AAAI Conference on Artificial Intelligence Β· 2140 papers
Doubly Debiased Test-Time Prompt Tuning for Vision-Language Models
Fei Song (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)
CodeClassificationKnowledge DistillationTransformerPrompt EngineeringVision Language ModelImageMultimodalityRetrieval-Augmented Generation
π― What it does: Proposes a dual debiasing test-time prompt tuning method called D2TPT, aiming to enhance the generalization ability of vision-language models on unlabeled test data by mitigating prompt optimization bias through dynamic retrieval and reliability-aware mechanisms.
Dhruv Sarkar (Indian Institute of Technology Kharagpur), Sayak Ray Chowdhury (Indian Institute of Technology Kanpur)
CodeSafty and PrivacyReinforcement Learning
π― What it does: Proposed a multi-armed bandit algorithm DP-NCB that simultaneously satisfies differential privacy (DP) and fairness (measured by Nash regret), providing theoretical guarantees and experimental validation under both global and local DP models.
DPRM: A Dual Implicit Process Reward Model in Multi-Hop Question Answering
Xinyi Wang (National University of Defense Technology), Minlie Huang (National University of Defense Technology)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextGraphRetrieval-Augmented GenerationChain-of-Thought
π― What it does: This paper proposes a Dual Implicit Process Reward Model (DPRM) to simultaneously evaluate chain-of-thought (CoT) and knowledge graph (KG) paths in multi-hop question answering tasks, achieving collaborative reasoning through mutual evaluation between the two.
DR.Experts: Differential Refinement of Distortion-Aware Experts for Blind Image Quality Assessment
Bohan Fu (Beijing Institute of Technology), Runze Hu (Beijing Institute of Technology)
CodeTransformerMixture of ExpertsVision Language ModelImage
π― What it does: Proposes DR.Experts, a blind image quality assessment framework that leverages distortion priors extracted by DA-CLIP, refines distortion features through the Distortion-Saliency Differential Module, and fuses them according to distortion importance in the Dynamic Distortion Weighting Module.
DRAFT-RL: Multi-Agent Chain-of-Draft Reasoning for Reinforcement Learning-Enhanced LLMs
Yuanhao Li, Wei Tan (Beijing University Of Posts And Telecommunications)
CodeReinforcement Learning from Human FeedbackReinforcement LearningAgentic AITextChain-of-Thought
π― What it does: Propose the DRAFT-RL framework, integrating multi-agent reinforcement learning with Chain-of-Draft (CoD) reasoning, enabling each agent to generate multiple concise reasoning drafts and selecting/learning through peer evaluation and reward models.
Yuan Zhou (Nanyang Technological University), Hanwang Zhang (Hefei University of Technology)
CodeDiffusion modelImageBenchmark
π― What it does: Proposed a region-level geometric transformation-based Drag-Based Image Editing (DBIE) framework named DragNeXt, utilizing Latent Space Optimization (LRO) and Progressive Backward Self-Intervention (PBSI) to achieve image drag editing.
Drift Estimation for Diffusion Processes Using Neural Networks Based on Discretely Observed Independent Paths
Yuzhen Zhao (Sorbonne Universite), Marc Hoffmann (Sorbonne Universite)
CodeTime SeriesStochastic Differential Equation
π― What it does: Nonparametric estimation of the drift function in time-homogeneous diffusion processes using deep ReLU neural networks, with non-asymptotic convergence rates provided.
CodeData SynthesisAnomaly DetectionLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
π― What it does: Studied the multi-level drift caused by the diversity of news generated by generative AI, and evaluated the robustness of large vision-language models in detecting multimodal misinformation, constructing a benchmark named DRIFTBENCH.
Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning
Yue Li (University of Science and Technology of China), Xinhai Zhao (Huawei Noah's Ark Lab)
CodeAutonomous DrivingOptimizationTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextSequentialChain-of-Thought
π― What it does: Propose the Drive-R1 model in autonomous driving scenarios, integrating vision-language models with planning, achieving interpretable trajectory prediction from visual inputs through supervised learning and reinforcement learning.
