AAAI Conference on Artificial Intelligence · 1442 papers
GLIC: General Format Learned Image Compression
MingSheng Zhou (XiHua University), MingMing Kong (XiHua University)
CodeCompressionAuto EncoderImage
🎯 What it does: A general format learning-based image compression model GLIC is designed, which can unify any channel image into a single channel for efficient compression, and proposes an adaptive attention residual block and a uniform grouping cross-channel context module to achieve high-quality progressive preview.
CodeAnomaly DetectionGraph Neural NetworkGraphFinance Related
🎯 What it does: A graph fraud detection framework based on Attribute-Association Pattern Aggregation (GAAP) is proposed, which extracts attribute patterns using dynamic binning embedding, obtains association patterns through GraphSAGE, and aggregates global graph patterns using global cross-attention, ultimately improving detection performance.
🎯 What it does: The GMAP framework is proposed to achieve a full process from instruction to perception to manipulation, used for the segmentation of joint objects, estimation of joint parameters, and prediction of interactive usability for robotic arms.
🎯 What it does: This paper proposes a new graph contrastive learning framework GTCA, which combines GCN and NodeFormer as two encoders, and generates reliable positive and negative samples through non-random augmentation and topology-based k-NN intersection, constructing a multi-positive sample contrastive loss.
GNS: Solving Plane Geometry Problems by Neural-Symbolic Reasoning with Multi-Modal LLMs
Maizhen Ning (Xi'an Jiaotong-Liverpool University), Kaizhu Huang (University of Liverpool)
CodeTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: The GNS (Geometry Neural-Symbolic) framework is proposed, utilizing multimodal LLMs for knowledge prediction, symbolic analysis, reasoning, and symbolic computation of plane geometry problems, ultimately obtaining answers through a symbolic solver.
GoHD: Gaze-oriented and Highly Disentangled Portrait Animation with Rhythmic Poses and Realistic Expressions
Ziqi Zhou (Institute of Automation, Chinese Academy of Sciences), Dong-Ming Yan (State Key Laboratory of Virtual Reality Technology and Systems, Beihang University)
🎯 What it does: The GoHD framework is proposed, which implements audio-based facial animation capable of generating high-quality, expressive, and controllable speaker videos from any reference identity.
CodeSegmentationDomain AdaptationConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: This paper proposes a test-time adaptation (TTA) method named GraTa, which utilizes gradient alignment and dynamic learning rates to enhance the performance of medical image segmentation models under domain shifts.
Gradient Weight-normalized Low-rank Projection for Efficient LLM Training
Jia-Hong Huang (University of Amsterdam), Evangelos Kanoulas (University of Amsterdam)
CodeComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A method called GradNormLoRP is proposed, which combines weight normalization and low-rank gradient projection to enhance the parameter and memory efficiency of LLM training.
🎯 What it does: A two-stage recommendation and matching optimization framework (RMO) is proposed to address the spatial crowdsourcing problem with subtask dependencies;
Graph Mixture of Experts and Memory-augmented Routers for Multivariate Time Series Anomaly Detection
Xiaoyu Huang (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
CodeAnomaly DetectionGraph Neural NetworkMixture of ExpertsTime Series
🎯 What it does: A multivariate time series anomaly detection framework based on Graph-MoE (Graph Mixture of Experts) is proposed, which can adaptively aggregate structural information from short-range to long-range in multi-layer GNNs and capture global historical features using a memory-enhanced router.
🎯 What it does: This paper presents GraphAvatar, which utilizes graph neural networks to generate 3D Gaussian points and combines graph-guided optimization and 3D-aware post-processing to achieve real-time high-quality rendering of facial avatars, requiring only a 10 MB GNN model without storing massive Gaussian points.
GraphMoRE: Mitigating Topological Heterogeneity via Mixture of Riemannian Experts
Zihao Guo (Beihang University), Jianxin Li (Beihang University)
CodeGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: The GraphMoRE framework is proposed, which dynamically generates personalized mixed curvature spaces for each node through various Riemannian experts and a topology distortion-based gating mechanism, effectively addressing the topological heterogeneity of graphs.
