guibin Zhang (Tongji University), Shirui Pan (Griffith University)
OptimizationComputational EfficiencyGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: To address the computational bottleneck of large-scale graphs, a Mixture of Graphs (MoG) method is proposed, which can dynamically select the most suitable sparsification expert for each node and mix the sparse subgraphs generated by it on the Grassmann manifold to obtain high-quality sparse graphs, while improving the inference speed and model performance of GNNs.
Graph Transformers Dream of Electric Flow
Xiang Cheng (Duke University), Suvrit Sra (Technical University of Munich)
Graph Neural NetworkTransformerGraph
🎯 What it does: This paper demonstrates that by precisely configuring the weights of linear Transformers, they can solve classic graph algorithms on graph data, such as current flow, Laplacian inverse, square root inverse, heat kernel, and graph Laplacian eigenvectors.
Graph-based Document Structure Analysis
Yufan Chen (Karlsruhe Institute of Technology), Rainer Stiefelhagen (Karlsruhe Institute of Technology)
Graph Neural NetworkTransformerGraph
🎯 What it does: This paper proposes the Graph-structured Document Structure Analysis task (gDSA), constructs a large-scale GraphDoc dataset, and designs an end-to-end Document Relationship Graph Generator (DRGG) to achieve joint reasoning of document layout detection and spatial/logical relationships.
Graph-Guided Scene Reconstruction from Images with 3D Gaussian Splatting
Chong Cheng (Hong Kong University of Science and Technology), Hao Wang (Hong Kong University of Science and Technology)
Autonomous DrivingOptimizationGaussian SplattingImage
🎯 What it does: Developed the GraphGS framework, which utilizes camera graph-guided 3D Gaussian Splatting to quickly and accurately reconstruct large scenes from uncalibrated images.
GraphArena: Evaluating and Exploring Large Language Models on Graph Computation
Jianheng Tang (Hong Kong University of Science and Technology), Jia Li (Hong Kong University of Science and Technology)
Graph Neural NetworkTransformerLarge Language ModelPrompt EngineeringGraphBenchmark
🎯 What it does: This paper proposes and implements GraphArena, a benchmark tool for graph computation problems, covering real-world graphs, ten multidimensional tasks, and a rigorous path-level evaluation process.
GraphBridge: Towards Arbitrary Transfer Learning in GNNs
Li Ju (National University of Singapore), Xinchao Wang (National University of Singapore)
Domain AdaptationRepresentation LearningGraph Neural NetworkContrastive LearningPoint CloudGraph
🎯 What it does: Proposes the GraphBridge framework, which constructs a two-stage pre-training + fine-tuning process, utilizing learnable input-output bridging and side networks (GSST, GMST) to achieve GNN transfer learning for any task and any domain;
GraphEval: A Lightweight Graph-Based LLM Framework for Idea Evaluation
Tao Feng (University of Illinois at Urbana Champaign), Jiaxuan You (University of Illinois at Urbana Champaign)
Graph Neural NetworkLarge Language ModelPrompt EngineeringTextGraph
🎯 What it does: Developed the GraphEval framework, which first uses a small LLM to decompose research ideas into viewpoint nodes, then constructs a viewpoint graph through BERT similarity or LLM relation extraction, and subsequently predicts paper review results using label propagation (LP) or graph neural networks (GNN), while incorporating novelty/plagiarism detection mechanisms.
GraphRouter: A Graph-based Router for LLM Selections
Tao Feng (University of Illinois Urbana Champaign), Jiaxuan You (University of Illinois Urbana Champaign)
Recommendation SystemGraph Neural NetworkLarge Language ModelTextGraph
🎯 What it does: Designed and implemented a heterogeneous graph-based LLM router called GraphRouter, which recommends the optimal language model for user queries in multi-task scenarios.
GravMAD: Grounded Spatial Value Maps Guided Action Diffusion for Generalized 3D Manipulation
Yangtao Chen (Nanjing University), Yang Gao (Nanjing University)
Robotic IntelligenceReinforcement Learning from Human FeedbackVision Language ModelDiffusion modelContrastive LearningPoint CloudBenchmark
🎯 What it does: This paper presents GravMAD—a sub-goal and language-conditioned action diffusion framework for generalization and precise execution of 3D manipulation tasks.
GReaTer: Gradients Over Reasoning Makes Smaller Language Models Strong Prompt Optimizers
Sarkar Snigdha Sarathi Das (Pennsylvania State University), Rui Zhang (Salesforce Research)
OptimizationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: A method named GREATER is proposed, which uses small language models to optimize prompts through gradient optimization, without relying on large proprietary LLMs for feedback.
Greener GRASS: Enhancing GNNs with Encoding, Rewiring, and Attention
Tongzhou Liao (Carnegie Mellon University), Barnabas Poczos (Carnegie Mellon University)
Graph Neural NetworkGraph
🎯 What it does: A new GNN architecture called GRASS is proposed, which enhances the learning effectiveness of graph data by combining relative random walk encoding, random reconnection, and a specially designed additive attention mechanism.