π― What it does: Proposes DriveFlow, a frequency decomposition method based on a pre-trained text-to-image flow model, to enhance the robustness of visual 3D object detection during training.
π― What it does: Propose the DriveSuprim method, which improves the trajectory selection accuracy of end-to-end planning through coarse-to-fine filtering, rotation data augmentation, and self-distillation.
π― What it does: This paper reformulates the Android malware detection problem as a first-order Markov Decision Process (MD-MDP) and trains a DRMD agent using deep reinforcement learning (PPO), integrating classification, active learning, and rejection decisions to achieve adaptive detection under concept drift.
π― What it does: Introducing dual-phase collaborative learning in direct SNN training: forward adaptive threshold (AT) and backward gradient optimization (TGO)
DSAP: Enhancing Generalization in Goal-Conditioned Reinforcement Learning
Yiming Wang (University of Macau), Leong Hou U (University of Macau)
CodeReinforcement LearningGraph
π― What it does: Propose the DSAP (Deconfounded State Abstraction for Policy learning) framework in goal-conditioned reinforcement learning, which learns a causal graph as a proxy for confounding variables, employs backdoor adjustment to achieve state abstraction and policy learning, thereby enhancing generalization capabilities for new goals and environments.
DSCodeBench: A Realistic Benchmark for Data Science Code Generation
Shuyin Ouyang (King's College London), Jie M. Zhang (King's College London)
CodeLarge Language ModelTextBenchmark
π― What it does: This paper proposes DSCodeBenchβa code generation benchmark based on real GitHub data science projects, containing 1,000 tasks involving 10 major Python data science libraries;
π― What it does: Propose DSFedMed, a dual-scale federated learning framework that collaborates to train a server-side large foundational model and client-side lightweight model through mutual knowledge distillation, achieving high accuracy in medical image segmentation with low communication/inference costs.
π― What it does: Propose a dual-stream perception point cloud quality assessment framework, DSP-PCQA, which simulates two human cognitive channels: technical rationality and semantic perception, addressing the issues of distortion and semantic conflicts caused by traditional single-view assessment methods.
π― What it does: This paper proposes the DualG framework, which simultaneously constructs package-level and instance-level dual graphs in multi-instance partial label learning to achieve joint optimization of feature learning and label disambiguation.
π― What it does: Proposes Dual Mamba-enhanced Graph Convolutional Network (DMbaGCN), which jointly models hierarchical evolution of node representations and global dependencies through two modules: Local State-Evolution Mamba (LSEMba) and Global Context-Aware Mamba (GCAMba), mitigating over-smoothing in deep GNNs.
π― What it does: This paper proposes a dual-branch asymmetric difference learning framework (DADL), which improves the detection performance of AI-generated images by amplifying inconsistencies within forged images.
π― What it does: Propose a dual-kernel graph community contrastive learning framework that leverages community structure and multi-kernel learning, simultaneously incorporating node-level and community-level information, while achieving efficient inference through decoupled GNN and knowledge distillation.
Dual-Phase Visual-Language Pretraining and Adaptation for Long-Tailed Multi-Label Recognition
Yongcheng Li (Tongji University), Cairong Zhao (Tongji University)
CodeRecognitionTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodality
π― What it does: Propose a two-stage vision-language pre-training and adaptation framework DP-VLPA, which generates rich descriptions for tail classes using large language models (LLMs) and improves long-tailed multi-label recognition performance in the second stage through dynamic query reweighting and co-occurrence aware loss.
Dual-stream Relation-modeling Disentanglement for Cloth-Changing Person Re-Identification
Shijuan Huang, Zhao Lv (Huazhong University Of Science And Technology)
CodeRecognitionRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Propose the DRDnet framework, utilizing dual-stream decomposition and relationship modeling to achieve person re-identification under clothing changes, overcoming the issues of traditional auxiliary modalities being overly dependent on single modes and label inconsistencies.