Haoran Zheng (Hong Kong Baptist University), Renchi Yang (Hong Kong Baptist University)
CodeOptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes a graph similarity prediction framework called GRASP, which estimates graph edit distance (GED) and maximum common subgraph (MCS) without cross-graph node interactions.
🎯 What it does: A globally robust ICP registration framework GRICP is proposed, which transforms point clouds into particle sphere clouds and uses multi-core mutual information optimization.
🎯 What it does: This paper proposes the Time Alignment Paradigm (TAP) and Gated Recurrent Spiking Neurons (GRSN), enabling spiking neural networks to perform single-step updates corresponding to single-step decisions in POMDP and multi-agent reinforcement learning, significantly reducing time steps and lowering energy consumption.
🎯 What it does: The GSDiff framework is proposed, which transforms floor plans into structural graphs and divides the process into two stages: node generation and edge prediction, to directly synthesize vector floor plans.
GTDE: Grouped Training with Decentralized Execution for Multi-agent Actor-Critic
Mengxian Li (Institute of Computing Technology Chinese Academy of Sciences), Yongjun Xu (Institute of Computing Technology Chinese Academy of Sciences)
🎯 What it does: A multi-agent reinforcement learning framework called 'Group Training and Decentralized Execution (GTDE)' is proposed and implemented, aiming to address the performance degradation issues of traditional Centralized Training with Decentralized Execution (CTDE) and Decentralized Training with Decentralized Execution (DTDE) in large-scale scenarios.
GTG: Generalizable Trajectory Generation Model for Urban Mobility
Jingyuan Wang (Beihang University), Yudong Li (Beihang University)
CodeGenerationDomain AdaptationGraph Neural NetworkGenerative Adversarial NetworkGraphTime Series
🎯 What it does: A transferable urban trajectory generation model GTG is proposed, utilizing spatial syntax and deep learning to learn city-invariant movement patterns, enabling cross-city trajectory generation.
🎯 What it does: Three methods are proposed: Guided Fusion, Variance-Corrected Fusion, and One-shot Style Alignment, to merge overlapping patches when generating large images using small diffusion models, eliminating seams, blurriness, and incoherent objects.
GuideNER: Annotation Guidelines Are Better than Examples for In-Context Named Entity Recognition
Shizhou Huang (East China Normal University), Xin Alex Lin (East China Normal University)
CodeRecognitionTransformerLarge Language ModelText
🎯 What it does: This paper proposes a GuideNER framework based on LLM, which achieves unsupervised and efficient named entity recognition by summarizing the label patterns of the training set into annotation guidelines.
🎯 What it does: This paper proposes using a Transformer language model to model vehicle diagnostic event sequences to predict the occurrence time and type of future fault patterns.
Harnessing Language Model for Cross-Heterogeneity Graph Knowledge Transfer
Jinyu Yang (Beijing University of Posts and Telecommunications), Chuan Shi (Beijing University of Posts and Telecommunications)
CodeKnowledge DistillationRepresentation LearningGraph Neural NetworkLarge Language ModelSupervised Fine-TuningGraph
🎯 What it does: A cross-heterogeneous graph knowledge transfer framework LMCH is proposed, which converts the meta-paths of heterogeneous graphs into language text, utilizes language models to extract general knowledge, and performs iterative self-supervised training on the target graph.
Harnessing Large Language Models for Knowledge Graph Question Answering via Adaptive Multi-Aspect Retrieval-Augmentation
Derong Xu (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraphRetrieval-Augmented Generation
🎯 What it does: The AMAR framework is proposed, which converts knowledge graph information from multi-perspective (entities, relations, subgraphs) retrieval into prompt embeddings through self-alignment and relevance gating, enhancing the reasoning and answer accuracy of large models in KGQA tasks.
Harnessing Multimodal Large Language Models for Multimodal Sequential Recommendation
Yuyang Ye (Rutgers University), Hui Xiong (Hong Kong University of Science and Technology)
CodeRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMultimodalitySequential
🎯 What it does: This paper proposes a multi-modal large language model enhanced sequence recommendation framework (MLLM‑MSR), which achieves sequential recommendations of multi-modal information (images + text) through a two-stage user preference summarization and supervised fine-tuning.