GridMix: Exploring Spatial Modulation for Neural Fields in PDE Modeling
Honghui Wang (Tsinghua University), Gao Huang (Tsinghua University)
OptimizationMeta LearningAuto EncoderTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: The MARBLE framework is proposed, which enhances the accuracy and generalization ability of PDE solutions based on Implicit Neural Representations (INR) through GridMix and spatial domain augmentation.
gRNAde: Geometric Deep Learning for 3D RNA inverse design
Chaitanya K. Joshi (University of Cambridge), Pietro Lio
GenerationOptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes gRNAde, a geometric deep learning pipeline that uses 3D RNA scaffolds (which can include multiple conformational states) to reverse design nucleotide sequences and generates candidate sequences through autoregressive decoding.
Grokking at the Edge of Numerical Stability
Lucas Prieto (Imperial College London), Tolga Birdal (Imperial College London)
Tabular
🎯 What it does: This study investigates the causes of the grokking phenomenon in deep learning, proposing explanations of Softmax Collapse and Naïve Loss Minimization, and addresses the issues of delay and numerical instability through the StableMax activation function and ⊥Grad optimizer.
GROOT-2: Weakly Supervised Multimodal Instruction Following Agents
Shaofei Cai (Peking University), Yitao Liang (Peking University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerAuto EncoderVideoTextMultimodality
🎯 What it does: GROOT-2 has been developed, an agent capable of following instructions under multimodal inputs (video, text, rewards, etc.), utilizing weak supervision learning combined with latent variable models to achieve instruction alignment and behavior generation.
Grounding by Trying: LLMs with Reinforcement Learning-Enhanced Retrieval
Sheryl Hsu (Stanford University), Archit Sharma (Stanford University)
RetrievalTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Improving the quality of retrieval queries generated by LLM through reinforcement learning, thereby enhancing the credibility of multi-hop retrieval and answer generation.
Grounding Continuous Representations in Geometry: Equivariant Neural Fields
David Wessels, Erik J Bekkers
ClassificationSegmentationGenerationRepresentation LearningMeta LearningNeural Radiance FieldImagePoint Cloud
🎯 What it does: This paper proposes an Equivariant Neural Field (ENF) framework that combines continuous function representation with geometrically interpretable latent point clouds to achieve geometrically aligned reversible decoding.
Grounding Multimodal Large Language Model in GUI World
Weixian Lei (National University of Singapore), Mike Zheng Shou (National University of Singapore)
TransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes an end-to-end GUI location positioning framework, which includes an automated data collection engine, a lightweight GUI Grounding model (AGG), and combines it with a multimodal large language model (MLLM) to build a visual agent capable of executing complex GUI tasks across various platforms.
Grounding Video Models to Actions through Goal Conditioned Exploration
Yunhao Luo (Georgia Tech), Yilun Du (Brown)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelVideo
🎯 What it does: Using intermediate frames generated by a pre-trained large-scale video model as visual targets to guide robots in self-supervised learning of goal-conditioned continuous action policies in environments without action labels.
Group Distributionally Robust Dataset Distillation with Risk Minimization
Saeed Vahidian (Duke University), Yiran Chen (Duke University)
Data SynthesisOptimizationKnowledge DistillationImage
🎯 What it does: This paper proposes a distributionally robust data distillation algorithm that combines clustering and risk minimization, aimed at enhancing the generalization and robustness of synthetic data in low-sample subgroups.
Group Downsampling with Equivariant Anti-aliasing
Md Ashiqur Rahman (Purdue University), Raymond A. Yeh (Purdue University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: A framework for uniform subgroup downsampling on finite groups is proposed, along with the corresponding anti-aliasing operations.
Group Ligands Docking to Protein Pockets
Jiaqi Guan (University of Illinois), Jianzhu Ma (Tsinghua University)
Drug DiscoveryDiffusion modelBiomedical Data
🎯 What it does: The GROUPBIND framework is proposed, which utilizes multiple ligands of the same protein pocket for molecular docking to enhance the accuracy of single-ligand docking.
Group-robust Sample Reweighting for Subpopulation Shifts via Influence Functions
Rui Qiao (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
Domain AdaptationOptimizationText
🎯 What it does: This paper proposes a group label-based sample reweighting method called GSR, which utilizes a small number of group labels as the target set to iteratively optimize the sample weights of unlabeled data through influence functions, thereby enhancing the model's robustness to changes in subgroup distributions.
Growth Inhibitors for Suppressing Inappropriate Image Concepts in Diffusion Models
Die Chen (East China Normal University), Yaliang Li (Alibaba Group)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: A non-fine-tuning 'growth inhibitor' method has been developed to suppress inappropriate concepts in the image space of diffusion models, achieving the elimination of unsafe content, styles, and objects.