DualSpeechLM: Towards Unified Speech Understanding and Generation via Dual Speech Token Modeling with Large Language Models
Yuanyuan Wang (Chinese University of Hong Kong), Xixin Wu (Chinese University of Hong Kong)
CodeRecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderTextAudio
π― What it does: This study proposes the DualSpeechLM model, achieving the unification of speech understanding and generation by bridging the modal gap between text and speech through a Understanding-Driven Speech Tokenizer (USTokenizer) and a dual token modeling framework.
π― What it does: Proposes DuGI-MAE, a self-supervised pre-training model for infrared images, combining entropy-based deterministic masking and dual-domain guidance modules to enhance representation learning for infrared images.
DUP: Detection-guided Unlearning for Backdoor Purification in Language Models
Man Hu (Beijing Electronic Science and Technology Institute), Shuai Zhao (Nanyang Technological University)
CodeSafty and PrivacyKnowledge DistillationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposed the DUP framework, combining feature-level backdoor detection with parameter-efficient unlearning (LoRA + knowledge distillation) for backdoor purification in language models.
π― What it does: Proposed a dynamic weighted dual-graph attention network (DW-DGAT) for fusing multimodal, multi-structure neuroimaging and phenotypic data to enable early diagnosis of Parkinson's disease and Alzheimer's disease.
π― What it does: This paper proposes Dynamic Agent Grouping ECBS (DAG-ECBS), which uses a bounded suboptimal solver within the Windowed Complete MAPF framework to solve windowed multi-agent pathfinding problems while maintaining completeness.
π― What it does: Proposes DGIMVCM, a dynamic deep graph learning framework for incomplete multi-view clustering, which compensates for missing views and performs clustering through global graph fusion, GCN embedding layer, GAT encoder, mask graph reconstruction loss, and pseudo-label self-supervised module.
Dynamic Deep Prompt Optimization for Defending Against Jailbreak Attacks on LLMs
Doniyorkhon Obidov (Michigan Technological University), Kaichen Yang (Michigan Technological University)
CodeSafty and PrivacyAdversarial AttackLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
π― What it does: Proposes a defense method based on dynamic deep hint optimization (DDPO), which generates input-dependent defense embeddings using intermediate features of the LLM itself and injects them into subsequent layers;
Dynamic Semantic Tokenization for Time Series via Elastic Sampling on Physics-aware Perception
Huaizhang Liao (National University of Defense Technology), Yongxiang Liu (National University of Defense Technology)
CodeClassificationRepresentation LearningRecurrent Neural NetworkTransformerContrastive LearningTime SeriesBiomedical DataPhysics Related
π― What it does: Investigated physics-aware semantic segmentation for time series, proposing the PATK framework to achieve elastic adaptive segmentation.
π― What it does: Empirical analysis using the real dynamics and Jacobian from MuJoCo Playground to evaluate global, state-dependent, and temporal sparsity, and to test the performance of simple MLPs in capturing these sparse structures.
π― What it does: Propose a dynamic weight adaptation mechanism (DWAM) based on BCM theory, enabling spiking neural networks to spontaneously achieve neuronal steady-state regulation during the inference phase;
π― What it does: This paper studies the task of unsupervised domain adaptation for video visible-infrared person re-identification (UDA-VVIReID) and proposes the DSC framework based on dynamic-static collaboration.
DynamicEarth: How Far Are We from Open-Vocabulary Change Detection?
Kaiyu Li, Zhi Wang (Xi'an Jiaotong University)
CodeClassificationObject DetectionSegmentationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Proposed the Open-Vocabulary Change Detection (OVCD) task and designed two unsupervised frameworks, M-C-I and I-M-C, based on pre-trained foundation models, for detecting changes of arbitrary categories in remote sensing images.
DynamicRTL: RTL Representation Learning for Dynamic Circuit Behavior
Ruiyang Ma (Peking University), Guojie Luo (Nanjing University of Aeronautics and Astronautics)
CodeRepresentation LearningRecurrent Neural NetworkGraph Neural NetworkGraphTime Series
π― What it does: A dynamic behavior-based RTL-level circuit representation learning model, DR-GNN, was researched and implemented, along with the construction of the first dynamic circuit dataset containing 6,300 Verilog designs and 63,000 simulation trajectories.
