AAAI Conference on Artificial Intelligence Β· 2140 papers
HuiduRep: A Robust Self-Supervised Framework for Learning Neural Representations from Extracellular Recordings
Feng Cao (Beihang University), Wei Shi (Beihang University)
CodeRepresentation LearningTransformerAuto EncoderContrastive LearningBiomedical Data
π― What it does: Propose HuiduRep, a self-supervised representation learning framework that combines contrastive learning with denoising autoencoders, for extracting robust synaptic waveform features and achieving unlabeled peak sorting;
Human Cognition Inspired RAG with Knowledge Graph for Complex Problem Solving
Yao Cheng, Xiang Li (Tencent Youtu Lab)
CodeTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose CogGRAG, a knowledge graph-based retrieval-augmented generation framework for complex KG question answering, simulating human cognition through tree-like mind maps for problem decomposition, structured retrieval, and self-verification.
π― What it does: Studied human action synthesis based on a unified scene representation (SSO) combining semantics and structure in 3D scenes, and proposed the SSOMotion framework.
π― What it does: Propose the HuMo framework to achieve human video generation with text, reference image, and audio multimodal collaborative control; construct an incomplete yet complementary multimodal dataset; adopt two-stage progressive training and stage-adaptive Classifier-Free Guidance (CFG) to realize fine-grained multimodal control.
Human-Corrected Labels Learning: Enhancing Labels Quality via Human Correction of VLMs Discrepancies
Zhongnian Li (China University of Mining and Technology), Xinzheng Xu (China University of Mining and Technology)
CodeClassificationRepresentation LearningData-Centric LearningTransformerVision Language ModelContrastive LearningImageMultimodality
π― What it does: This paper proposes a Human-Corrected Labels (HCL) setup, which leverages label consistency predicted by multi-modal vision-language models (VLM) to filter samples requiring manual correction, thereby significantly reducing manual annotation costs while ensuring label quality.
HumanSense: From Multimodal Perception to Empathetic Context-Aware Responses Through Reasoning MLLMs
Zheng Qin, Le Wang (Xi'an Jiaotong University)
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmarkAudio
π― What it does: This paper proposes the HumanSense benchmark for systematically evaluating the human-centric perception, contextual understanding, and response strategy capabilities of multimodal large language models in comprehension and interaction.
Hybrid Restricted Master Problem for Boolean Matrix Factorisation
Ellen Visscher (University of Oxford), Christopher Yau (University of Oxford)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyTabularBiomedical DataBenchmark
π― What it does: Proposed the bfact algorithm, achieving low-rank Boolean matrix decomposition via hybrid combinatorial optimization (generating candidate factors based on prior clustering);
Yitong Huang (Xiamen University), Cheng Wang (Xiamen University)
CodeLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextBenchmark
π― What it does: Propose the HotMoE framework, which combines LoRA with Mixture of Experts, using hierarchical routing to achieve multi-task instruction fine-tuning.
Hybrid-DMKG: A Hybrid Reasoning Framework over Dynamic Multimodal Knowledge Graphs for Multimodal Multihop QA with Knowledge Editing
Li Yuan (South China University of Technology), Tao Wang (King's College London)
CodeGraph Neural NetworkTransformerLarge Language ModelVision Language ModelMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the MMQAKE benchmark to evaluate knowledge editing and reasoning processes in multi-modal multi-hop question answering; simultaneously design the Hybrid-DMKG framework to dynamically maintain a multi-modal knowledge graph and combine LLM decomposition, cross-modal retrieval, and hybrid reasoning to improve answer quality.
π― What it does: Propose an unsupervised domain adaptation framework named HARL, which integrates high-quality near-eye images and low-quality facial images to accurately regress 3D gaze direction from a single facial image.
HyFI: Hyperbolic Feature Interpolation for Brain-Vision Alignment
Sangmin Jo (Korea University), Heung-Il Suk (Korea University)
CodeRepresentation LearningContrastive LearningMultimodalityBiomedical Data
π― What it does: This paper proposes a framework (HyFI) that interpolates semantic and perceptual visual features in hyperbolic space to better align brain signals with visual features, addressing the modal gap and feature entanglement issues.
