AAAI Conference on Artificial Intelligence Β· 2140 papers
GloCTM: Cross-Lingual Topic Modeling via a Global Context Space
Nguyen Tien Phat (Hanoi University of Science and Technology), Thien Huu Nguyen (University of Oregon)
CodeRepresentation LearningAuto EncoderText
π― What it does: Proposes GloCTM, leveraging a global context space and a dual-channel VAE architecture to achieve unified learning and alignment of cross-lingual topics.
π― What it does: Proposed an online low-frame-rate multi-object tracking framework called GLoMOT, which realizes real-time association using graph neural networks.
GMAI-VL & GMAI-VL-5.5M: A Large Vision-Language Model and a Comprehensive Multimodal Dataset Towards General Medical AI
Tianbin Li (Shanghai Artificial Intelligence Laboratory), Junjun He (Shanghai Artificial Intelligence Laboratory)
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBiomedical Data
π― What it does: Constructed a general-purpose medical vision-language model named GMAI-VL with 7B parameters, and developed a high-quality multimodal medical dataset named GMAI-VL-5.5M containing 5.5M samples based on over 200 specialized medical datasets, supporting cross-modal learning and medical question answering from images to text.
Goal-Oriented Time-Series Forecasting: Foundation Framework Design
Luca-Andrei Fechete (Γcole Polytechnique), Tareq Si Salem (Huawei Technologies)
CodeConvolutional Neural NetworkTransformerTime Series
π― What it does: This paper proposes a time series prediction framework that performs fine-grained partitioning and dynamic reweighting of the prediction space during training, enabling the model to flexibly focus on target intervals during inference according to different application requirements without needing retraining.
GOMPSNR: Reflourish the Signal-to-Noise Ratio Metric for Audio Generation Tasks
Lingling Dai (Institute of Acoustics, Chinese Academy of Sciences), Chengshi Zheng (Institute of Acoustics, Chinese Academy of Sciences)
CodeGenerationMultimodalityAudio
π― What it does: Propose an improved audio quality evaluation metric called GOMPSNR, and design new phase-guided and joint amplitude-phase optimization loss functions based on this metric to enhance the generation quality of neural vocoders and audio codecs.
π― What it does: Developed an audio generation framework named SS2A, capable of perceiving and fusing multiple sound sources based on three modalitiesβvisual, text, and audioβto achieve more immersive and controllable audio generation.
π― What it does: By first separating the target object within the 3D Gaussian Splatting framework, using Point-MAE to complete the geometry of unobserved regions, then performing brightness and texture correction on reference images projected to different viewpoints, and finally fine-tuning 3DGS with the corrected images, achieving multi-view consistent 3D object removal.
π― What it does: Proposed a binary encoding scheme based on diffusion models that can reconstruct and enhance the readability and author style consistency of 3D space handwritten characters from a small number of examples.
GraFT: Infusing Pre-trained Transformers with Relational Structure for Time Series Forecasting
Yuqi Yuan (University of Science and Technology Beijing), Wenbing Zhao (Cleveland State University)
CodeGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTime SeriesBenchmark
π― What it does: Improving long-term prediction for multivariate time series by constructing a Heterogeneous Patch Relationship Graph (HPRG) and using R-GCN to inject structural priors into pre-trained Transformer models.
GRAM-RΒ²: Self-Training Generative Foundation Reward Models for Reward Reasoning
Chenglong Wang (Northeastern University), Tong Xiao (Northeastern University)
CodeGenerationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
π― What it does: Designed and trained a self-supervised pre-trained generative reward model, GRAM-RΒ², which enhances reward reasoning capabilities through self-training iterations using unlabeled data, and provides a dedicated preference proof model to generate reasoning processes for unlabeled labels.
π― What it does: Propose a hybrid multi-agent path planning framework (LaGAT) that uses graph attention neural networks as heuristic guidance for search.