DynaQuant: Dynamic Mixed-Precision Quantization for Learned Image Compression
Youneng Bao (Shenzhen University), Yongsheng Liang (Shenzhen University)
CodeCompressionImage
π― What it does: Propose DynaQuant, a dynamic mixed-precision quantization framework for learned image compression, achieving content-aware quantization parameters and dynamic bit-width selection;
DySy-Det: A Synergistic Framework with Dynamic Reconstruction-Path Consistency for AI-Generated Image Detection
Fanli Jin (Zhejiang University), Zhisheng Yan (George Mason University)
CodeAnomaly DetectionTransformerVision Language ModelDiffusion modelImage
π― What it does: Propose a unified framework named DySy-Det for detecting AI-generated images, integrating multi-dimensional features such as semantic consistency, local reconstruction error, and dynamic generation trajectory consistency.
E-Logic Prompt: Unified Energy-Logic Framework for Continual Visual Question Answering
Jiayao Tan (Tianjin University), Liang Wan (Tianjin University)
CodePrompt EngineeringVision Language ModelScore-based ModelMultimodality
π― What it does: This paper proposes a unified energy logic prompt framework (E-Logic Prompt) based on energy models to address the knowledge forgetting problem in continuous visual question answering.
π― What it does: Propose a lightweight E3SAM2 framework for cardiac ultrasound video segmentation, combining entropy-guided attention, entropy regularization, and edge supervision;
π― What it does: Proposed an edge-aware reconstruction-guided network (EARG-Net), which achieves image tampering detection and localization by masking suspected tampered regions, reconstructing using a pre-trained image inpainting model, and then extracting fine-grained forgery traces through reconstruction residuals.
Easy for Children, Hard for AI: The Limits of Multimodal LLMs in Early Childhood Learning
Jingping Liu (Sun Yat-sen University), Huacan Wang (University of Chinese Academy of Sciences)
CodeSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Proposed and constructed ChildBench β a multi-modal benchmark for evaluating models' cognitive abilities in early childhood learning, including spatial reasoning, visual reasoning, visual discrimination, counting, and visual tracking;
EasyText: Controllable Diffusion Transformer for Multilingual Text Rendering
Runnan Lu (National University of Singapore), Yiren Song (National University of Singapore)
CodeGenerationTransformerDiffusion modelImageText
π― What it does: Propose the EasyText framework, which achieves controllable multilingual text rendering based on Diffusion Transformer; combines character positioning encoding with position interpolation to support text generation at arbitrary positions and shapes.
π― What it does: This paper proposes EC-MVSNet, an enhanced cascaded multi-view stereo framework, which improves depth estimation accuracy by leveraging cross-scale feature joint cost volume construction, probabilistic guidance enhancement, and monocular feature refinement.
π― What it does: Proposes a medical image segmentation model named EccoMamba based on a U-shaped encoder-decoder architecture, addressing the sequential and single-scale limitations of the Mamba structure when processing 2D/3D medical images.
EchoEdit: Consistent Multi-Hop Question Answering via Ripple Control in Knowledge Editing
Jinwei Shi, Tieke He (Nanjing University)
CodeGraph Neural NetworkLarge Language ModelTextGraphChain-of-Thought
π― What it does: This paper proposes EchoEdit, an impact-control-based knowledge editing framework that maintains coherence and consistency in multi-hop reasoning without retraining the model.
π― What it does: Propose a precomputation framework named Echoless-LP to address the echo leakage problem caused by label precomputation in heterogeneous graph learning;
π― What it does: Propose EchoMimicV3, a unified multi-task, multi-modal human animation framework with 1.3B parameters, capable of performing multiple animation tasks such as text-to-video, image-to-video, and audio-driven lip-sync within the same model.
