These 343 IJCAI 2025 papers come with a code repository. Each shows an AI one-line summary below β get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every IJCAI 2025 paper, free trial on arXivSub.
A Correlation Manifold Self-Attention Network for EEG Decoding
Chen Hu (Jiangnan University), Ziheng Chen (University of Trento)
CodeClassificationTransformerBiomedical Data
π― What it does: Propose Correlation Attention Network (CorAtt), implementing self-attention on full-rank correlation matrices, and design corresponding transformation layers, aggregation layers, and tangent mapping layers.
A Finite-State Controller Based Offline Solver for Deterministic POMDPs
Alex Schutz (University of Oxford), Nick Hawes (University of Oxford)
CodeOptimizationReinforcement Learning
π― What it does: Designed an offline solver called DetMCVI based on a finite state controller for solving deterministic partially observable Markov decision processes (DetPOMDP).
A Hybrid Multi-Factor Network with Dynamic Sequence Modeling for Early Warning of Intraoperative Hypotension
Mingyue Cheng (University of Science and Technology of China), Chunli Liu (Hefei University of Technology)
CodeAnomaly DetectionTransformerTime SeriesBiomedical Data
π― What it does: Redefine the early warning of hypotension during surgery as a multivariate time series prediction problem and propose the Hybrid Multi-Factor (HMF) network.
π― What it does: Propose a multi-stage neuro-symbolic framework for joint utilization of visual perception, constraint reasoning, and temporal reasoning in sequence classification tasks involving relational and temporal knowledge.
π― What it does: Proposed a new local search algorithm called CELS based on (k,l,S)-cluster for solving the Vertex Bipartition Minimization Problem (VBMP).
π― What it does: Propose a prior-based discrete diffusion model (PDDM) that generates more realistic social propagation graphs through a discrete forward process and a reverse starting point based on user similarity.
A Reduction-Based Algorithm for the Clique Interdiction Problem
Chenghao Zhu (University of Electronic Science and Technology of China), Haoyu Jiang (University of Electronic Science and Technology of China)
CodeOptimizationComputational EfficiencyGraph
π― What it does: For the Clique Interdiction Problem, we propose an algorithm called RECIP based on data reduction, and design various new reduction rules in the preprocessing phase.
π― What it does: This paper proposes a counting method based on SAT that can directly count all single-point attractors in a given Boolean network without enumerating individual single-point attractors.
π― What it does: This paper proposes a Temporal Adaptive Frequency Enhancement Framework (TFDSR) to improve image super-resolution based on diffusion models.
A Weighted-Based Fast Local Search for Ξ±-Neighbor p-Center Problem
Qingyun Zhang (Huazhong University of Science and Technology), Zhouxing Su (Huazhong University of Science and Technology)
CodeOptimizationGraphBenchmark
π― What it does: Propose an algorithm based on Weighted Fast Local Search (WFLS) to solve the Ξ±-neighbor p-center problem by transforming the original problem into a series of decision subproblems with given radii and solving them step by step.
π― What it does: This paper constructs two new Raven Advanced Matrix (RPM) benchmarksβA-I-RAVEN and I-RAVEN-Meshβto systematically evaluate the generalization and knowledge transfer capabilities of deep learning models in abstract visual reasoning.
Accurate Sublayer Pruning for Large Language Models by Exploiting Latency and Tunability Information
Seungcheol Park (Seoul National University), U Kang (Seoul National University)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper proposes a novel sublayer pruning method called SPRINT, specifically designed to enhance inference speed for large language models (LLMs) while maintaining accuracy.
ActiveHAI: Active Collection Based Human-AI Diagnosis with Limited Expert Predictions
Xuehan Zhao (Northwestern Polytechnical University), Bin Guo (Northwestern Polytechnical University)
CodeClassificationConvolutional Neural NetworkTransformerBiomedical DataElectronic Health Records
π― What it does: Proposed ActiveHAI, combining median window active sampling with an evaluation module to achieve efficient human-machine collaborative diagnosis when expert predictions are limited.