GS-CPR: Efficient Camera Pose Refinement via 3D Gaussian Splatting
Changkun Liu (Hong Kong University of Science and Technology), Tristan Braud
Pose EstimationComputational EfficiencyNeural Radiance FieldGaussian SplattingImage
🎯 What it does: This paper presents GS-CPR, a method for scene modeling using 3D Gaussian Splatting, which refines the rough camera pose in a one-shot manner by rendering synthesized images and matching them with query images through 2D–2D correspondence.
GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting
Junzhe Jiang (Fudan University), Li Zhang (Fudan University)
GenerationData SynthesisAutonomous DrivingGaussian SplattingPoint Cloud
🎯 What it does: A framework utilizing two-dimensional Gaussian pulses for panoramic Gaussian splitting (GS-LiDAR) is proposed to generate realistic and controllable LiDAR point clouds, supporting perspective synthesis of dynamic scenes.
GSBA$^K$: $top$-$K$ Geometric Score-based Black-box Attack
Md Farhamdur Reza (North Carolina State University), Huaiyu Dai (North Carolina State University)
Adversarial AttackConvolutional Neural NetworkScore-based ModelImage
🎯 What it does: This paper proposes a model-free attack method based on geometric decision boundaries, GSBA K, which can generate imperceptible adversarial examples in a top-K manner for single-label multi-class and multi-label learning tasks.
GSE: Group-wise Sparse and Explainable Adversarial Attacks
Shpresim Sadiku (Zuse Institute Berlin), Sebastian Pokutta (Zuse Institute Berlin)
OptimizationExplainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: A two-stage algorithm GSE is proposed to generate adversarial attacks that are sparse and interpretable based on pixel groups.
GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models
Seyed Iman Mirzadeh, Mehrdad Farajtabar (Apple)
Large Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Two new mathematical reasoning benchmarks, GSM-Symbolic and GSM-NoOp, are proposed to systematically evaluate the reasoning performance of various LLMs under different instances, numerical variations, and additional irrelevant information.
GTR: Improving Large 3D Reconstruction Models through Geometry and Texture Refinement
Peiye Zhuang, Hsin-Ying Lee
GenerationData SynthesisConvolutional Neural NetworkTransformerNeural Radiance FieldPoint CloudMesh
🎯 What it does: This study proposes a fast 3D reconstruction framework GTR based on multi-view images, capable of generating high-quality meshes and textures in seconds.
Guaranteed Generation from Large Language Models
Minbeom Kim (Seoul National University), Marc Dymetman (Independent Researcher)
GenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes the GUARD framework, which combines autoregressive approximation during training with rejection sampling during inference to achieve constrained generation (guaranteed generation) for large language models.
GUI-World: A Video Benchmark and Dataset for Multimodal GUI-oriented Understanding
Dongping Chen (Huazhong University of Science and Technology), Lichao Sun (Microsoft Research)
RecognitionGenerationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Created the GUI-WORLD large-scale video dataset and established a benchmark for evaluating and enhancing the capabilities of multimodal large models in dynamic multi-window GUI understanding tasks.
Guided Score identity Distillation for Data-Free One-Step Text-to-Image Generation
Mingyuan Zhou (University of Texas at Austin), Hai Huang (University of Texas at Austin)
GenerationData SynthesisKnowledge DistillationDiffusion modelScore-based ModelImageText
🎯 What it does: A data-free resolution generation method is proposed, which combines Stable Diffusion with Score Identity Distillation (SiD) and Classifier-Free Guidance (CFG) to obtain a single-step text-to-image generator.
Gumbel Counterfactual Generation From Language Models
Shauli Ravfogel (New York University), Ryan Cotterell (ETH Zurich)
GenerationLarge Language ModelText
🎯 What it does: This paper proposes to reconstruct language models as structural equation models using the Gumbel-max technique, thereby achieving string-based causal counterfactual generation.
Gyrogroup Batch Normalization
Ziheng Chen (University of Trento), Nicu Sebe (University of Trento)
ClassificationRecognitionVideoGraph
🎯 What it does: A general GyroBN batch normalization framework is proposed, suitable for pseudo-reductive Gyrogroups, and implemented on Grassmannian and hyperbolic spaces;
h4rm3l: A Language for Composable Jailbreak Attack Synthesis
Moussa Koulako Bala Doumbouya (Stanford University), Christopher D Manning (Stanford University)
Adversarial AttackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A domain-specific language (DSL) h4rm3l for composable jailbreak attacks is proposed, and a program synthesis framework based on bandit algorithms is constructed to automatically generate high-success-rate jailbreak attacks targeting large language models (LLMs);
HADAMRNN: BINARY AND SPARSE TERNARY ORTHOGONAL RNNS
Armand Foucault (Institut de Mathématiques de Toulouse), Franck Mamalet (Institut de Recherche Technologique Saint Exupéry)
Recurrent Neural NetworkSequentialBenchmark
🎯 What it does: Proposed a binary and sparse ternary orthogonal recurrent neural network based on Hadamard matrices (HadamRNN and Block-HadamRNN), achieving an efficient and lightweight RNN model on edge devices.