EcoAgent: An Efficient Device-Cloud Collaborative Multi-Agent Framework for Mobile Automation
Biao Yi (Zhejiang University), Fan Wu (Hong Kong Polytechnic University)
CodeSafty and PrivacyComputational EfficiencyAI Code AssistantTransformerLarge Language ModelVision-Language-Action ModelMultimodalityBenchmarkChain-of-Thought
π― What it does: Propose EcoAgent, a closed-loop device-cloud collaborative multi-agent framework for achieving privacy-friendly, efficient, and responsive mobile automation;
π― What it does: Proposed the ECPv2 algorithm for global optimization under unknown Lipschitz constants, improving the efficiency and scalability of ECP.
π― What it does: Propose an edge self-adversarial enhanced graph contrastive learning framework named EDA-GCL, which can generate two adversarial views with neighborhood inconsistency under unsupervised settings and perform contrastive learning using them.
CodeClassificationConvolutional Neural NetworkTime Series
π― What it does: Propose EdgeMTSC, a lightweight ConvNet with large convolutional kernels for multivariate time series classification (MTSC) on edge devices, achieving inter-channel information passing and temporal feature learning through the IMP-KB module.
EEG Agent: A Unified Framework for Automated EEG Analysis Using Large Language Models
Sha Zhao (Zhejiang University), Shijian Li (Zhejiang University)
CodeClassificationAnomaly DetectionTransformerLarge Language ModelAgentic AIBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Built EEGAgent, a unified framework based on large language models, for sensing, exploring, detecting EEG, and enabling user interaction with automatic report generation.
CodeComputational EfficiencyKnowledge DistillationTransformerAuto EncoderContrastive LearningBiomedical Data
π― What it does: To address pretraining of large-scale EEG foundation models, this paper proposes a data distillation framework, EEG-DLite, which constructs a concise subset occupying only 5% of the original 2,500-hour EEG dataset through three steps: self-supervised autoencoder compression, anomaly sample removal, and diversity sampling, while maintaining or even exceeding the performance of the full dataset on multiple downstream tasks.
Efficient and Adaptive Simultaneous Speech Translation with Fully Unidirectional Architecture
Biao Fu (Xiamen University), Xiaodong Shi (Xiamen University)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextAudio
π― What it does: Proposed an efficient adaptive simultaneous interpretation system EASiST with a fully unidirectional architecture, achieving end-to-end streaming translation through multi-delay data mining, interleaved read/write tagging, lightweight decision heads, and three-stage training.
Efficient and Reliable Hitting-Set Computations for the Implicit Hitting Set Approach
Hannes Ihalainen (University of Helsinki), Matti JΓ€rvisalo
CodeOptimizationBenchmark
π― What it does: Under the Implicit Hitting Set (IHS) framework, a more efficient, reliable, and provable hitting set solving method is provided for Pseudo-Boolean Optimization (PBO) problems.
π― What it does: Propose an efficient solving method called EFLOCO based on discrete flow matching, which is used to quickly generate high-quality solutions for combinatorial optimization problems (such as TSP, ATSP).
Efficient LLM-Jailbreaking via Multimodal-LLM Jailbreak
Haoxuan Ji (Xi'an Jiaotong University), Gang Hua (Xidian University)
CodeAdversarial AttackPrompt EngineeringVision Language ModelMultimodality
π― What it does: Propose a method that constructs a multimodal large language model (MLLM) and generates text suffixes through image attacks on the MLLM, thereby efficiently achieving jailbreaking of large language models (LLMs).
π― What it does: This paper proposes a multimodal generative model called MAC-WAE based on arbitrary conditions and Wasserstein regularization, which uses a single shared encoder to handle arbitrary missing modalities, addressing the insufficient expressiveness caused by posterior collapse in traditional multimodal VAEs.
Efficient Plug-and-Play Weight Refinement for Sparse Large Models
Jingcheng Xie (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Propose EIWR, a training-free, plug-and-play weight refinement method aimed at efficiently recovering sparse weights generated after one-time pruning.