π― What it does: Propose the BalanceGCL framework, which improves the representation quality of unsupervised graph contrastive learning by generating class-balanced hard negative samples and fine-grained semantic positive samples.
Graph Meets Deep Unfolding: An Interpretable Mutual-benefit Multi-view Learning Network
Renjie Lin (University of Electronic Science and Technology of China), Le Zhang (University of Electronic Science and Technology of China)
CodeOptimizationExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkGraph
π― What it does: Designed an interpretable deep unrolling network, IMML-Net, for multi-view graph learning, integrating sparse, low-rank constraints, and noise handling to achieve joint learning of multi-view graphs.
Graph Neural Field with Spatial-Correlation Augmentation for HRTF Personalization
De Hu (Inner Mongolia University), Cuicui Jiang (Inner Mongolia University)
CodeGenerationGraph Neural NetworkAudio
π― What it does: Propose the GraphNF-SCA framework, which first generates position-agnostic personalized HRTF using graph neural networks, and then refines it through a spatial correlation graph network to achieve high-quality HRTF prediction for unseen subjects.
CodeExplainability and InterpretabilityLarge Language ModelTextBenchmark
π― What it does: Propose the Graph of Verification (GoV) framework, which uses directed acyclic graphs (DAGs) and configurable node blocks to perform hierarchical step-by-step verification of the LLM reasoning process.
π― What it does: Proposes a boundary-aware calibration method called BaCa based on dual dynamic dictionaries, achieving test-time OOD detection for graph data without fine-tuning pre-trained graph neural networks or using auxiliary OOD samples.
π― What it does: This paper proposes a point cloud analysis framework named GSPoint, combining a graph smoothing module and a local geometric learning module to address issues caused by traditional spherical query methods, such as sparse boundary connections and cross-node noise;
π― What it does: Proposed the Graph VQ-Transformer (GVT) two-stage generation framework: the first stage uses Graph VQ-VAE to compress molecular graphs into high-fidelity discrete latent sequences; the second stage trains a self-attention Transformer to generate new molecules based on these latent sequences.
π― What it does: Propose the GOCC framework, which utilizes sentence-level and cluster-level graph structures to enhance representation learning and contrastive learning for short text unsupervised clustering.
π― What it does: This paper proposes a generative model based on graph conditional flow matching, which encodes the foreign key graph of relational databases using graph neural networks to generate all table data content in an end-to-end manner in one go.
π― What it does: This paper proposes a graph-based domain collaborative adaptation framework called GDCIQA for cross-domain image quality assessment.
Graph-of-Mark: Promote Spatial Reasoning in Multimodal Language Models with Graph-Based Visual Prompting
Giacomo Frisoni (University of Bologna), Gianluca Moro (University of Bologna)
CodeObject DetectionSegmentationDepth EstimationRepresentation LearningPrompt EngineeringVision Language ModelImageTextGraph
π― What it does: This paper proposes Graph-of-Mark (GoM), a pixel-level visual prompting technique that overlays scene graphs (nodes as detected objects, edges as their spatial relationships) on images to enhance the zero-shot performance of multimodal language models in spatial reasoning tasks.
Graph2Video: Leveraging Video Models to Model Dynamic Graph Evolution
Hua Liu (Southern University of Science and Technology), Yu Zhang (City University of Hong Kong)
CodeGraph Neural NetworkTransformerVideoGraph
π― What it does: This paper proposes the Graph2Video framework, which serializes the spatial-temporal neighborhood of target edges in dynamic graphs into 'graph videos,' and extracts spatiotemporal embeddings using a frozen video base model as a lightweight, plug-and-play edge-level memory unit to enhance dynamic link prediction performance.
π― What it does: Proposed a lightweight graph-guided 6-DoF grasping pose estimation network called GraphGrasp, which can directly construct scene graphs, object graphs, and grasp graphs from point clouds to achieve grasping point prediction and pose reasoning.