HC-LLM: Historical-Constrained Large Language Models for Radiology Report Generation
Tengfei Liu (Beijing University of Technology), Baocai Yin (University of Science and Technology of China)
CodeGenerationTransformerLarge Language ModelContrastive LearningTextBiomedical Data
🎯 What it does: The HC-LLM framework is proposed, utilizing historically constrained large language models to achieve longitudinal generation of chest X-ray reports.
🎯 What it does: This paper proposes a fully atomic heterogeneous multi-channel E(3) equivariant graph neural network called HeMeNet, which jointly predicts six tasks: ligand binding affinity (LBA), protein-protein affinity (PPA), as well as the enzymatic classification number (EC) and gene ontology (GO) functions (MF, BP, CC). A unified Protein-MT multi-task benchmark set is constructed.
🎯 What it does: This paper proposes HEP-NAS, which achieves more accurate performance evaluation and higher precision network search by hierarchically partitioning hypernetwork edges and gradually shrinking the search space.
Heterogeneous Graph Neural Network on Semantic Tree
Mingyu Guan (Georgia Institute of Technology), Taesoo Kim (Georgia Institute of Technology)
CodeClassificationGraph Neural NetworkGraph
🎯 What it does: A new heterogeneous graph neural network called HETTREE is designed to simultaneously utilize graph structure and heterogeneous characteristics. It performs offline aggregation of features and labels for all meta-paths before training, constructs a semantic tree structure, and encodes it using subtree attention to obtain node representations and complete the node classification task.
🎯 What it does: The HGSFusion network is proposed to address the issues of sparse radar point clouds and angular errors through the fusion of radar and camera data.
HI-DR: Exploiting Health Status-Aware Attention and an EHR Graph+ for Effective Medication Recommendation
Taeri Kim (Hanyang University), Sang-Wook Kim (Hanyang University)
CodeRecommendation SystemDrug DiscoveryGraph Neural NetworkTransformerMultimodalityGraphBiomedical DataElectronic Health Records
🎯 What it does: A new drug recommendation framework, HI-DR, is proposed, utilizing a health-status-aware attention mechanism and an enhanced EHR graph (EHR Graph+) to improve the accuracy and safety of drug recommendations.
Lei Zhang (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Min Yang
CodeOptimizationAI Code AssistantLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This study investigates warehouse-level code completion and proposes the Hierarchical Context Pruning (HCP) strategy to construct high-quality prompts, significantly reducing context length and improving completion accuracy.
Hierarchical Divide-and-Conquer for Fine-Grained Alignment in LLM-Based Medical Evaluation
Shunfan Zheng (East China Normal University), Linlin Wang (East China Normal University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical Data
🎯 What it does: A hierarchical split-merge framework HDCEval is proposed and implemented for fine-grained evaluation of medical LLMs, including the construction of specialized evaluation criteria, splitting evaluation tasks, and using expert models along with ADTO training based on preference data.
Hierarchical Mixture of Experts: Generalizable Learning for High-Level Synthesis
Weikai Li (University of California Los Angeles), Yizhou Sun (University of California Los Angeles)
CodeGraph Neural NetworkMixture of ExpertsTabular
🎯 What it does: A dual-layer hierarchical Mixture of Experts (MoE) model is proposed to enhance the cross-domain generalization capability of high-level synthesis (HLS) performance prediction.
Hierarchical Multi-Source Uncertainty Aggregation for Interactive Video Captioning
Ervine Zheng (Rochester Institute of Technology), Qi Yu (Rochester Institute of Technology)
CodeGenerationData-Centric LearningTransformerLarge Language ModelVideoText
🎯 What it does: A hierarchical multi-source uncertainty aggregation framework is proposed for interactive video subtitle generation and active learning.
Hierarchical Vector Quantization for Unsupervised Action Segmentation
Federico Spurio (University of Bonn), Juergen Gall (Toyota Motor Europe)
CodeSegmentationAuto EncoderVideo
🎯 What it does: This study proposes an unsupervised video temporal action segmentation method—Hierarchical Vector Quantization (HVQ), which achieves semantic action segmentation of long videos by learning embeddings, a dual-layer codebook, and clustering from frames to sub-actions and then to actions.