Efficient Post-Training Refinement of Latent Reasoning in Large Language Models
Xinyuan Wang (Arizona State University), Yanjie Fu (Arizona State University)
CodeComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelContrastive LearningText
π― What it does: Proposed a lightweight post-training framework that dynamically improves LLM's reasoning trajectories in an implicit reasoning space by leveraging contrastive reasoning feedback and residual embedding refinement.
CodeOptimizationDrug DiscoveryProtein Structure PredictionTransformerBiomedical Data
π― What it does: Propose a structure-aware Bayesian optimization framework called HADES, which samples in continuous sequence space using Hamiltonian dynamics and generates discrete protein sequences through position discretization, thereby efficiently searching for the optimal combination of protein structure and function.
Efficient Rule Induction by Ignoring Pointless Rules
Andrew Cropper (ELLIS Institute Finland), David M. Cerna (Dynatrace Research)
CodeComputational EfficiencyImageTabular
π― What it does: This paper proposes an ILP system called REDUCER, which can identify and ignore 'useless rules' (including reducible rules and indifferent rules), thereby performing effective pruning of the hypothesis space during the search process.
CodeSegmentationComputational EfficiencyTransformerLarge Language ModelImageTextMultimodality
π― What it does: Propose an efficient segmentation framework based on a multi-modal large language model (MLLM), which significantly reduces computational load while maintaining segmentation performance through dynamic scheduling of image and mask token participation.
Efficient Tensorized Multi-View Anchor Graph Clustering with Affinity Propagation for Remote Sensing Data
Yongshan Zhang (China University of Geosciences), Zhihua Cai (China University of Geosciences)
CodeOptimizationRepresentation LearningImageMultimodalityAgriculture Related
π― What it does: Propose an efficient tensor quantized multi-view anchor graph clustering method ETAP, which utilizes superpixel preprocessing, joint learning of anchor graph and compressed anchor graph, tensor Schatten p-norm regularization, and connectivity constraints, and infers pixel clustering from anchors directly via affinity propagation.
Efficient Transcoder Adaptation for Fine-Tuned Models: Revealing Medical Reasoning Mechanisms in Large Language Models
Zhouxing Tan (Peking University), Junfei Liu (Peking University)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderBiomedical Data
π― What it does: Fine-tune large language models for medical tasks by freezing attention layers and only updating feed-forward network parameters to construct interpretable encoders.
Efficient Verification and Falsification of ReLU Neural Barrier Certificates
Dejin Ren (Chinese Academy of Sciences), Bai Xue (Chinese Academy of Sciences)
CodeOptimizationSafty and PrivacyComputational Efficiency
π― What it does: This paper proposes a necessary and sufficient condition for verifying and refuting the forward invariance of ReLU neural network barrier certificates under continuous-time systems.
π― What it does: Proposed a multi-scale adaptive observation window framework (ALW) based on wavelet transform, which can dynamically learn the optimal historical observation length for each instance and each frequency band in time series prediction, and reconstruct new input features by weighting and restoring the selected features.
Ego-PMOVE: Prompt-aware Mixture of View Experts Network for Egocentric Gaze Prediction
Heqian Qiu (University of Electronic Science and Technology of China), Hongliang Li (University of Electronic Science and Technology of China)
CodeRecognitionTransformerPrompt EngineeringMixture of ExpertsVision Language ModelContrastive LearningImageVideo
π― What it does: This paper proposes a prompt-aware Mixture of Experts (Ego-PMOVE) network to predict gaze points in first-person perspective, fully leveraging cross-perspective (internal and external) information.
EgoCross: Benchmarking Multimodal Large Language Models for Cross-Domain Egocentric Video Question Answering
Yanjun Li (East China Normal University), Xiaoling Wang (Sofia University St Kliment Ohridski)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
π― What it does: This paper proposes the EgoCross benchmark, specifically designed to evaluate the generalization ability of multi-modal large language models in cross-domain first-person video question answering (EgocentricQA), covering four domains: surgery, industry, extreme sports, and animal perspectives, with a total of 798 videos and 957 QA pairs.