HaDeMiF: Hallucination Detection and Mitigation in Large Language Models
Xiaoling Zhou (Peking University), Shikun Zhang (Peking University)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The HADEMIF framework is proposed, utilizing two small networks (an interpretable D3T and MLP) to detect hallucinations in the output space and internal hidden states of LLMs, and calibrating logits through network outputs, achieving a unified process for hallucination detection and calibration.
HAINAN: Fast and Accurate Transducer for Hybrid-Autoregressive ASR
Hainan Xu (NVIDIA Corporation), Boris Ginsburg (NVIDIA Corporation)
RecognitionTransformerAudio
🎯 What it does: A hybrid autoregressive inference transformer named HAINAN is proposed, capable of performing speech recognition in autoregressive, non-autoregressive, and semi-autoregressive modes.
HALL-E: Hierarchical Neural Codec Language Model for Minute-Long Zero-Shot Text-to-Speech Synthesis
Yuto Nishimura (University of Tokyo), Nakamasa Inoue (University of Tokyo)
GenerationData SynthesisKnowledge DistillationTransformerLarge Language ModelAudio
🎯 What it does: Two technologies, MReQ and HALL-E, are proposed to achieve minute-level zero-shot text-to-speech (TTS) synthesis, breaking through the frame rate bottleneck of traditional LLM-TTS in long audio generation.
Hallo2: Long-Duration and High-Resolution Audio-Driven Portrait Image Animation
Jiahao Cui (Fudan University), Jingdong Wang (Baidu Inc.)
GenerationData SynthesisDiffusion modelImageVideoTextAudio
🎯 What it does: Achieved long-duration (several minutes to even hours) 4K resolution audio-driven portrait animation, with adjustable text prompts to control expressions and actions.
Halton Scheduler for Masked Generative Image Transformer
Victor Besnier (Valeo.ai), Matthieu Cord (Valeo.ai)
GenerationTransformerImage
🎯 What it does: A new Halton scheduler is proposed for the token decoding order in the Masked Generative Image Transformers (MaskGIT) generation process, improving the sampling strategy;
HAMSTER: Hierarchical Action Models for Open-World Robot Manipulation
Yi Li (NVIDIA), Ankit Goyal (NVIDIA)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityPoint Cloud
🎯 What it does: A hierarchical visual language action model (HAMSTER) is proposed, which first fine-tunes a large-scale pre-trained VLM on out-of-domain data to generate 2D paths, and then executes robot operations based on that path using a low-level 3D policy.
Handling Delay in Real-Time Reinforcement Learning
Ivan Anokhin (Mila), Samira Ebrahimi Kahou (CIFAR AI Chair)
Reinforcement LearningSequential
🎯 What it does: This study investigates the delay problem in real-time reinforcement learning and proposes a network architecture that introduces temporal skip connections and historical observation enhancement within a parallel layer computing framework, validated through theoretical and experimental evidence of its advantages.
HARDMath: A Benchmark Dataset for Challenging Problems in Applied Mathematics
Jingxuan Fan (Harvard University), Michael Brenner
Large Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: An automatically generated dataset HARDMATH containing 1060 advanced applied mathematics problems was created, and a subset HARDMATH-MINI (366 problems) along with 40 context-based semantic problems were constructed through manual verification.
HarmAug: Effective Data Augmentation for Knowledge Distillation of Safety Guard Models
Seanie Lee (KAIST), Sung Ju Hwang (KAIST)
Safty and PrivacyComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Train and deploy a security protection model with only 435M parameters, utilizing knowledge distillation and self-made harmful instruction data augmentation techniques to improve detection performance in low-resource environments such as mobile devices.
Harnessing Diversity for Important Data Selection in Pretraining Large Language Models
Chi Zhang (Beijing Institute of Technology), Conghui He (Renmin University of China)
TransformerLarge Language ModelText
🎯 What it does: A data selection method called Quad is proposed, aimed at improving the pre-training effectiveness of large language models by balancing data quality and diversity.
Harnessing Webpage UIs for Text-Rich Visual Understanding
Junpeng Liu (Chinese University of Hong Kong), Xiang Yue (Carnegie Mellon University)
RecognitionObject DetectionData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelTextMultimodality
🎯 What it does: Utilizing the accessibility tree of web UI and text large language models (LLM) to synthesize multimodal instructions, generating 7.3M UI task samples (MultiUI), and enhancing text-visual understanding capabilities on multimodal LLM through a two-stage training approach.
HART: Efficient Visual Generation with Hybrid Autoregressive Transformer
Haotian Tang (Massachusetts Institute of Technology), Song Han (Massachusetts Institute of Technology)
GenerationTransformerDiffusion modelImageText
🎯 What it does: The HART model is proposed, which combines autoregressive Transformers with lightweight residual diffusion, capable of directly generating high-quality images of 1024×1024 from text prompts.