Hyper-Opinion Vagueness Quantification for Robust Multimodal Learning
Disen Hu (Tongji University), Xing Xu (Tongji University)
CodeAdversarial AttackMultimodality
π― What it does: Propose a new robust multimodal learning framework HOVQ, which can maintain high accuracy even when multimodal data is affected by noise or adversarial attacks.
π― What it does: Propose a hierarchical clustering method HypCSE that optimizes continuous structural entropy (CSE) in hyperbolic space, combining graph structure learning and contrastive learning to achieve end-to-end differentiable hierarchical clustering.
Hyperbolic Hierarchical Alignment Reasoning Network for Text-3D Retrieval
Wenrui Li (Harbin Institute of Technology), Xiaopeng Fan (Harbin Institute of Technology)
CodeRetrievalContrastive LearningTextPoint Cloud
π― What it does: Proposed the H2ARN model, which maps text and 3D point clouds into the hyperbolic space of the Lorentz model, and achieves cross-modal retrieval through hierarchical ranking loss and contribution-aware aggregation.
π― What it does: Proposed the first large-scale hyperspectral concealed object detection benchmark, HyperCOD, and designed the HSC-SAM framework based on SAM, achieving spatial-spectral decomposition and adaptive prompting for hyperspectral images to complete concealed object segmentation.
π― What it does: Propose a diagnostic framework named HyperDiag that integrates time evolution and regional hypergraphs for dynamic functional network modeling and classification of brain diseases using rs-fMRI.
HyperGOOD: Towards Out-of-Distribution Detection in Hypergraphs
Tingyi Cai (Zhejiang Normal University), Zhonglong Zheng (Zhejiang Normal University)
CodeAnomaly DetectionGraph
π― What it does: This paper proposes the hypergraph-oriented discrete distribution detection task (HOOD) and introduces the HyperGOOD framework to realize this task.
HyperSHAP: Shapley Values and Interactions for Explaining Hyperparameter Optimization
Marcel Wever (Leibniz University Hannover), Marius Lindauer (Leibniz University Hannover)
CodeExplainability and InterpretabilityHyperparameter SearchBenchmark
π― What it does: This paper proposes HyperSHAP, a game-theoretic explanation framework based on Shapley values and interactions, for local and global explainability analysis of the hyperparameter optimization (HPO) process.
I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders
Andrey V. Galichin (AIRI), Ivan Oseledets (AIRI)
CodeExplainability and InterpretabilityLarge Language ModelAuto EncoderText
π― What it does: Researchers decomposed the internal activations of large language models using sparse autoencoders (SAE) and constructed an automatic metric called ReasonScore to identify interpretable features related to reasoning.
π― What it does: Proposed the I2E framework, which can convert static images into high-fidelity event streams in real-time, thereby addressing the scarcity of event data in SNN training.
ICAD-LLM: One-for-All Anomaly Detection via In-Context Learning with Large Language Models
Zhongyuan Wu (Beihang University), Changqing Ma (Capinfo Co., Ltd.)
CodeAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextMultimodalityTabularTime Series
π― What it does: This paper proposes a unified anomaly detection framework called ICAD-LLM, which leverages large language models to achieve consistent processing and anomaly determination for multi-modal data by providing normal sample context during inference.
ICL-Router: In-Context Learned Model Representations for LLM Routing
Chenxu Wang (Fudan University), Shuyue Hu (Beijing Institute of Technology)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelTextBenchmark
π― What it does: Proposes a context vector-based model routing framework called ICL-Router, which can dynamically assign the most suitable LLM for each query and supports adding new models without training.
ID-Splat: Propagating Object Identities for Segmenting 3D Aerial-view Scenes
Yijing Wang (Xidian University), Jingjing Ma (Xidian University)
CodeSegmentationGaussian SplattingPoint Cloud
π― What it does: Proposed a 3D aerial view segmentation framework called ID-Splat based on multi-view object identity, which assigns semantics to 3D Gaussian Splatting points through Mask-Object Tracking and Object Integration & Propagation.