π― What it does: A novel EHL* algorithm is proposed for the Euclidean shortest path problem, capable of building an index under a given memory budget and supporting workload-aware compression using known query distributions.
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTabularBiomedical DataElectronic Health RecordsBenchmark
π― What it does: Proposed the EHRStruct framework, constructed a standardized structured EHR benchmark containing 11 types of clinical tasks and 2,200 samples, systematically evaluated 20 large language models and 11 enhancement methods, and finally proposed the EHRMaster code enhancement approach to achieve optimal performance.
EigenShield: Inference-Time, Model-Agnostic Jailbreaking Defense via Causal Subspace Filtering
Nastaran Darabi (University of Illinois Chicago), Amit Ranjan Trivedi (University of Illinois Chicago)
CodeSafty and PrivacyRepresentation LearningTransformerVision Language ModelTextMultimodality
π― What it does: Propose a model-free defense method called EigenShield during inference, which uses random matrix theory (RMT) to filter out causal subspaces in the high-dimensional embeddings of LLMs and VLMs, thereby suppressing adversarial perturbations while preserving semantic structure;
ElementaryNet: A Non-Strategic Neural Network for Predicting Human Behavior in Normal-Form Games
Greg d'Eon (University of British Columbia), James R. Wright (University of Alberta)
CodeExplainability and InterpretabilityConvolutional Neural NetworkTabular
π― What it does: Proposed a novel neural network structure called ElementaryNet to replace the level-0 model in GameNet that simulates strategic reasoning, and proved that it cannot express any strategic behavior;
π― What it does: This paper proposes the Elite-Pattern Reinforcement (EPR) strategy to improve neural network-based construction methods for solving the vehicle routing problem (VRP).
Ellipsoid-Based Decision Boundaries for Open Intent Classification
Yuetian Zou (Tsinghua University), Long Xiao (Hebei University of Science and Technology)
CodeClassificationTransformerContrastive LearningTextFinance Related
π― What it does: To address open-intent classification in dialogue systems, this paper proposes EliDecide, a model based on an elliptical decision boundary, which simultaneously accurately identifies known intents and reliably rejects unknown intents.
ELSPR: Evaluator LLM Training Data Self-Purification on Non-Transitive Preferences via Tournament Graph Reconstruction
Yan Yu, Fuliang Li (Huawei)
CodeData-Centric LearningLarge Language ModelTextBenchmark
π― What it does: Propose the ELSPR framework, which uses a tournament graph to self-purify bidirectional comparison data from LLM evaluations, eliminating non-transitive preferences;
EMAformer: Enhancing Transformer Through Embedding Armor for Time Series Forecasting
Zhiwei Zhang (Beijing Jiaotong University), Wenjuan Han (Beijing Jiaotong University)
CodeTransformerTime Series
π― What it does: Propose EMAformer, which enhances the performance of multivariate time series forecasting by introducing channel, phase, and joint channel-phase embeddings into a Transformer with variant tokenization.
EmbryoDiff: A Conditional Diffusion Framework with Multi-Focal Feature Fusion for Fine-Grained Embryo Developmental Stage Recognition
Yong Sun, Qiang Nie (Hong Kong University of Science and Technology (Guangzhou))
CodeRecognitionConvolutional Neural NetworkTransformerDiffusion modelVideoBiomedical Data
π― What it does: Proposed a two-stage conditional diffusion framework called EmbryoDiff, which utilizes multi-focal plane time-stretched videos to achieve fine-grained identification of embryonic development stages.
Yuning Chen (Zhejiang University), Gang Pan (Zhejiang University)
CodeRecognitionComputational EfficiencyRepresentation LearningTransformerContrastive LearningBiomedical Data
π― What it does: Propose the EMOD framework, which utilizes the emotional dimensions (Valence-Arousal) to unify labels and structures across different EEG emotion datasets, constructing a cross-dataset emotion representation model;
EMODIS: A Benchmark for Context-Dependent Emoji Disambiguation in Large Language Models
Jiacheng Huang (Hubei Normal University), Xiaoyin Yi (Chongqing University of Posts and Telecommunications)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: Propose the EMODIS benchmark to evaluate the ambiguity resolution capability of large language models in contexts containing emojis.