Has the Deep Neural Network learned the Stochastic Process? An Evaluation Viewpoint
Harshit Kumar (Georgia Institute of Technology), Saibal Mukhopadhyay (Georgia Institute of Technology)
Convolutional Neural NetworkRecurrent Neural NetworkAuto EncoderTime SeriesSequentialFinance Related
🎯 What it does: A new evaluation framework is proposed for assessing deep neural networks (DNN) in predicting the evolution of random complex systems, which includes the statistical ground truth (Statistic-GT) and a fidelity metric to stochastic processes (Fidelity to Stochastic Process, F2SP). It is demonstrated that the expected calibration error (ECE) is the only metric that can test F2SP based solely on a single observation.
HASARD: A Benchmark for Vision-Based Safe Reinforcement Learning in Embodied Agents
Tristan Tomilin (Eindhoven University of Technology), Mykola Pechenizkiy (Eindhoven University of Technology)
Safty and PrivacyReinforcement LearningImageBenchmark
🎯 What it does: A new safety reinforcement learning benchmark based on ViZDoom, called HASARD, has been proposed and made public, which includes six visual perception tasks with three difficulty levels.
Have the VLMs Lost Confidence? A Study of Sycophancy in VLMs
Shuo Li (Fudan University), Xuanjing Huang (Fudan University)
Supervised Fine-TuningPrompt EngineeringVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This study investigates the phenomenon of sycophancy in visual language models and proposes the MM-SY benchmark evaluation method.
HD-Painter: High-Resolution and Prompt-Faithful Text-Guided Image Inpainting with Diffusion Models
Hayk Manukyan (Picsart AI Research), Humphrey Shi (Picsart AI Research)
RestorationGenerationDiffusion modelImage
🎯 What it does: A framework for high-resolution text-guided image inpainting (HD-Painter) is proposed, which achieves this without training by transforming self-attention into Prompt-Aware Introverted Attention (PAIntA) and incorporating a post-guidance mechanism called Reweighting Attention Score Guidance (RASG) to enhance the consistency between the inpainting area and the text prompts.
HeadMap: Locating and Enhancing Knowledge Circuits in LLMs
Xuehao Wang (Southern University of Science and Technology), Yu Zhang (Tencent Technology Co., Ltd)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A layer-conditioned localization algorithm is proposed to identify knowledge circuits composed of key attention heads in large language models, and based on this, a parameter-efficient fine-tuning method called HeadMap is designed.
Heavy-Tailed Diffusion Models
Kushagra Pandey (NVIDIA), Morteza Mardani (NVIDIA)
GenerationData SynthesisDiffusion modelTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a diffusion model using a multivariate Student-t distribution as a noise prior, which can better capture the tail behavior of high-dimensional data.
Heavy-Tailed Diffusion with Denoising Levy Probabilistic Models
Dario Shariatian (INRIA), Alain Oliviero Durmus (École Polytechnique)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A Denoising Lévy Probabilistic Model (DLPM) based on α-stable distribution and its deterministic sampling version DLIM are proposed, extending the traditional DDPM.
HELM: Hierarchical Encoding for mRNA Language Modeling
Mehdi Yazdani-Jahromi (University of Central Florida), Rui Liao (Johnson & Johnson Innovative Medicine)
GenerationData SynthesisProtein Structure PredictionTransformerLarge Language ModelTextBiomedical Data
🎯 What it does: A hierarchical encoding pre-training method for mRNA sequences, called HELM, is proposed and evaluated on tasks such as attribute prediction, sequence generation, and antibody mRNA region annotation.
HELMET: How to Evaluate Long-context Models Effectively and Thoroughly
Howard Yen (Princeton University), Danqi Chen (Princeton University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: The HELMET benchmark is proposed for the systematic evaluation of the multidimensional performance of Long Context Language Models (LCLMs).
HelpSteer2-Preference: Complementing Ratings with Preferences
Zhilin Wang (NVIDIA), Yi Dong (NVIDIA)
Recommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Released a supplementary preference annotation dataset and conducted a head-to-head comparison of the Bradley-Terry and regression-based reward models on the same data for the first time, proposing a training method that combines both models; trained a reward model that achieves 94.1% on RewardBench, comparable in alignment performance to GPT-4o and Claude 3.5.
Herald: A Natural Language Annotated Lean 4 Dataset
Guoxiong Gao (Peking University), Bin Dong (Peking University)
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: A natural language to Lean 4 formal statement translation pipeline based on hierarchical retrieval enhancement is proposed, and the HeralD dataset is generated.
HERO: Human-Feedback Efficient Reinforcement Learning for Online Diffusion Model Finetuning
Ayano Hiranaka (Sony AI), Yuki Mitsufuji (Sony AI)
GenerationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelContrastive LearningImage
🎯 What it does: Proposes the HERO framework, which utilizes online human feedback to fine-tune Stable Diffusion through reinforcement learning, achieving efficient controllable text-to-image generation;
Hessian-Free Online Certified Unlearning
Xinbao Qiao (Zhejiang University), Ermin Wei (Northwestern University)
OptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A Hessian-free online machine learning model forgetting method is proposed, which can quickly delete training samples without retraining.