GraphIF: Enhancing Multi-Turn Instruction Following for Large Language Models with Relation Graph Prompt
Zhenhe Li (University of Science and Technology of China), Qi Song (University of Science and Technology of China)
CodeTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextBenchmark
π― What it does: Proposed the GraphIF framework, which utilizes relational graphs to model multi-turn dialogues and generate graph prompts, enhancing the multi-turn instruction following capability of large language models.
π― What it does: Propose a novel knowledge graph reasoning framework named GRAPHORACLE based on a relation dependency graph (RDG), which achieves efficient and accurate chain-of-thought reasoning in fully inductive (both entities and relations unseen) and cross-domain reasoning scenarios.
GraphRAG-Induced Dual Knowledge Structure Graphs for Personalized Learning Path Recommendation
Xinghe Cheng (Jinan University), Weiqi Luo (Griffith University)
CodeRecommendation SystemGraph Neural NetworkLarge Language ModelReinforcement LearningAgentic AITextGraphRetrieval-Augmented Generation
π― What it does: Propose the KnowLP framework, which combines prerequisite relationships and similar relationships to construct a dual knowledge structure graph, using EDU-GraphRAG to automatically generate knowledge graphs and DLRL (including three agents: prerequisite, similar, and difficulty) to generate personalized learning paths.
Ground What You See: Hallucination-Resistant MLLMs via Caption Feedback, Diversity-Aware Sampling, and Conflict Regularization
Miao Pan (Zhejiang University), Xuhong Zhang (National Certification Technology Co Ltd)
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelContrastive LearningVideoTextMultimodality
π― What it does: Propose an RL optimization framework based on visual calibration, reward variance sampling, and conflict regularization, significantly reducing hallucinations in multimodal large language models.
CodeRobotic IntelligenceTransformerVision Language ModelVision-Language-Action ModelDiffusion modelImageMultimodality
π― What it does: Proposes an OCVLA framework that aligns robot action prediction to camera space, enabling direct prediction of actions in the observation space and addressing the mismatch between perception and action spaces.
Group Causal Policy Optimization for Post-Training Large Language Models
Ziyin Gu (Institute of Software Chinese Academy of Sciences), Wenwen Qiang (Institute of Software Chinese Academy of Sciences)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningText
π― What it does: Proposes the Group Causal Policy Optimization (GCPO) algorithm, which enhances the reasoning and generation performance of large language models through causal-structured group policy optimization during the post-training phase.
Atasi Panda (Indian Institute of Science), Prajakta Nimbhorkar (Chennai Mathematical Institute)
CodeOptimizationFlow-based ModelTabular
π― What it does: Proposes the problem of project-platform matching on bipartite graphs, achieving softened fairness constraints by introducing overall and group-specific convex cost functions, aiming to minimize total cost while meeting a given utility threshold.
π― What it does: This paper proposes a parameter-efficient fine-tuning framework based on Group Orthogonal Low-Rank Adaptation (GOLA), specifically designed for RGB-T visual object tracking. The framework decomposes the LoRA low-rank matrix via singular value decomposition (SVD), separating key ranks and redundant ranks, and then clusters the redundant ranks into groups. Subsequently, orthogonal constraints are applied across different groups, forcing them to learn complementary features, thereby fully activating the redundant rank space and enhancing model expressiveness.
GROVER: Graph-guided Representation of Omics and Vision with Expert Regulation for Adaptive Spatial Multi-omics Fusion
Yongjun Xiao (Great Bay University), Xubin Zheng (Great Bay University)
CodeRepresentation LearningGraph Neural NetworkMixture of ExpertsContrastive LearningMultimodalityBiomedical Data
π― What it does: Propose the GROVER framework, which performs single-point adaptive fusion of spatial transcriptomics, proteomics, and tissue imaging based on graph convolutional networks, addressing challenges of multi-modal heterogeneity and spatial alignment.