Hang Yang (Macau University of Science and Technology), Witold Pedrycz (University of Alberta)
CodeClassificationImage
π― What it does: Propose AdaCrowd, a probability model-based adaptive learning framework that can dynamically select the most informative instances during the label collection process, effectively estimating worker parameters and training the target model with only a small amount of annotations.
π― What it does: This paper proposes an adaptive gradient learning framework (MPD-AGL), which dynamically adjusts the width of synaptic gradients to align with the distribution of membrane potential dynamics (MPD) across different time steps, thereby alleviating the problem of gradient vanishing or mismatch in SNN training.
Siyan Fang (Huazhong University of Science and Technology), Yuehuan Wang (Huazhong University of Science and Technology)
CodeRestorationVision Language ModelMultimodalityBenchmark
π― What it does: Propose Adaptive Language-Aware Network (ALANet) for removing complex reflections in single images, even when the provided language descriptions contain errors;
Adaptive Wizard for Removing Cross-Tier Misconfigurations in Active Directory
Huy Q. Ngo (University of Adelaide), Hung X. Nguyen (University of Adelaide)
CodeOptimizationReinforcement LearningGraph
π― What it does: Propose the Adaptive Path Removal (APR) problem and design precise, approximate, and scalable heuristic algorithms to guide IT administrators in progressively removing attack paths in Active Directory graphs using a wizard-based interface;
AdaR: An Adaptive Gradient Method with Cyclical Restarting of Moment Estimations
Yangchuan Wang (Beijing Wuzi University), Peng Shi (University of Science and Technology Beijing)
CodeOptimizationImageText
π― What it does: Propose AdaR, which suppresses long-tail gradients by periodically restarting Adam's momentum estimation, thereby enhancing the optimizer's generalization and convergence speed.
π― What it does: Proposes an Anchor-Driven Deformable and Compressed Gaussian Splatting (ADC-GS) framework, achieving efficient compression and rendering of dynamic scenes through anchor-driven hierarchical deformation and multi-dimensional entropy models.
Advancing Community Detection with Graph Convolutional Neural Networks: Bridging Topological and Attributive Cohesion
Anjali de Silva (Victoria University of Wellington), Xingquan Zuo (Beijing University of Posts and Telecommunications)
CodeGraph Neural NetworkGraph
π― What it does: Propose a new graph convolutional network community detection framework called TAS-Com, which trains GCN by combining high-quality community structures generated by the Leiden algorithm with the topological connectivity of human-annotated communities, ultimately achieving better community partitioning in attributed graphs.
Advancing Embodied Agent Security: From Safety Benchmarks to Input Moderation
Ning Wang (Chongqing University), Tao Xiang (Chongqing University)
CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Proposes a complete input review framework for embodied agents, including the safety benchmark EAsafetyBench and a lightweight Prompt-decoupled review method called Pinpoint.
π― What it does: Proposed and evaluated the Pathways of Normalized Group Convolution (PoNG) model, conducting experiments on the generalization performance of abstract visual reasoning (AVR) tasks (e.g., Raven Progressive Matrices, Visual Analogies) under both i.i.d. and out-of-distribution (o.o.d.) settings.
Advancing Stain Transfer for Multi-Biomarkers: A Human Annotation-Free Method Based on Auxiliary Task Supervision
Siyuan Xu (East China Normal University), Qingli Li (East China Normal University)
CodeImage TranslationConvolutional Neural NetworkTransformerGenerative Adversarial NetworkBiomedical Data
π― What it does: Investigated an H&E to IHC staining transfer method that does not require manual annotation and is applicable to multiple biomarkers, capable of generating virtual IHC images while maintaining pathological and structural consistency.
π― What it does: Propose an end-to-end adaptive kernel representation (AKBR) framework that uses an attention mechanism to weight substructures in traditional R-convolution kernels (e.g., WLSK, SPGK), automatically learning the importance of substructures and constructing a learnable kernel matrix for graph classification tasks;
An Approach to Quantify Plans Robustness in Real-world Applications
Francesco Percassi (University of Huddersfield), Mauro Vallati (University of Huddersfield)
CodeOptimizationRobotic IntelligenceTabular
π― What it does: Proposes a statistical framework for evaluating plan robustness in uncertain environments, and introduces the concepts of execution-invariant tasks and B-robustness.