HiRED: Attention-Guided Token Dropping for Efficient Inference of High-Resolution Vision-Language Models
Kazi Hasan Ibn Arif (Virginia Tech), Bo Ji (University College Dublin)
CodeComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: This paper proposes HiRED, an early visual token dropping scheme based on CLS attention mechanism for efficient inference of high-resolution Vision-Language Models (VLMs);
🎯 What it does: This paper proposes a navigation trajectory generation framework called HOSER, which synthesizes high-quality human movement trajectories under the conditions of a given starting point, departure time, and destination.
How Do Position Encodings Affect Length Generalization? Case Studies On In-Context Function Learning
Di-Nan Lin (National Cheng Kung University), Hung-Yu Kao (National Tsing Hua University)
CodeMeta LearningTransformerLarge Language ModelTabularSequential
🎯 What it does: In this study, the authors utilized a from-scratch trained GPT-2 model to perform in-context learning (ICL) on synthetic data such as linear regression and Boolean functions, systematically comparing the effects of different position encodings (NoPE, ALiBi, FIRE, Dynamic YaRN) on out-of-distribution (OOD) length extrapolation.
🎯 What it does: This paper studies and compares two model stitching methods—Task Loss Matching and Direct Matching—in terms of their performance in measuring the similarity of internal representations in deep networks, and contrasts them with traditional structural similarity metrics (such as CCA, CKA, OPD).
🎯 What it does: The RPLPO framework is proposed, which utilizes an encoding module to reconstruct learnable high-resolution states and uses PDE loss with a transition module to address the challenge of training physical system models under partial observations.
HSEvo: Elevating Automatic Heuristic Design with Diversity-Driven Harmony Search and Genetic Algorithm Using LLMs
Pham Vu Tuan Dat (Hanoi University of Science and Technology), Huynh Thi Thanh Binh (George Mason University)
CodeOptimizationTransformerLarge Language ModelText
🎯 What it does: This study investigates the introduction of diversity metrics within the LLM-EPS framework and designs the HSEvo framework to balance diversity and target performance, thereby enhancing the effectiveness of automatic heuristic design.
🎯 What it does: A hierarchical self-regulating diffusion model, HSRDiff, is proposed for generating diverse and reliable semantic segmentation hypotheses in safety-critical domains.
🎯 What it does: This paper proposes a dual transformer model called HVDualformer, based on histograms and Vision Transformers, for end-to-end image white balance correction.
Jialiang Tang (Nanjing University of Science and Technology), Chen Gong (Shanghai Jiao Tong University)
CodeClassificationGenerationData SynthesisKnowledge DistillationGenerative Adversarial NetworkImageBiomedical Data
🎯 What it does: A hybrid data-free knowledge distillation framework HiDFD is proposed, which trains a GAN to generate high-quality synthetic data using a very small number of real samples in conjunction with a teacher network, and then trains the student network using a mix of synthetic and real samples, referred to as 'mixed data'.
🎯 What it does: A hybrid decentralized optimization framework is proposed, allowing for collaborative learning on distributed nodes using both gradient (first-order) and non-gradient (zero-order) estimators simultaneously.
HybridReg: Robust 3D Point Cloud Registration with Hybrid Motions
Keyu Du (University of Electronic Science and Technology of China), Shuaicheng Liu (University of Electronic Science and Technology of China)
CodeRecognitionPose EstimationPoint Cloud
🎯 What it does: The HybridReg method is proposed, which combines uncertainty mask learning to achieve robust 3D point cloud registration for mixed motion (rigid background and non-rigid foreground).
🎯 What it does: We propose and implement HYGENE, a diffusion-based hypergraph generation model that can sample realistic hypergraphs from a given data distribution.
🎯 What it does: This paper proposes and implements HyperDefender, a robust defense framework for hyperbolic graph neural networks (Hy-GNN) to resist adversarial attacks and noise on node features and graph structures, and validates its effectiveness in node classification tasks.