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkLarge Language ModelText
π― What it does: An evaluation framework based on convolutional neural networks (CNN) for classifying attention weights is proposed to address the character-by-character memorization phenomenon in large language models, with a redesigned classification system for memory samples based on this framework;
CodeOptimizationTransformerReinforcement LearningVision Language ModelGaussian SplattingMultimodality
π― What it does: Designed the GUI-G2 reward framework, modeling GUI elements with a continuous 2D Gaussian distribution to achieve denser learning signals.
π― What it does: Developed the GUIDE framework, leveraging 3D Gaussian sparse representation to achieve instance-level occupancy prediction, and performing object detection and tracking within the same framework.
Guided Distillation and Risk Adaptive Evolution for Multi-Robot Navigation
Xuyang Li (Xi'an Jiaotong University), Jianru Xue (Xi'an Jiaotong University)
CodeOptimizationKnowledge DistillationRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
π― What it does: This paper proposes a two-phase offline reinforcement learning framework called GUIDER, which first utilizes a large language model (LLM) to distill knowledge into a lightweight policy guidance model (PGM), providing reliable guidance when the agent is uncertain; subsequently, the LLM is transformed into a semantic evolution engine to automatically search and generate environment-specific risk-adaptive strategies, achieving safe and efficient multi-robot navigation.
π― What it does: Developed GuideGen, a controllable generation framework capable of generating full-body CT volumes and corresponding anatomical masks using only structured medical text prompts.
GUIDER: Uncertainty Guided Dynamic Re-ranking for Large Language Models Based Recommender Systems
Cai Xu (Xidian University), Meng Yan (Xidian University)
CodeRecommendation SystemTransformerLarge Language ModelTextSequential
π― What it does: Design and implement the GUIDER framework to quantify and decompose uncertainty in LLM-generated recommendations, and enhance recommendation reliability through four-quadrant dynamic re-ranking.
π― What it does: Propose a point cloud denoising framework based on structural prior guidance, first extracting local geometric structural priors through vector quantization, and then achieving fine-grained feature reconstruction using FiLM-modulated attention.
π― What it does: Generate realistic noise using a single-pair guided diffusion model, achieving noise synthesis without camera metadata and requiring only a pair of noisy/clean images
π― What it does: Proposed a generic unified structured light optimization framework, GUSLO, which can accomplish geometric calibration and illumination compensation under a single projection, achieving high-precision 3D reconstruction for various structured light patterns, including binary stripes, speckles, and color-encoded patterns;
CodeExplainability and InterpretabilityComputational EfficiencyLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Proposed and implemented the Aspect-Based Causal Abstention (ABCA) framework, which can decide whether to refuse answering by analyzing the diversity of internal knowledge in LLMs before generation, thereby reducing hallucinations.
HalluClean: A Unified Framework to Combat Hallucinations in LLMs
Yaxin Zhao (Harbin Institute of Technology), Yu Zhang (Harbin Institute of Technology)
CodeExplainability and InterpretabilityTransformerPrompt EngineeringTextBenchmarkFinance RelatedChain-of-Thought
π― What it does: Proposes the hallucination removal framework HalluClean, which detects and corrects hallucinations in LLM-generated text using structured reasoning.
HALO: Hardware-Aware Quantization with Low Critical-Path-Delay Weights for LLM Acceleration
Rohan Juneja (National University of Singapore), Li-Shiuan Peh (National University of Singapore)
CodeComputational EfficiencyTransformerText
π― What it does: This paper proposes HALO, a hardware-aware post-training quantization framework that efficiently accelerates LLM inference by combining weight quantization with dynamic voltage and frequency scaling (DVFS).