π― What it does: Developed the IdealTSF framework, integrating three stages: negative sample pre-training, positive sample generation, and ECOS optimization, to enhance the robustness and accuracy of time series forecasting.
Identifying and Analyzing Performance-Critical Tokens in Large Language Models
Yu Bai (Beijing Institute Of Technology), Jackie Chi Kit Cheung (Mila Quebec Artificial Intelligence Institute)
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper conducts a fine-grained classification of prompts in large language models (LLMs) during few-shot learning (ICL), identifying and analyzing which types of tokens (template tokens, stop tokens, content tokens) most directly affect model performance, proposes the concept of 'performance-critical tokens,' and investigates their characteristics (semantic meaning, repetitiveness, structural features).
CodeAnomaly DetectionExplainability and InterpretabilityTransformerVision Language ModelImageMultimodalityChain-of-Thought
π― What it does: Designed a personalized vision-language model that integrates low-level visual cues with high-level semantic consistency for facial forgery detection and providing interpretable explanations.
IdentityStory: Taming Your Identity-Preserving Generator for Human-Centric Story Generation
Donghao Zhou (Hong Kong University of Science and Technology), Pheng-Ann Heng (Hong Kong University of Science and Technology)
CodeSegmentationGenerationData SynthesisTransformerVision Language ModelDiffusion modelContrastive LearningImageTextBenchmark
π― What it does: Propose the IdentityStory framework to achieve text-based continuous image generation, ensuring identity consistency of characters across multiple frames.
π― What it does: IGIANet aims to address weak alignment and modality imbalance issues in UAV RGB-IR object detection by constructing a three-module unified framework comprising illumination-guided frequency domain modulation, frequency domain difference enhancement, and implicit alignment dynamic fusion.
IGT4ETH: An Isotropic Pre-trained Graph Transformer for Ethereum Account Classification
Ao Liu, Qiang Duan (Central University of Finance and Economics)
CodeClassificationTransformerContrastive LearningGraphFinance Related
π― What it does: Propose IGT4ETH, a pre-trained graph Transformer for Ethereum account classification, which explicitly models transaction network topology and eliminates embedding directional imbalance through post-processing;
iMAD: Intelligent Multi-Agent Debate for Efficient and Accurate LLM Inference
Wei Fan (Virginia Tech), Bo Ji (Virginia Tech)
CodeExplainability and InterpretabilityComputational EfficiencyLarge Language ModelAgentic AIPrompt EngineeringTextMultimodalityChain-of-Thought
π― What it does: Propose an intelligent multi-agent debate framework called iMAD, which significantly reduces token consumption while maintaining or improving answer accuracy by performing structured self-critique after a single LLM output and deciding whether to trigger a multi-agent debate based on linguistic features.
π― What it does: Propose a plug-and-play (PnP) framework based on flow-matching generative models, which employs the PrimalβDual Hybrid Gradient (PDHG) algorithm to handle β1 and β2 data fidelity terms for non-Gaussian noise, enabling multi-task image restoration.
π― What it does: Proposed a text-to-video generation method called ImagerySearch based on adaptive test-time search, which enhances video quality under long-distance semantic prompts.
ImageSet2Text: Describing Sets of Images Through Text
Piera Riccio (University of Amsterdam), Nuria M Oliver (Johannes Kepler University Linz)
CodeGenerationExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose ImageSet2Text, capable of automatically generating natural language descriptions for large-scale image collections.
Implicit Neural Representation with Multi-Scale Sine Activation
Jufeng Han (Chinese Academy of Sciences), Yan Pang (Chinese Academy of Sciences)
CodeRepresentation LearningImageVideoMesh
π― What it does: Proposed a multi-scale sine activation function (MSA) to enhance the modeling capability of implicit neural representations (INR) in high-frequency details and multi-scale structures.
Improved Algorithms for Trip-Vehicle Assignment in Ride-Sharing
Jingyang Zhao (University of Electronic Science and Technology of China), Yonghang Su (University of Electronic Science and Technology of China)
CodeOptimizationComputational Efficiency
π― What it does: This paper proposes three new approximation algorithms for the passenger-vehicle assignment problem (RSAP) and provides improved approximation ratios for two cases: n = mk and n < mk.