Hyperparametric Robust and Dynamic Influence Maximization
Arkaprava Saha (DesCartes Program), Laks V. S. Lakshmanan (National University of Singapore)
CodeOptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes the RIME algorithm, which addresses the robust influence maximization problem based on hyperparameters in dynamic networks with node/edge additions and deletions, and can return a near-optimal seed set at each time step.
IAA: Inner-Adaptor Architecture Empowers Frozen Large Language Model with Multimodal Capabilities
Bin Wang (360 AI Research), Yuhui Yin (360 AI Research)
CodeOptimizationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes an Inner-Adaptor Architecture (IAA) that extends multimodal capabilities by inserting trainable adapter layers within a frozen large language model, while preserving the original NLP performance of the language model.
🎯 What it does: A contrastive learning framework named ICE-T is proposed, treating each column as a separate modality, using inter-column average aggregation as anchors to obtain column-specific and global embeddings through linear time contrastive learning.
ICM-Assistant: Instruction-tuning Multimodal Large Language Models for Rule-based Explainable Image Content Moderation
Mengyang Wu (Chinese University of Hong Kong), Chi-Wing Fu (Chinese University of Hong Kong)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImageTextMultimodalityChain-of-Thought
🎯 What it does: Designed and implemented a rule-based multimodal large language model ICM-Assistant for image content moderation, and constructed the ICM-Instruct dataset that can be generated according to rules;
Identifying Query-Relevant Neurons in Large Language Models for Long-Form Texts
Lihu Chen (Imperial College), Francesca Toni (Imperial College)
CodeTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an architecture-agnostic Query-Relevant Neuron Clustering Attribution (QRNCA) framework for locating query-relevant neurons in large decoder language models, supporting long text answers.
Bin Huang (Tsinghua University), Wenwu Zhu (Tsinghua University)
CodeRetrievalTransformerLarge Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes the Identity-Text Video Semantic Retrieval task (ITVCG), constructs the TVR-IT dataset, and designs the Video-Locator model to achieve joint understanding and retrieval of videos, text, and identity images.
🎯 What it does: The Virtual Dressing (VD) task is proposed, and the IMAGDressing-v1 model is developed to generate customizable portraits with editable clothing images; simultaneously, the IGPair large-scale clothing-outfit pair dataset is constructed, and a comprehensive affinity metric CAMI is designed.
CodeGenerationData SynthesisLarge Language ModelPrompt EngineeringDiffusion modelImageMultimodalityBenchmark
🎯 What it does: The ImageRepainter framework is proposed, which evaluates the generation quality of text-to-image models through the image regeneration task (generating an image with the same content given a reference image) and utilizes a multimodal LLM for image understanding and iterative prompt generation and revision, ultimately assessing similarity through image-to-image comparison.
🎯 What it does: In unsupervised domain adaptation from images to videos, the authors propose a two-stage method: the first stage models outlier frames by introducing an 'outlier' category in the classifier and employing batch nuclear norm maximization along with pseudo outlier class loss; the second stage selects high-quality pseudo labels using a consistency measure based on label propagation and employs FixMatch for semi-supervised video-level learning to alleviate modality gaps.
iMoT: Inertial Motion Transformer for Inertial Navigation
Son Minh Nguyen (University of Twente), Paul Havinga (University of Twente)
CodeTransformerSimultaneous Localization and MappingTime Series
🎯 What it does: A Transformer-based iMoT model is proposed to achieve more accurate inertial positioning through cross-modal information of acceleration and angular velocity.
Implicit Location-Caption Alignment via Complementary Masking for Weakly-Supervised Dense Video Captioning
Shiping Ge (Nanjing University), Qing Gu (Nanjing University)
CodeRecognitionGenerationTransformerLarge Language ModelVideoText
🎯 What it does: A weakly supervised dense video captioning method is proposed, utilizing complementary masking to achieve implicit alignment between event locations and captions, avoiding the traditional complex event proposal process.