HAMLET4Fairness: Enhancing Fairness in AI Pipelines Through Human-Centered AutoML and Argumentation
Joseph Giovanelli (Alma Mater Studiorum - University of Bologna), Roberta Calegari (Alma Mater Studiorum - University of Bologna)
CodeOptimizationExplainability and InterpretabilityHyperparameter SearchTabularBenchmark
π― What it does: Developed HAMLET4Fairness, a framework that integrates structured argumentation with human-centric AutoML for fairness-driven AI pipeline optimization;
HaNa: Hardness and Noise-Aware Robust Cross-modal Retrieval
Fangming Zhong (Dalian University of Technology), Suhua Zhang (Dalian University of Technology)
CodeRetrievalTransformerVision Language ModelContrastive LearningMultimodality
π― What it does: Proposes the HaNa method, which combines an adaptive reweighting mechanism for clean samples and asymmetric regularization for noisy samples to enhance cross-modal retrieval performance under noisy conditions.
HanjaBridge: Resolving Semantic Ambiguity in Korean LLMs via Hanja-Augmented Pre-Training
Seungho Choi (Wisenut), Bongsu Kim (Wisenut)
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelContrastive LearningText
π― What it does: Propose HanjaBridge, a semantic enhancement technique that inserts Hanja (Chinese characters) candidate words into Korean LLM pre-training processes, leveraging multiple candidates and context to resolve semantic ambiguity in Korean homonyms; based on this, continuous pre-training (CPT) and token-level knowledge distillation are adopted to prevent catastrophic forgetting, and no additional Hanja tokens are required during inference, maintaining original inference efficiency.
Hard vs. Noise: Resolving Hard-Noisy Sample Confusion in Recommender Systems via Large Language Models
Tianrui Song (Hong Kong University of Science and Technology (Guangzhou)), Hao Liu (Hong Kong University of Science and Technology)
CodeRecommendation SystemLarge Language ModelContrastive LearningTextGraph
π― What it does: This work proposes the LLMHNI framework, which leverages two auxiliary signals provided by large language models (LLMs)βsemantic relevance and logical relevanceβto address the confusion between hard samples and noisy samples in recommendation systems. It achieves more robust implicit recommendations through semantic-guided hard negative sampling and logic relevance-driven interaction denoising.
HarmoQ: Harmonized Post-Training Quantization for High-Fidelity Image Super-Resolution
Hongjun Wang (University of Tokyo), Yinqiang Zheng (University of Tokyo)
CodeRestorationSuper ResolutionTransformerImage
π― What it does: Propose HarmoQ, a unified post-training quantization framework specifically designed for Transformer-based super-resolution models, combining structural residual calibration, quantization scale balance, and adaptive boundary refinement to achieve coordinated optimization of weights and activations.
Harnessing Textual Semantic Priors for Knowledge Transfer and Refinement in CLIP-Driven Continual Learning
Lingfeng He, Nannan Wang (Huawei Technologies Co Ltd)
CodeKnowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelMultimodality
π― What it does: This paper proposes a unified framework called SECA, which leverages text semantic priors to guide knowledge transfer and visual prototype refinement in CLIP-driven continual learning.
Harnessing the Unseen: The Hidden Influence of Intrinsic Knowledge in Long-Context Language Models
Yu Fu (University of California Riverside), Yue Dong (University of California Riverside)
CodeRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: Investigate the impact of parameter knowledge (intrinsic knowledge) in long context language models during generation, and propose a hybrid Needle-in-a-Haystack evaluation method to simultaneously assess parameter recall and external retrieval capabilities.
Harnessing Vision-Language Models for Time Series Anomaly Detection
Zelin He (Pennsylvania State University), Matthew Reimherr (Pennsylvania State University)
CodeAnomaly DetectionTransformerVision Language ModelTime Series
π― What it does: This paper proposes a two-stage zero-shot time series anomaly detection framework called VLM4TS, which converts one-dimensional sequences into two-dimensional images. It first uses ViT4TS to locate candidate anomalies and then employs VLM4TS to verify precise boundaries.
π― What it does: This paper proposes HC2-GNN, a graph neural network based on hierarchical clustering and reduction, for efficient text classification.