HexGen-2: Disaggregated Generative Inference of LLMs in Heterogeneous Environment
YOUHE JIANG, Binhang Yuan (Hong Kong University of Science and Technology)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Developed the HEXGEN-2 system, which implements discrete inference for large language models (LLM) during the prefill and decoding phases on heterogeneous GPU clusters, and proposed a scheduling algorithm for this scenario.
HG-Adapter: Improving Pre-Trained Heterogeneous Graph Neural Networks with Dual Adapters
Yujie Mo (National University of Singapore), Xinchao Wang (University of Electronic Science and Technology of China)
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: The HG-Adapter framework is proposed, which improves the generalization performance of pre-trained heterogeneous graph neural networks in downstream tasks through a dual-structure-aware adapter, contrastive loss with label propagation, and two types of self-supervised losses.
HGM³: Hierarchical Generative Masked Motion Modeling with Hard Token Mining
Minjae Jeong (Pohang University of Science and Technology), Won Hwa Kim (Pohang University of Science and Technology)
GenerationData SynthesisKnowledge DistillationTransformerVideoText
🎯 What it does: A text-driven motion generation framework HGM3 is proposed, which combines Hard Token Mining and hierarchical semantic graph conditions.
HiBug2: Efficient and Interpretable Error Slice Discovery for Comprehensive Model Debugging
Muxi Chen (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
ClassificationObject DetectionPose EstimationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerLarge Language ModelPrompt EngineeringImage
🎯 What it does: HiBug2 is proposed, an automated error slice discovery and model repair framework designed to enhance the robustness of visual models in real-world scenarios.
Hidden in the Noise: Two-Stage Robust Watermarking for Images
Kasra Arabi (New York University), Niv Cohen (New York University)
GenerationData SynthesisRetrievalDiffusion modelImage
🎯 What it does: A two-stage robust watermarking method (WIND) based on the initial noise of the diffusion model is proposed, which can achieve watermark embedding and detection without affecting image quality.
Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models
Pit Neitemeier (Aleph Alpha Research), Lukas Balles (Aleph Alpha Research)
TransformerLarge Language ModelText
🎯 What it does: A hierarchical autoregressive Transformer architecture is proposed, combining character-level encoders/decoders with a word-level backbone, eliminating the dependence on fixed subword vocabularies.
Hierarchical Uncertainty Estimation for Learning-based Registration in Neuroimaging
Xiaoling Hu (Massachusetts General Hospital and Harvard Medical School), Juan Eugenio Iglesias (Massachusetts Institute of Technology)
Convolutional Neural NetworkSupervised Fine-TuningBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: A hierarchical uncertainty estimation framework is proposed, which can propagate the local uncertainty predicted by deep learning to the global transformation model and downstream tasks;
Hierarchical World Models as Visual Whole-Body Humanoid Controllers
Nicklas Hansen (University of California San Diego), Hao Su (University of California San Diego)
Robotic IntelligenceReinforcement LearningWorld ModelImageBenchmark
🎯 What it does: A hierarchical world model named Puppeteer has been designed and implemented for visual full-body humanoid control. It first pre-trains a low-level tracking agent using MoCap data, and then generates commands through a high-level agent based on visual input, enabling the robot to perform various complex motion tasks.
Hierarchically Encapsulated Representation for Protocol Design in Self-Driving Labs
Yu-Zhe Shi (Peking University), Qining Wang (Peking University)
Autonomous DrivingOptimizationLarge Language ModelTextBiomedical Data
🎯 What it does: A hierarchical packaging experimental protocol representation method is proposed, which includes instance actions, operation abstractions, and product flow models, and uses this representation to assist LLM in protocol planning, modification, and adjustment.
High-dimension Prototype is a Better Incremental Object Detection Learner
Yanjie Wang (Huazhong University of Science and Technology), Xu Zou (Huazhong University of Science and Technology)
Object DetectionKnowledge DistillationImage
🎯 What it does: A knowledge distillation framework based on high-dimensional mixed Gaussian prototypes is proposed for incremental object detection.
High-dimensional Analysis of Knowledge Distillation: Weak-to-Strong Generalization and Scaling Laws
Muhammed Emrullah Ildiz (University of Michigan), Samet Oymak (University of Michigan)
Knowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper studies knowledge distillation and weak-to-strong generalization under high-dimensional linear regression, providing a non-asymptotic risk definition and deriving the form of the optimal teacher model;
High-Dimensional Bayesian Optimisation with Gaussian Process Prior Variational Autoencoders
Siddharth Ramchandran (Aalto University), Harri Lähdesmäki
OptimizationDrug DiscoveryAuto EncoderTabular
🎯 What it does: Utilize GP prior VAE to learn a structured latent space and perform Bayesian Optimization in that space to efficiently search for the objective function in high-dimensional black-box optimization.