π― What it does: Proposed a knowledge-enhanced masked image generation framework, KA-MIG, which improves the generation quality of Masked Image Generation by introducing three token-level knowledge graphs based on training data (co-occurrence graph, semantic similarity graph, position-token incompatibility graph).
π― What it does: This paper proposes a one-pass streaming algorithm for fair k-center clustering, which supports setting an upper limit on the number of centers for each group. It provides a 3+Ξ΅ approximation algorithm for semi-structured data streams (arriving in group order) and further extends its application to offline scenarios.
π― What it does: Proposes an adaptive method based on a learnable BN layer, combining geometric confidence maximization (GCM) and entropy minimization (EM) losses, and achieving robust test-time adaptation through a two-stage adaptive strategy with semantic consistency.
π― What it does: Propose the CTMRL framework, leveraging cross-task context to enhance the generalization ability of offline meta-reinforcement learning.
Improving Implicit Discourse Relation Recognition with Natural Language Explanations from LLMs
Heng Wang (East China Jiaotong University), Changxing Wu (East China Jiaotong University)
CodeClassificationRecognitionExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: Leverage large language models to generate natural language explanations, and jointly train relation prediction and explanation generation in a lightweight IDRR model to enhance performance and explainability.
Improving Large Molecular Language Model via Relation-aware Multimodal Collaboration
Jinyoung Park (Korea Advanced Institute of Science and Technology), Hyunwoo J. Kim (Korea Advanced Institute of Science and Technology)
CodeDrug DiscoveryGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityGraphBiomedical Data
π― What it does: Propose the CoLLaMo model, which synergistically encodes three modalitiesβ1D SELFIES, 2D molecular graphs, and 3D spatial conformationsβinto a unified token space interpretable by language models through a relation-aware multi-layer projector; simultaneously introduce molecule-centered evaluation metrics (CHARM/RCHARM) and a GPT-4o evaluator to detect hallucinations and description quality in model outputs.
Improving Region Representation Learning from Urban Imagery with Noisy Long-Caption Supervision
Yimei Zhang (Zhejiang University of Technology), Xiangjie Kong (Zhejiang University of Technology)
CodeKnowledge DistillationRepresentation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextTabular
π― What it does: This paper proposes a cross-modal pre-training framework named UrbanLN, aiming to utilize long-text descriptions generated by multi-modal large language models to achieve fine-grained semantic alignment between urban street scenes and satellite images, while suppressing noise in urban region representation learning.
Improving the Accuracy of Dense Retrieval on the Quantized Indexes via Gradient Optimization of the Target Embeddings
Cong Tan (Shanghai Jiao Tong University), Tao Fang (Shanghai Jiao Tong University)
CodeRetrievalTextBenchmark
π― What it does: Propose an scalable training method that significantly improves dense retrieval performance under quantized index (QHNSW) by directly updating the cached target embeddings via gradient and combining similarity-based approximate negative sampling.
π― What it does: Proposed and implemented two hard-label attack algorithms based on zeroth-order optimization, ARS-OPT and PARS-OPT, leveraging the Nesterov acceleration idea and momentum mechanism. By using lookahead to pre-estimate gradients when searching for optimal ray directions, query efficiency is significantly improved.
Jaesung Lim (University of Seoul), Jong-June Jeon (University of Seoul)
CodeData-Centric LearningAuto EncoderTabular
π― What it does: Propose U-VAE, a missing value uncertainty modeling method based on VAE, which can provide reliable statistical inference after single or multiple imputations.
In-Token Rationality Optimization: Towards Accurate and Concise LLM Reasoning via Self-Feedback
Mingye Zhu, Yongdong Zhang (University Of Science And Technology Of China)
CodeOptimizationExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmarkChain-of-Thought
π― What it does: Propose the InTRO framework for LLM chain-of-thought reasoning, aligning the model's own answer conditional posterior with generation strategies to achieve token-level exploration and self-feedback, enhancing reasoning accuracy and conciseness.