HCC-3D: Hierarchical Compensatory Compression for 98% 3D Token Reduction in Vision-Language Models
Liheng Zhang (China University of Petroleum (East China)), Weifeng Liu (China University of Petroleum (East China))
CodeCompressionComputational EfficiencyVision Language ModelPoint Cloud
π― What it does: Proposed a hierarchical compensation compression framework named HCC-3D, which significantly reduces the computational cost of 3D-VLM and improves inference efficiency by compressing 3D point cloud features into 12 tokens.
HDΒ²-SSC: High-Dimension High-Density Semantic Scene Completion for Autonomous Driving
Zhiwen Yang (Peking University), Yuxin Peng (Peking University)
CodeAutonomous DrivingTransformerPoint Cloud
π― What it does: Propose the HD-SSC framework to address the dimension and density gaps in camera-based semantic scene completion, achieving more accurate 3D semantic completion through pixel semantic decoupling and voxel refinement.
π― What it does: Proposed a Hierarchical Dynamic Gaussian Slicing (HDGS) framework that efficiently models city-scale 4D dynamic scenes using an anchor-based Gaussian model and achieves real-time novel view synthesis.
HDRMovieformer: A Transformer Framework and Benchmark for Cinematic SDR-to-HDR Conversion
Xianwei Li (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
CodeImage TranslationTransformerVideoBenchmark
π― What it does: Propose the HDRMovieformer framework to achieve movie-level SDR-to-HDR conversion, leveraging brightness-guided transformers and color refinement modules to restore high dynamic range images.
π― What it does: This paper proposes HACK, a training-agnostic head-aware KV cache compression framework, specifically optimized for KV cache accumulation and attention complexity issues in multi-scale generation processes of visual autoregressive (VAR) models.
HeartLLM: Discretized ECG Tokenization for LLM-Based Diagnostic Reasoning
Jinning Yang (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences), Wen Shi (Massachusetts General Hospital)
CodeRepresentation LearningConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTime SeriesBiomedical DataElectrocardiogram
π― What it does: Propose the HeartLLM framework, which discretizes 12-lead ECG signals into discrete symbolic tokens that share a vocabulary with LLMs, enabling open-ended question answering and report generation;
π― What it does: Proposed Heterogeneous Complementary Distillation (HCD), which maps the distinct features of teachers and students into a shared logits space via Complementary Feature Mapper, and promotes diverse knowledge transfer through Sub-logit Decoupled Distillation and Orthogonality Loss.
Heterogeneous Graph Neural Networks for Assumption-Based Argumentation
Preesha Gehlot (Imperial College London), Francesca Toni (Imperial College London)
CodeGraph Neural NetworkGraphBenchmark
π― What it does: Proposed a dependency-based heterogeneous graph neural network for predicting suspicious acceptability under stable semantics in assumption-based argumentation frameworks, and implemented a polynomial-time extended reconstruction algorithm.
Heterogeneous Uncertainty-Guided Composed Image Retrieval with Fine-Grained Probabilistic Learning
Haomiao Tang, Shu-Tao Xia (Hong Kong University Of Science And Technology)
CodeRetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodality
π― What it does: Propose a heterogeneous uncertainty-guided fine-grained probabilistic learning framework HUG, which uses Gaussian embeddings to represent queries and targets, achieving explicit modeling of content quality and multi-modal coordination uncertainties, thereby enhancing the robustness of synthetic image retrieval.
HGATSolver: A Heterogeneous Graph Attention Solver for FluidβStructure Interaction
Qin-Yi Zhang (Chinese Academy of Sciences), Zeng-Guang Hou (Chinese Academy of Sciences)
CodeOptimizationGraph Neural NetworkGraphPhysics Related
π― What it does: Propose a learning-based FSI solver HGATSolver based on heterogeneous graph attention, which can simultaneously capture the heterogeneous dynamics of fluid, solid, and interface coupling within the same model;
Hidden in the Noise: Unveiling Backdoors in Audio LLMs Alignment Through Latent Acoustic Pattern Triggers
Liang Lin (Institute of Information Engineering Chinese Academy of Sciences), Yang Liu (Nanyang Technological University)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelBenchmarkAudio
π― What it does: This paper investigates backdoor attacks on Audio Large Language Models (ALLM) triggered by audio, proposes the Hidden in the Noise (HIN) framework, and constructs the AudioSafe benchmark;
π― What it does: Propose a framework called DiPVNet based on the atomic dot product operator for learning rotation-invariant point cloud representations under arbitrary rotations.