High-Dynamic Radar Sequence Prediction for Weather Nowcasting Using Spatiotemporal Coherent Gaussian Representation
Ziye Wang (Sun Yat-sen University), Ruimao Zhang (Sun Yat-sen University)
Recurrent Neural NetworkGaussian SplattingOptical FlowTime SeriesSequential
🎯 What it does: A framework for spatiotemporal consistency Gaussian light scattering (STC-GS) and memory-enhanced Mamba (GauMamba) based on 3D high-dynamic radar sequences is proposed to achieve 3D sequence prediction of radar echoes.
High-Precision Dichotomous Image Segmentation via Probing Diffusion Capacity
Qian Yu (Dalian University of Technology), Huchuan Lu (Dalian University of Technology)
SegmentationDiffusion modelImage
🎯 What it does: This paper proposes DiffDIS, a single-step binary image segmentation framework based on diffusion models, aimed at achieving high-resolution and fine-grained object segmentation.
High-Quality Joint Image and Video Tokenization with Causal VAE
Dawit Mureja Argaw (Korea Advanced Institute of Science and Technology), Fitsum Reda (NVIDIA)
GenerationCompressionConvolutional Neural NetworkAuto EncoderOptical FlowImageVideo
🎯 What it does: This paper proposes a causal continuous variational autoencoder (Causal Video VAE) that can simultaneously compress images and videos, achieving high-quality image and video reconstruction and generation through spatiotemporal convolution and attention.
High-quality Text-to-3D Character Generation with SparseCubes and Sparse Transformers.
Jiachen Qian (DreamTech), Feihu Zhang (DreamTech)
GenerationData SynthesisTransformerMesh
🎯 What it does: Proposed SparseCubes, a sparse differentiable grid representation, and Sparse Cube Transformer, achieving high-quality detail capture in text-to-3D anime character generation.
Higher-Order Graphon Neural Networks: Approximation and Cut Distance
Daniel Herbst (Technische Universität München), Stefanie Jegelka (Massachusetts Institute of Technology)
Graph Neural NetworkGraph
🎯 What it does: This paper extends higher-order graph neural networks (such as k-WL level IGN) to the graphon space, proposing Invariant Graphon Networks (IWNs) and defining signal-weighted homomorphic density to characterize the structure of graphon signals.
Highly Efficient Self-Adaptive Reward Shaping for Reinforcement Learning
Haozhe Ma (National University of Singapore), Tze-Yun Leong (National University of Singapore)
Reinforcement Learning
🎯 What it does: Proposes an Adaptive Success Rate Shaping (SASR) method to address the issue of extremely sparse rewards in reinforcement learning.
HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts
Hongjun Wang (Visual AI Lab, University of Hong Kong), Kai Han (Visual AI Lab, University of Hong Kong)
Domain AdaptationTransformerContrastive LearningImage
🎯 What it does: A new framework HiLo is proposed for general category discovery in the presence of domain shift;
HiRA: Parameter-Efficient Hadamard High-Rank Adaptation for Large Language Models
Qiushi Huang (Southern University of Science and Technology), Yu Zhang (University of Surrey)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The HiRA method is proposed, which combines pre-trained weights with low-rank matrices using the Hadamard product to achieve high-rank adaptation of parameters while maintaining the parameter and computational advantages of PEFT.
HiSplat: Hierarchical 3D Gaussian Splatting for Generalizable Sparse-View Reconstruction
Shengji Tang (Fudan University), Wanli Ouyang (Shanghai AI Laboratory)
RestorationGenerationDepth EstimationTransformerGaussian SplattingImage
🎯 What it does: This paper presents HiSplat, a scalable 3D Gaussian rendering framework for sparse two-view scenarios, capable of simultaneously reconstructing large-scale structures and fine details through a multi-scale hierarchical structure.
HMoRA: Making LLMs More Effective with Hierarchical Mixture of LoRA Experts
Mengqi Liao (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
TransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: A new LLM fine-tuning method called HMoRA is proposed, which combines LoRA experts with hierarchical mixed routing to achieve efficient multi-task learning.
Holistic Reasoning with Long-Context LMs: A Benchmark for Database Operations on Massive Textual Data
Seiji Maekawa (Megagon Labs), Nikita Bhutani (Megagon Labs)
TransformerLarge Language ModelTextTabularBenchmarkChain-of-Thought
🎯 What it does: A benchmark called HoloBench is proposed for system evaluation of long-context language models' global reasoning capabilities on large-scale text data.
Holistically Evaluating the Environmental Impact of Creating Language Models
Jacob Morrison (Allen Institute for AI), Jesse Dodge (Allen Institute for AI)
Large Language ModelMixture of ExpertsText
🎯 What it does: A comprehensive assessment of energy, carbon emissions, and water consumption during the development, training, and inference processes of large-scale language models.
Holographic Node Representations: Pre-training Task-Agnostic Node Embeddings
Beatrice Bevilacqua (Purdue University), Bruno Ribeiro (Purdue University)
Recommendation SystemRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Designed and implemented HoloGNN, a method capable of pre-training task-agnostic node representations, which can quickly adapt to different order tasks (node-level, edge-level, and higher-order) on new tasks through a lightweight reduction map.