Incoherence as Oracle-less Measure of Error in LLM-Based Code Generation
Thomas Jean-Michel Valentin (ENS Paris-Saclay), Marcel BΓΆhme (Max Planck Institute for Security and Privacy)
CodeAI Code AssistantTransformerLarge Language ModelText
π― What it does: Proposed an oracle-free code correctness evaluation metric called incoherence to estimate the error probability of programs generated by LLMs.
Incomplete Multi-View Unsupervised Federated Feature Selection via Cooperative Particle Swarm Optimization and Tensor-Aligned Learning
Zhiwei Ye (Hubei University of Technology), Jixin Zhang (Hubei University of Technology)
CodeOptimizationFederated LearningMultimodality
π― What it does: Proposes a federated learning-based 'IMUFFS' framework for unsupervised feature selection in the presence of multi-view missing data.
π― What it does: Introduce weighted Chamfer distance into the ColBERT multi-vector retrieval framework to enhance retrieval recall by leveraging token importance.
Incremental Data-Driven Policy Synthesis via Game Abstractions
Irmak SaΔlam (Max Planck Institute for Software Systems), Anne-Kathrin Schmuck (Max Planck Institute for Software Systems)
CodeOptimizationSequential
π― What it does: Propose an incremental data-driven abstraction and synthesis framework that can learn upper and lower bounds of reachable sets from sampled data in unknown stochastic dynamical systems, construct finite stochastic (2.5-player) games, and incrementally update winning regions and control strategies as new data arrives.
IndexTTS2: A Breakthrough in Emotionally Expressive and Duration-Controlled Auto-Regressive Zero-Shot Text-to-Speech
Siyi Zhou (bilibili), Jingchen Shu (bilibili)
CodeGenerationTransformerLarge Language ModelFlow-based ModelTextAudio
π― What it does: This paper proposes IndexTTS2, a zero-shot emotionally controllable, precisely duration-controllable autoregressive text-to-speech model.
Indoor Multi-View Radar Object Detection via 3D Bounding Box Diffusion
Ryoma Yataka (Mitsubishi Electric Corporation), Ryuhei Takahashi (Mitsubishi Electric Research Laboratories)
CodeObject DetectionDiffusion modelPoint Cloud
π― What it does: Propose a multi-perspective indoor radar object detection framework REXO, which achieves object detection through 3D bounding box diffusion in radar space.
π― What it does: This paper proposes the SpecGR framework, which combines generative recommendation with an induced draft-validation mechanism, enabling the model to recommend new items in an incremental, unlabeled setting.
π― What it does: Propose an online two-threshold hierarchical reasoning (H2T2) strategy for cost-sensitive inference and dynamic offloading to remote models on edge devices for binary classification tasks.
Inferring Heterogeneous Private Valuations from Offline Market Data via Entropic Risk-Sensitive Utility Maximization
Xingyu Qian (Beijing Institute of Technology), Haoran Yu (Beijing Institute of Technology)
CodeOptimizationRecurrent Neural NetworkReinforcement LearningTabularTime SeriesFinance Related
π― What it does: Propose a method to infer heterogeneous private valuations of participants in continuous double auctions through entropy risk-sensitive utility maximization.
Inferring Implicit Goals Across Differing Task Models
Silvia Tulli (Sorbonne University), Sarath Sreedharan (Colorado State University)
CodeReinforcement Learning
π― What it does: Proposed an algorithm to infer and align implicit subgoals across different task models, utilizing bottleneck state identification and query MDP to minimize query costs.
InfiGUI-G1: Advancing GUI Grounding with Adaptive Exploration Policy Optimization
Yuhang Liu (Zhejiang University), Fei Wu (Zhejiang University)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMultimodality
π― What it does: Proposed an AEPO framework based on multi-answer generation and adaptive exploration reward to improve semantic alignment of multimodal large language models in GUI localization tasks.
InfoDecom: Decomposing Information for Defending Against Privacy Leakage in Split Inference
Ruijun Deng (Fudan University), Qiang Duan (Pennsylvania State University)
CodeSafty and PrivacyConvolutional Neural NetworkImage
π― What it does: Propose the InfoDecom method, which mitigates data reconstruction attacks and reduces privacy leakage by decomposing redundant information during segmentation inference.