Hierarchical Dual-Domain Fusion with Frequency-Guided Spatial Modeling for Pan-Sharpening
Huangqimei Zheng (Yunnan University), Xin Jin (Yunnan University)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: Proposed MS-FSNet, a multi-scale frequency-domain and spatial collaborative fusion framework for synthesizing high-resolution multispectral images.
Hierarchical Frequency-Guided Alignment Transformer for Compressed Video Quality Enhancement
Liuhan Peng, Chong Lv (Shandong University)
CodeRestorationTransformerVideo
π― What it does: Propose a Hierarchical Frequency-Guided Alignment Transformer, enhancing multi-frame video quality through frequency domain decomposition and frequency-domain self-attention
Hierarchical Prompt Learning for Image- and Text-Based Person Re-Identification
Linhan Zhou (Kunming University of Science and Technology), Huafeng Li (Nanjing University of Science and Technology)
CodeRetrievalTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper proposes a unified framework called Hierarchical Prompt Learning (HPL), which simultaneously optimizes image-to-image (I2I) and text-to-image (T2I) person re-identification through task-aware prompt learning;
Hierarchical Schedule Optimization for Fast and Robust Diffusion Model Sampling
Aihua Zhu (Macau University of Science and Technology), Shibo He (Macau University of Science and Technology)
CodeGenerationOptimizationDiffusion modelImage
π― What it does: This paper proposes a two-layer optimization framework called HSO that accelerates diffusion model sampling through adaptive scheduling.
Hierarchical Structure-Property Alignment for Data-Efficient Molecular Generation and Editing
Ziyu Fan (Central South University), Lei Deng (Central South University)
CodeDrug DiscoveryTransformerAuto EncoderContrastive LearningBiomedical Data
π― What it does: Designed and implemented a data-efficient molecular generation and editing framework called HSPAG, which leverages hierarchical structure-attribute alignment learning to capture fine-grained relationships between SMILES and properties, and significantly reduces pre-training data requirements through active learning-based sample selection.
π― What it does: Propose HierSearch: a hierarchical deep search framework that utilizes search agents and planning agents from local and network knowledge sources to achieve multi-source information retrieval and reasoning.
π― What it does: Proposed a dual-stream w-Laplacian enhanced HiFi-Mamba architecture for reconstructing high-fidelity MRI images from undersampled k-space.
π― What it does: Utilize a hierarchical guided diffusion model for data augmentation on fine-grained visual classification data, generating synthetic images that are both diverse and preserve class details.
Hilbert Curve-Encoded Rotation-Equivariant Oriented Object Detector with Locality-Preserving Spatial Mapping
Qi Ming (Beijing University of Technology), Yufei Guo (East China Normal University)
CodeObject DetectionAutonomous DrivingConvolutional Neural NetworkTransformerImagePoint CloudBiomedical Data
π― What it does: Proposes HERO-Det, a two-stage arbitrary orientation object detection framework that utilizes Hilbert curves to maintain spatial locality;
π― What it does: Proposed a unified heterogeneous graph pooling framework called HINPool for graph-level classification tasks in molecular and protein attribute prediction.
Ziwei Wang (Zhejiang University), Yong Li (Ant Group)
CodeLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision-Language-Action ModelMultimodality
π― What it does: This paper proposes a history-aware reasoning framework (HAR) that enhances the short-term memory and reasoning capabilities of GUI agents through reflective learning;
π― What it does: Proposes a History-Enhanced Two-Stage Transformer (HETT) framework for unmanned aerial vehicles (UAVs) to accomplish target localization and navigation tasks in large-scale urban environments based on natural language instructions.