Homomorphism Counts as Structural Encodings for Graph Learning
Linus Bao (University of Oxford), Matthias Lanzinger (TU Wien)
Graph Neural NetworkTransformerGraph
🎯 What it does: A structure encoding method based on graph isomorphism counting—Motif Structural Encoding (MoSE)—is proposed to provide inductive bias of graph structure for graph Transformers.
Homomorphism Expressivity of Spectral Invariant Graph Neural Networks
Jingchu Gai (Peking University), Liwei Wang (Peking University)
Graph Neural NetworkGraph
🎯 What it does: This paper quantitatively characterizes the expressive power of Spectral Invariant GNNs through the framework of homomorphic expressiveness and provides a complete family of distinguishable graphs—parallel trees.
HOPE for a Robust Parameterization of Long-memory State Space Models
Annan Yu (Cornell University), N. Benjamin Erichson (Lawrence Berkeley National Laboratory)
Time SeriesSequential
🎯 What it does: This paper proposes a novel parameterization method based on Hankel operator Markov parameters, called HOPE, for reconstructing the state space model (SSM) of linear time-invariant (LTI) systems. The advantages of HOPE in initialization, training stability, and long-term memory are validated through theoretical and experimental evidence.
Horizon Generalization in Reinforcement Learning
Vivek Myers (University of California Berkeley), Benjamin Eysenbach (Princeton University)
Reinforcement LearningContrastive LearningTabular
🎯 What it does: This paper studies the horizon generalization of goal-oriented reinforcement learning by defining planning invariance and quasimetric to prove and experimentally validate its feasibility.
Hot-pluggable Federated Learning: Bridging General and Personalized FL via Dynamic Selection
Lei Shen (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
Federated LearningSafty and PrivacyConvolutional Neural NetworkImage
🎯 What it does: A hot-plug federated learning framework (HPFL) is proposed, which utilizes a shared feature extractor and client-trained pluggable plugins to dynamically select the most suitable plugin during inference to enhance global generalization performance.
Hotspot-Driven Peptide Design via Multi-Fragment Autoregressive Extension
Jiahan Li (Tsinghua University), Jianzhu Ma (Tsinghua University)
GenerationDrug DiscoveryBiomedical DataStochastic Differential Equation
🎯 What it does: A hotspot-driven autoregressive generative model, PepHAR, is proposed for designing peptide bundles that meet geometric and interaction requirements under a given target protein.
How Discrete and Continuous Diffusion Meet: Comprehensive Analysis of Discrete Diffusion Models via a Stochastic Integral Framework
Yinuo Ren (Stanford University), Lexing Ying (Stanford University)
Diffusion modelStochastic Differential Equation
🎯 What it does: This paper constructs a Lévy-type stochastic integral framework by introducing a variable intensity Poisson random measure, systematically analyzes the errors of discrete diffusion models, and provides the error upper bounds for τ-leaping and homogenization algorithms under KL divergence.
How DNNs break the Curse of Dimensionality: Compositionality and Symmetry Learning
Arthur Jacot (New York University), Yuxiao Wen (New York University)
🎯 What it does: This paper demonstrates that deep neural networks (DNNs) can effectively learn function compositions with bounded F1 norms, thereby breaking the curse of dimensionality and surpassing the capabilities of shallow networks.
How Do Large Language Models Understand Graph Patterns? A Benchmark for Graph Pattern Comprehension
Xinnan Dai (Michigan State University), Caihua Shan (Microsoft Research)
Large Language ModelPrompt EngineeringGraphBenchmarkChain-of-Thought
🎯 What it does: A complete benchmark for graph pattern understanding is proposed and constructed, covering 11 sub-tasks (pattern translation, isomorphic mapping, graph modification, pattern detection, k-core detection, frequent subgraph mining, discriminative pattern learning, etc.) and evaluating 7 mainstream LLMs.
How Does Critical Batch Size Scale in Pre-training?
Hanlin Zhang (Harvard University), Sham M. Kakade
Hyperparameter SearchTransformerLarge Language ModelText
🎯 What it does: This study investigates and quantifies the scaling laws of the critical batch size (CBS) in large-scale pre-training, conducting systematic experiments on autoregressive Transformers with parameters ranging from 85M to 1.2B, and derives empirical and theoretical models for CBS.
How Does Vision-Language Adaptation Impact the Safety of Vision Language Models?
Seongyun Lee (KAIST AI), Minjoon Seo (KAIST AI)
Safty and PrivacyTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: This paper studies the impact of visual language adaptation (VL adaptation) on the security of large visual language models (LVLM) and proposes a model weight merging method to enhance security while maintaining multimodal capabilities.
How efficient is LLM-generated code? A rigorous & high-standard benchmark
Ruizhong Qiu (University of Illinois Urbana-Champaign), Hanghang Tong (University of Illinois Urbana-Champaign)
Computational EfficiencyAI Code AssistantLarge Language ModelTextBenchmark
🎯 What it does: The ENAMEL benchmark is proposed to rigorously evaluate the efficiency of code generated by large language models.