Importance Weighting Can Help Large Language Models Self-Improve
Chunyang Jiang (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
CodeGenerationOptimizationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes a DS weight to measure the degree of distribution shift in self-generated data, and combines it with self-consistency to construct the IWSI self-improvement framework;
Improving Cancer Gene Prediction by Enhancing Common Information Between the PPI Network and Gene Functional Association
Chao Deng (Central South University), Jianxin Wang (Central South University)
CodeClassificationDrug DiscoveryGraph Neural NetworkGraphBiomedical Data
🎯 What it does: The DISFusion method is proposed, which integrates multi-omics features, protein-protein interaction networks (PPI), and gene functional associations (hypergraphs) to identify cancer genes, and enhances the common information between PPI and functional associations through cross-view decorrelation loss.
Improving Complex Reasoning over Knowledge Graph with Logic-Aware Curriculum Tuning
Tianle Xia (Wuhan University), Dacheng Tao (Nanyang Technological University)
CodeTransformerLarge Language ModelSupervised Fine-TuningGraph
🎯 What it does: Using large language models (LLM) through instruction tuning and knowledge graph (KG) context, a logic-aware curriculum learning framework (LACT) is proposed to achieve reasoning for complex logical queries (EFO 1) on incomplete KGs.
🎯 What it does: This paper proposes and validates the Synaptic Neural Balance theory, which maintains a balance between input and output weight costs during the training process by proportionally balancing the input and output synaptic weights of individual neurons, thereby accelerating convergence and improving accuracy.
Improving Factuality in Large Language Models via Decoding-Time Hallucinatory and Truthful Comparators
Dingkang Yang (Fudan University), Lihua Zhang (Fudan University)
CodeGenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: By introducing a hallucination and truth comparator during the inference phase, comparative decoding is performed on the next token prediction of large language models, thereby reducing the hallucinated content generated by the model.
🎯 What it does: This paper proposes a decoupling framework that enhances the generalization ability of AI-generated speech detection by extracting domain-independent synthetic speech forgery features.
🎯 What it does: A dynamic max-min optimization framework DM-UAP is proposed to generate universal adversarial perturbations with stronger generalization capabilities across samples and models.
🎯 What it does: Improved transferable adversarial attacks based on integrated gradients by refining the integral path (multi-path, diversification, monotonicity) to enhance the attack success rate.
🎯 What it does: A retrieval-based multimodal social media popularity prediction framework SKAPP is proposed, which enhances prediction accuracy by retrieving relevant UGC and filtering effective information.
Improving Pareto Set Learning for Expensive Multi-objective Optimization via Stein Variational Hypernetworks
Minh-Duc Nguyen (VinUniversity), Dung D. Le (Singapore University of Technology and Design)
CodeOptimizationTabular
🎯 What it does: A method named SVH-PSL is proposed, which combines Stein Variational Gradient Descent (SVGD) with Hypernetwork to achieve Pareto set learning for expensive multi-objective optimization problems through multi-sample sampling and local kernel functions.
Zhangtao Cheng (University of Electronic Science and Technology of China), Fan Zhou (University of Electronic Science and Technology of China)
CodeRetrievalRecommendation SystemTransformerPrompt EngineeringVision Language ModelVideoMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper studies the prediction of micro-video popularity and proposes a retrieval-enhanced contextual prompt framework (ICPF) that improves prediction performance without fine-tuning the parameters of the pre-trained model.
In-depth Analysis of Low-rank Matrix Factorisation in a Federated Setting
Constantin Philippenko (Inria Paris), Laurent Massoulié (Inria Paris)
CodeOptimizationFederated LearningTabular
🎯 What it does: In a federated learning environment, this paper proposes a low-rank matrix factorization algorithm based on random power iteration initialization and local gradient descent.
Infer Human’s Intentions Before Following Natural Language Instructions
Yanming Wan (University of Washington), Natasha Jaques (University of Washington)
CodeRobotic IntelligenceTransformerLarge Language ModelMultimodalityChain-of-Thought
🎯 What it does: The FISER framework is proposed, modeling human intentions as intermediate reasoning steps, separating social reasoning from embodied reasoning to better execute ambiguous natural language instructions.