HiVA: Self-organized Hierarchical Variable Agent via Goal-driven Semantic-Topological Evolution
Jinzhou Tang (Sun Yat-sen University), Keze Wang (Sun Yat-sen University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmarkChain-of-Thought
π― What it does: Propose a self-organizing multi-agent framework HiVA that can simultaneously evolve agents' semantic behaviors and collaborative topological structures from scratch;
π― What it does: Propose a spiking neural network framework HLML-SNN that integrates Hebbian learning with meta-learning, utilizing a two-stage mechanism to achieve fast and stable continual learning.
HLPD: Aligning LLMs to Human Language Preference for Machine-Revised Text Detection
Fangqi Dai (Shandong University), Zizhuang Deng (Shandong University)
CodeClassificationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
π― What it does: To address the traceability of text generated or rewritten by LLMs, this paper proposes the Human Language Preference Detection (HLPD) method. HLPD first trains a scoring model to align with human writing style through comparative analysis of human original texts and their machine-rewritten versions (HLPO), followed by text detection using Human Language Preference-Conditional Probability Curvature (HLP-CPC).
π― What it does: Propose a hierarchical multi-scale Transformer model (HMformer), which captures multiple periodicities and long-term dependencies in time series through hierarchical resolution branches and cross-scale mixing;
HN-MVTS: HyperNetwork-based Multivariate Time Series Forecasting
Andrey Savchenko (Sber AI Lab), Oleg Kachan (Sber AI Lab)
CodeTime SeriesBenchmark
π― What it does: Proposed the HN-MVTS architecture, which dynamically generates the final layer weights of the multivariate time series prediction model through a hypernetwork based on learnable channel embeddings, integrating the advantages of channel independence (CI) and channel dependence (CD);
HouseTune: Two-Stage Floorplan Generation with LLM Assistance
Ziyang Zong (Sun Yat-sen University), Guang Tan (Sun Yat-sen University)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageTextChain-of-Thought
π― What it does: Propose a two-stage text-to-floorplan generation framework: first, leverage a large language model (LLM) to generate an initial layout (Layout-Init) via chain-of-thought (CoT) prompting, then refine this initial layout into the final layout (Layout-Final) using a dual-noise prior preserving diffusion model (DNPP-Diffusion).
CodeExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: Investigated the impact of Multilingual Alignment (MAPO) on language neurons in large language models, proposed a three-class neuron classification scheme, and analyzed their roles across four reasoning stages.
π― What it does: Using an information theory framework to estimate the hidden dimensions and number of layers of spectral GNNs to alleviate the information compression (over-squashing) problem.
HPSU: A Benchmark for Human-Level Perception in Real-World Spoken Speech Understanding
Chen Li (Sun Yat-sen University), Jian Yin (Sun Yat-sen University)
CodeClassificationRecognitionTransformerLarge Language ModelVision Language ModelVideoMultimodalityBenchmarkAudio
π― What it does: Constructed and publicly released a large-scale human-level perceptual speech understanding benchmark, HPSU, containing over 20,000 multi-language, multi-task expert-verified samples, along with the HPSC dataset comprising 50,000 audio-caption pairs;
π― What it does: Proposed HQ-SVC, a low-resource zero-shot singing voice conversion framework that combines decoupled audio codec, EVA module, DDSP, and diffusion models to achieve high-quality singing voice conversion and super-resolution.
HTG-GCL: Leveraging Hierarchical Topological Granularity from Cellular Complexes for Graph Contrastive Learning
Qirui Ji (National Key Laboratory of Space Integrated Information System Institute of Software Chinese Academy of Sciences), Jiangmeng Li (National Key Laboratory of Space Integrated Information System Institute of Software Chinese Academy of Sciences)