Emotion-Coherent Reasoning for Multimodal LLMs via Emotional Rationale Verifier
Hyeongseop Rha (KAIST), Yong Man Ro (KAIST)
CodeExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodality
π― What it does: Guide multimodal large language models to generate natural language explanations consistent with target emotions through an Emotion Reasonableness Validator (ERV) and reward mechanism.
π― What it does: Proposed the DiveSeg framework based on DINOv2, achieving high-precision prediction for underwater instance segmentation through AquaStyle Aligner and ObjectPrior Prompter.
CodeGraph Neural NetworkTransformerLarge Language ModelTextGraph
π― What it does: Proposed the Geo-Graph Geometry Problem Solving (GΒ³PS) model, which converts geometric diagrams into geo-graphs and encodes them using a multi-attention graph Transformer. It then fuses the graph features with text features to generate problem-solving steps and answers.
π― What it does: Propose an end-to-end contrastive language-voice retrieval model, CLSR, for retrieving the most relevant segments in long audio to support long-text question answering.
End-to-End Multi-Person Pose Estimation with Pose-Aware Video Transformer
Yonghui Yu (Zhejiang Gongshang University), Wenwu Yang (Zhejiang Gongshang University)
CodePose EstimationTransformerVideo
π― What it does: Propose a full end-to-end multi-person human pose estimation framework called PAVE-Net, eliminating heuristic steps such as detection, RoI cropping, and NMS; achieve precise estimation of multi-person poses in videos through a spatial encoder, pose-aware spatiotemporal decoder, and joint decoder.
Energy-based Autoregressive Generation for Neural Population Dynamics
Ningling Ge (Institute of Automation, Chinese Academy of Sciences), Shan Yu (Institute of Automation, Chinese Academy of Sciences)
CodeGenerationTransformerDiffusion modelScore-based ModelAuto EncoderTime SeriesBiomedical Data
π― What it does: Proposed an energy-based autoregressive generation framework (EAG) for efficiently generating temporal samples of neural population dynamics in latent space, and achieving conditional generation to generalize to unseen behavioral contexts.
Enhanced Privacy Leakage from Noise-Perturbed Gradients via Gradient-Guided Conditional Diffusion Models
Jiayang Meng (Renmin University of China), Guolong Zheng (Minjiang University)
CodeSafty and PrivacyAdversarial AttackDiffusion modelImage
π― What it does: Proposes a gradient-guided conditional diffusion model (GG-CDM) to achieve gradient inversion attacks, capable of recovering high-resolution images from noisy perturbed gradients.
π― What it does: Propose a new attack strategy for All-to-X backdoor attacks, significantly improving attack success rates by optimizing source class grouping and target class mapping, while maintaining robustness against existing defenses.
Enhancing Chemical Explainability Through Counterfactual Masking
Εukasz JanisiΓ³w (Jagiellonian University), Tomasz Danel (Jagiellonian University)
CodeExplainability and InterpretabilityDrug DiscoveryGraph Neural NetworkDiffusion modelGraph
π― What it does: Propose a counterfactual masking method based on generative models, which enhances the interpretability and evaluation metrics of molecular property prediction models by replacing masked substructures with chemically reasonable fragments;
π― What it does: Proposed the ASAP method, achieving action smoothing without structural loss by predicting actions using the transition distribution of the previous state and incorporating a second-order difference penalty.
π― What it does: Propose a dialogue recommendation framework PCRS-TKA that integrates pre-trained language models (PLMs) with knowledge graphs, utilizing retrieval-augmented generation to construct dialogue-specific tree-structured knowledge and serialize it, while combining user multi-round collaborative preferences and semantic alignment to achieve fine-grained reasoning and response generation.
π― What it does: Propose a new variant of differential privacy stochastic gradient descent (DP-SGD) called DP-PMLF, which combines per-sample momentum smoothing and low-pass filtering to simultaneously mitigate DP noise and clipping bias.