Infinite-Canvas: Higher-Resolution Video Outpainting with Extensive Content Generation
Qihua Chen (University of Science and Technology of China), Wei Liu (Tencent)
CodeGenerationData SynthesisDiffusion modelVideo
🎯 What it does: A high-resolution video outpainting method based on diffusion models is proposed, capable of achieving 9× content generation while maintaining spatial and temporal consistency.
Information-Theoretic Generative Clustering of Documents
Xin Du (Waseda University), Kumiko Tanaka-Ishii (Waseda University)
CodeGenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: A generative document clustering method is proposed that uses large language models to generate text and implements KL divergence from information theory.
🎯 What it does: A low-light image enhancement method based on initial illumination adaptive compensation and optimization, called IniRetinex, is proposed.
Instruction-guided Multi-Granularity Segmentation and Captioning with Large Multimodal Model
Xu Yuan (Hong Kong Polytechnic University), Jinsong Lan (TAO Technology)
CodeSegmentationGenerationTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: A multi-granularity segmentation and captioning model (MGLMM) is proposed, which can adaptively switch between fine-grained and overall segmentation and description based on user instructions. The MGSCData dataset and a unified SegCap data format are also created.
InstructOCR: Instruction Boosting Scene Text Spotting
Chen Duan (Meituan), Junfeng Luo (Meituan)
CodeRecognitionTransformerVision Language ModelImageText
🎯 What it does: A scene text recognition framework called InstructOCR is proposed, which guides the model to better understand the text in images through human language instructions.
🎯 What it does: Proposes Int*-Match, a pseudo-label selection method that simultaneously considers intra-class compactness and inter-class differences in semi-supervised speaker recognition.
🎯 What it does: Designed and implemented Int2Planner, an intent-driven multimodal motion planner that integrates prediction and planning, capable of providing multiple predicted and planned trajectories in a single forward inference.
Integrating Inference and Experimental Design for Contextual Behavioral Model Learning
Gongtao Zhou (Beijing Institute of Technology), Haoran Yu (Beijing Institute of Technology)
CodeRecommendation SystemAnomaly DetectionOptimizationTabularFinance Related
🎯 What it does: The study investigates how to learn behavioral models in user contexts through sequential experimental design combined with Bayesian inference.
🎯 What it does: This paper proposes a next location recommendation model integrating personalized spatiotemporal clustering, iPCM, to predict the user's next visit point in location-based services.
Integrating Sequence and Image Modeling in Irregular Medical Time Series Through Self-Supervised Learning
Liuqing Chen (Zhejiang University), Lingyun Sun (Zhejiang University)
CodeClassificationRepresentation LearningRecurrent Neural NetworkTransformerGenerative Adversarial NetworkContrastive LearningTime SeriesBiomedical Data
🎯 What it does: A joint learning framework is proposed that simultaneously uses sequence representation and image representation for the classification of irregular medical time series, and three self-supervised learning strategies are designed to fuse the two representations.
Interacted Object Grounding in Spatio-Temporal Human-Object Interactions
Xiaoyang Liu (Shanghai Jiao Tong University), Yong-Lu Li (Shanghai Jiao Tong University)
CodeObject DetectionSegmentationTransformerVision Language ModelVideoPoint CloudBenchmark
🎯 What it does: A new open-world human-object interaction video benchmark GIO is proposed, and an object localization task is defined on it. Subsequently, a 4D-QA framework based on SAM candidates and 4D question answering is introduced to address this task.
Interpretable Failure Detection with Human-Level Concepts
Kien X. Nguyen (University of Delaware), Xi Peng (University of Delaware)
CodeAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImage
🎯 What it does: A post-processing failure detection framework based on human visual concepts, ORCA, is proposed, which uses concept-activated ordinal ranking to estimate model confidence and can explain the reasons for errors.
Interpretable Solutions for Multi-Physics PDEs Using T-NNGP
Lulu Cao (Xiamen University), Min Jiang (Xiamen University)
CodeOptimizationExplainability and InterpretabilityRecurrent Neural NetworkReinforcement LearningTime SeriesPhysics Related
🎯 What it does: A T-NNGP method is proposed, which combines traditional numerical solving with deep reinforcement learning in a hybrid genetic programming approach to automatically generate interpretable symbolic expressions that satisfy multi-physics PDE systems.