π― What it does: Propose InfoQ, a mixed-precision quantization framework based on information flow, which estimates the impact of hierarchical quantization on global information flow through a single forward pass and allocates bit-widths via integer linear programming;
π― What it does: This study proposes RISE-DDI, an interpretable subgraph extraction framework based on deep reinforcement learning, for predicting drug-drug interactions (DDI) and explaining their mechanisms.
π― What it does: Propose a drone target detection framework that utilizes infrared images as privileged information during training. It employs cross-modal vector quantization (CMVQ) to predict and reconstruct infrared codebook indices from RGB features, enabling the acquisition of infrared complementary features without infrared input during inference.
Instance Dependent Testing of Samplers Using Interval Conditioning
Rishiraj Bhattacharyya (University of Birmingham), Sayantan Sen (National University of Singapore)
CodeAnomaly DetectionComputational Efficiency
π― What it does: Proposes an instance-dependent sampler verification method based on the interval conditioning (interval conditioning) model, which can efficiently verify the distribution correctness of samplers on infinite domains, and implements a practical tester Lachesis for inverse transform samplers (e.g., geometric, binomial, Poisson distributions).
π― What it does: Proposed a method generating diverse training instances through latent space reverse engineering (LSRE), constructing a new Diverse-BBO benchmark to enhance the generalization capability of Meta-Black-Box Optimization (MetaBBO).
CodeGenerationRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningVideoMultimodalityAudio
π― What it does: Proposes InstructDubber, an instruction-based movie dubbing alignment method that generates detailed speaking rate and emotion instructions via multimodal large language models, and achieves lip-sync and emotion alignment without visual preprocessing through instruction-driven duration extraction and emotion calibration modules.
CodeRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Construct a unified and scalable Oracle Bone character set named InteChar, and build the OracleCS corpus based on this character set along with expert annotations and LLM-assisted generation, thereby enhancing the training and evaluation of ancient Chinese language models.
Integrating Diverse Assignment Strategies into DETRs
Yiwei Zhang (Chinese Academy of Sciences), Zhipeng Zhang (Shanghai Jiao Tong University)
CodeObject DetectionTransformerImage
π― What it does: Insert a lightweight LoRA branch into the DETR decoder to parallelly integrate various one-to-many label assignment strategies, enhancing the diversity of training supervision.
π― What it does: Proposed a plug-and-play image restoration framework based on diffusion models, leveraging the gGSM model for noise modeling and employing the βqβ data fidelity term to achieve robust denoising against arbitrary noise (including impulse noise);
π― What it does: Designed and implemented a code generation framework called RoutingGen based on difficulty-aware dynamic routing, using few-shot prompts for simple tasks and Intention Chain-of-Thought (ICoT) two-phase reasoning for complex tasks.
π― What it does: This paper proposes a pedestrian trajectory prediction framework based on diffusion models, which can simultaneously model short-term fine-tuning intentions and long-term goal intentions, and achieve more accurate trajectory generation through soft mask guidance and residual noise correction.
Intention-Guided Cognitive Reasoning for Egocentric Long-Term Action Anticipation
Qiaohui Chu (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelVision-Language-Action ModelVideoChain-of-Thought
π― What it does: Propose a two-stage long-term action anticipation framework named INSIGHT, which first extracts semantically rich visual features from hand-object interaction (HOI) regions and uses a verb-noun co-occurrence matrix for semantic correction, then employs a reinforcement learning-based explicit cognitive reasoning module (think β reason β answer) on a large vision-language model to complete action prediction.
π― What it does: Propose the IntentMotion framework to generate 3D human motion that aligns with natural language instructions, significantly enhancing the realism of physical contact.
Inter-Client Dependency Recovery with Hidden Global Components for Federated Traffic Prediction
Hang Zhou (Nanjing University of Science and Technology), Chen Gong (Shanghai Jiao Tong University)
CodeFederated LearningGraph Neural NetworkTime Series
π― What it does: Proposed a traffic flow prediction method FedHINT under the federated learning framework, which generates proxy nodes from hidden global components within each client's internal data to restore cross-client dependencies and improve prediction accuracy.
π― What it does: Propose the ELECTOR framework to address the interest drift and computational complexity issues in logical reasoning methods for long-sequence recommendations.
Intermediate N-Gramming: Deterministic and Fast N-Grams for Large N and Large Datasets
Ryan R. Curtin (Booz Allen Hamilton), Priyanka Ranade (Laboratory for Physical Sciences)
CodeComputational EfficiencyTextBiomedical Data
π― What it does: This paper proposes a multi-stage algorithm called Intergrams for rapidly and deterministically retrieving the top-k most frequent n-grams from large-scale datasets.
InterMoE: Individual-Specific 3D Human Interaction Generation via Dynamic Temporal-Selective MoE
Lipeng Wang (Beihang University), Lu Sheng (Beihang University)
CodeGenerationTransformerMixture of ExpertsDiffusion modelText
π― What it does: Propose a text-driven 3D human interactive generation framework called InterMoE based on dynamic time selection Mixture-of-Experts (MoE), which can achieve high semantic fidelity while preserving individual identity features.
Shuyi Zhang (University of Science and Technology of China), Xiang Wang (University of Science and Technology of China)
CodeExplainability and InterpretabilityRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAuto EncoderText
π― What it does: This study proposes the Sparse Autoencoder-Enhanced Reward Model (SARM), which maps LLM hidden states to a sparse, disentangled feature space by embedding a pre-trained sparse autoencoder into the reward model, and uses a linear head to aggregate these features to obtain interpretable reward scores.
Interpreting Fedspeak with Confidence: A LLM-Based Uncertainty-Aware Framework Guided by Monetary Policy Transmission Paths
Rui Yao (Hong Kong University of Science and Technology Guangzhou), Hao Wang (Hong Kong University of Science and Technology Guangzhou)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextFinance Related
π― What it does: This paper proposes a 'Fedspeak' parsing framework based on large language models (LLMs), integrating monetary policy transmission paths for domain reasoning and incorporating a dynamic uncertainty decoding module.
Intervention-Aware Time Series Modeling: Capturing and Evaluating Feature Dependencies
Ibrahim Delibasoglu (Linkoping University), Mattias Tiger (Linkoping University)
CodeExplainability and InterpretabilityAuto EncoderTime Series
π― What it does: Designed a conditional variational autoencoder, CFOR-VAE, for interpretable reconstruction when performing feature-level interventions in multivariate time series.
π― What it does: This paper proposes a lightweight few-shot OOD detection framework named IIM (Intra-Image Mining), which selects the top k local patches most relevant to the class prototype within a single image as positive samples to optimize local feature prototypes. It also selects the top k patches from the remaining patches that have high similarity to the prototype but mismatch the category as ID-like OOD samples to construct OOD prototypes. Meanwhile, it introduces a bidirectional reasoning strategy called Symmetric Maximum Concept Matching (S-MCM), which considers both image-text and text-image similarity.
Introducing Visual Scenes and Reasoning: A More Realistic Benchmark for Spoken Language Understanding
Di Wu (Xinjiang University), Xuelong Li (Institute of Artificial Intelligence of China Telecom (TeleAI))
CodeGenerationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmarkAudio
π― What it does: Constructed the first voice-language understanding dataset VRSLU that simultaneously includes user environment images and reasoning processes, and proposed the LR-Instruct instruction template to let the model first perform label prediction and then generate reasoning.
Intuitive Thinking: Expanding Large Language Modelsβ Thinking for Rapid Decision-Making on Candidate Corrections in Chinese Grammar Error Correction
Lintao Long (Guizhou University), Qihang Fu (Guizhou University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Propose a lightweight Expanding Intuitive Thinking Model (ExIT), which injects confidence information from large language models (LLMs) into the embedding space of error correction candidates and performs joint evaluation of source sentences and correction candidates at a global level, mimicking human intuitive rapid decision-making processes;
π― What it does: Propose a framework called IC-MOL for conditional molecular generation under distribution drift conditions, combining invariant learning with graph diffusion models;
π― What it does: Proposed the I-Mem framework, which learns transferable memory behavior models without requiring environmental labels by leveraging self-supervised contrastive learning and decorrelation constraints.