Interweaving Memories of a Siamese Large Language Model
Xin Song (East China Normal University), Wei Lu (Wuhan University)
CodeClassificationGenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a model-agnostic parameter-efficient fine-tuning framework called IMSM, which enhances the adaptability to downstream tasks by constructing a dual-tower (siamese) LLM and interweaving the memories of the two towers, while alleviating catastrophic forgetting.
Intra and Inter Parser-Prompted Transformers for Effective Image Restoration
Cong Wang (Shenzhen Campus of Sun Yat-sen University), Wei Wang (Dalian University of Technology)
CodeRestorationTransformerImage
🎯 What it does: A PPTformer model is proposed, which generates parsing information based on the visual foundation model (SAM) to guide image restoration, integrating a Transformer structure with parsing prompts.
Inverse Reinforcement Learning by Estimating Expertise of Demonstrators
Mark Beliaev (University of California Santa Barbara), Ramtin Pedarsani (University of California Santa Barbara)
CodeReinforcement LearningTabular
🎯 What it does: A new inverse reinforcement learning framework IRLEED is proposed to learn true rewards and policies from suboptimal and heterogeneous demonstrations.
InverseCoder: Self-improving Instruction-Tuned Code LLMs with Inverse-Instruct
Yutong Wu (Institute of Computing Technology), Yunji Chen
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A self-supervised data augmentation method called Inverse-Instruct is proposed, which expands the instruction-tuning dataset for code LLMs by transforming existing code snippets into new natural language instructions and pairing them with the original code, resulting in the training of the InverseCoder series models.
Invertible Projection and Conditional Alignment for Multi-Source Blended-Target Domain Adaptation
Yuwu Lu (South China Normal University), Xue Hu (South China Normal University)
CodeDomain AdaptationFlow-based ModelImage
🎯 What it does: This paper proposes an Inverse Projective and Conditional Alignment method (IPCA) for Multi-Source Blended Domain Adaptation (MBDA). It maps source domain and blended target domain features through reversible projection, enhances model robustness using projection consistency regularization and conditional entropy maximization, and achieves domain-invariant feature learning through CKB metric-driven unsupervised adversarial learning.
🎯 What it does: This paper proposes a Minimization of Abstract State based on Relative Spatial Relationships (MARC), which transforms multi-agent observations into a graph structure and uses R-GCN for encoding, achieving cooperative learning within a centralized training and distributed execution actor-critic framework.
Investigating the Security Threat Arising from “Yes-No” Implicit Bias in Large Language Models
Yanrui Du (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
CodeAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper quantifies and verifies the existence of implicit bias in 'yes-no' responses by statistically analyzing the probability differences in large language models (LLMs) when answering 'yes' and 'no'. It proposes a covert jailbreak attack method based on this bias (IB-ICM), which induces the model to convert malicious instructions that should be rejected into affirmative responses by inserting high 'yes' inclination instructions into the attack context.
🎯 What it does: This paper proposes the Image-enhanced Prompt Decoding Network (IPDN), which enhances 3D point cloud semantics through multi-view image information and utilizes task-driven prompts to guide the decoder to focus on the target object, thereby achieving 3D Referring Expression Segmentation.
Is Sarcasm Detection a Step-by-Step Reasoning Process in Large Language Models?
Ben Yao (University of Copenhagen), Jing Qin (Hong Kong Polytechnic University)
CodeTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the SarcasmCue framework, integrating four prompt sub-methods (CoC, GoC, BoC, ToC) to enable large language models to reason step-by-step and concurrently synthesize multiple cues in sarcasm detection.
Xiujie Song (Shanghai Jiao Tong University), Kenny Q. Zhu (University of Texas at Arlington)
CodeTransformerLarge Language ModelVision Language ModelImageChain-of-Thought
🎯 What it does: This paper proposes the Image Semantic Assessment (ISA) task, constructs the first ISA dataset containing two scoring metrics: entity complexity and semantic complexity, and designs the VLISA method based on visual language models.