Investigating Data Pruning for Pretraining Biological Foundation Models at Scale
Yifan Wu (Chinese University of Hong Kong), Yu Li (Chinese University of Hong Kong)
CodeData-Centric LearningBiomedical Data
π― What it does: This paper proposes a post-hoc influence-guided data reduction framework to select a small yet effective subset during the pre-training of biological foundation models (BioFMs), maintaining or even improving model performance under extreme data pruning (99%).
Investigating Prosocial Behavior Theory in LLM Agents Under Policy-Induced Inequities
Yujia Zhou (Tsinghua University), Yiqun Liu (Tsinghua University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelAgentic AITextTabular
π― What it does: Simulate prosocial behaviors in multiple scenarios using LLM agents through the PROSIM framework, and systematically evaluate their sensitivity to unfair policies.
π― What it does: Designed and implemented a visualization-based 3D detection adversarial creation attack (AdvRoad) that places naturally appearing adversarial stickers on roads to induce false targets.
IPFormer: Instance Prompt-guided Transformer for Multi-modal Multi-shot Video Understanding
Yujia Liang (JD Explore Academy Deepeleph Intelligent Technology), Zhicheng Wang (Huazhong University Of Science And Technology)
CodeRecognitionComputational EfficiencyTransformerPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
π― What it does: Propose the MultiClip-Bench dataset and design IPFormer (Instance Prompt-guided Transformer) to achieve instance-aware understanding of multimodal multi-camera videos.
π― What it does: For scenes reconstructed by 2D Gaussian Splatting, the IQGS (Instance Query-based Gaussian Segmentation) framework is proposed to achieve panoramic instance segmentation.
IROTE: Human-like Traits Elicitation of Large Language Model via In-Context Self-Reflective Optimization
Yuzhuo Bai (Tsinghua University), Xing Xie (Microsoft Research Asia)
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes the IROTE method, which enables large language models (LLMs) to stably and transferably exhibit specified human traits through self-reflective text without fine-tuning.
π― What it does: Proposed the LaT-IB method, which splits representations into clean label space and noisy label space, and designs a three-stage training framework to learn Minimal-Sufficient-Clean representations under label noise environments.
Is Your (Reasoning) Multimodal Language Model Vulnerable Toward Distractions?
Ming Liu (Iowa State University), Wensheng Zhang (Iowa State University)
CodeAdversarial AttackPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Constructed the I-ScienceQA benchmark to systematically evaluate the robustness of Vision-Language Models under visual and textual perturbation conditions.
CodeSafty and PrivacyRobotic IntelligenceTransformerPrompt EngineeringVision Language ModelTextMultimodalityBenchmarkChain-of-Thought
π― What it does: This paper proposes the IS-Bench benchmark, specifically designed to evaluate the interaction safety of vision-language model-driven embodied agents in daily household tasks, constructing 161 interactive scenarios, 388 safety risks, and introducing a process-oriented evaluation method;
π― What it does: Propose a multi-granularity retrieval-augmented generation framework (MGranRAG) that combines local iterative retrieval with a lightweight context hierarchy graph, leveraging LLM to extract fine-grained evidence and propagate global semantics on the graph.
ITPP: Learning Disentangled Event Dynamics in Marked Temporal Point Processes
Wang-Tao Zhou (University of Electronic Science and Technology of China), Ling Tian (Shenzhen Institute for Advanced Study)
CodeTransformerTime SeriesSequentialElectronic Health RecordsOrdinary Differential Equation
π― What it does: Proposed the ITPP model, which utilizes channel-independent encoding and type-aware inverted self-attention to model marked point processes.
CodeClassificationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Proposed the Judge of Edit-Level Validity (JELV) framework for automated evaluation of the validity of revisions in grammar error correction (GEC), and implemented two judgment implementations based on LLM and DeBERTa;