π― What it does: Designed an end-to-end simple clustering hierarchical pooling method called SCHPool, which progressively compresses graphs through layer-by-layer compression using Top-K node selection combined with multi-perspective importance scoring and position attention, achieving O(E+N) computational complexity via sparse matrices.
Approximate Verification of Strategic Abilities under Imperfect Information Using Local Models
Damian Kurpiewski (Polish Academy of Sciences), Yan Kim (University of Luxembourg)
Code
π― What it does: This paper proposes a scheme that utilizes proxy local models for approximation verification to check strategy capabilities under incomplete information conditions.
ARMR: Adaptively Responsive Network for Medication Recommendation
Feiyue Wu (Southeast University), Shenqi Jing (Nanjing Medical University)
CodeRecommendation SystemTransformerTabularTime SeriesBiomedical DataElectronic Health Records
π― What it does: Propose the ARMR adaptively responsive network for dynamic modeling of historical and new drug recommendations, enhancing personalized medication suggestions.
ARPDL: Adaptive Relational Prior Distribution Loss as an Adapter for Document-Level Relation Extraction
Huangming Xu (Northeastern University), Xin Li (Northeastern University)
CodeRepresentation LearningTransformerLarge Language ModelText
π― What it does: Propose a new multi-label loss function called ARPDL for document-level relation extraction tasks, and apply it as a generic adapter to existing threshold loss functions;
Attention-based Conditional Random Field for Financial Fraud Detection
Xiaoguang Wang (Xi'an Jiaotong University), Tao Qin (Xi'an Jiaotong University)
CodeAnomaly DetectionRecurrent Neural NetworkTabularTime SeriesFinance Related
π― What it does: The paper proposes an attention-based conditional random fields recurrent neural network (ACRF-RNN) for detecting financial statement fraud.
AttentionDrag: Exploiting Latent Correlation Knowledge in Pre-trained Diffusion Models for Image Editing
Biao Yang (Fudan University), Hualei Liu (Alibaba Group)
CodeGenerationDiffusion modelImageBenchmark
π― What it does: Developed a single-step point-based image editing method called AttentionDrag based on the self-attention latent correlations of pre-trained diffusion models, which can automatically generate masks and maintain image semantic consistency through semantic interpolation;
π― What it does: This paper proposes an improved multi-modal learning method based on training sequencesβBalance-Aware Sequence Sampling (BSS)βwhich mitigates performance degradation caused by modal imbalance by evaluating sample balance and iteratively training from balanced to unbalanced orders.
Balancing Imbalance: Data-Scarce Urban Flow Prediction via Spatio-Temporal Balanced Transfer Learning
Xinyan Hao (Beijing Jiaotong University), Youfang Lin (Beijing Jiaotong University)
CodeDomain AdaptationAuto EncoderContrastive LearningTime Series
π― What it does: Proposes the STBaT framework for cross-city traffic flow prediction based on regional imbalance perception, correcting source city distribution bias and enabling knowledge transfer through the Region Imbalance Acquisition Module and Spatio-Temporal Balanced Learning Module, achieving precise prediction in data-scarce cities.
Balancing User-Item Structure and Interaction with Large Language Models and Optimal Transport for Multimedia Recommendation
Haodong Li (China University of Petroleum (East China)), Xiaokang Zhou (Kansai University)
CodeRecommendation SystemGraph Neural NetworkTransformerLarge Language ModelMultimodality
π― What it does: Propose the BLAST framework, which leverages large language models to generate user/item-side information, constructs user-user and item-item graphs, and combines Optimal Transport data augmentation with negative sample constraints to achieve structural and interactive balance in multi-modal recommendation.
Balans: Multi-Armed Bandits-based Adaptive Large Neighborhood Search for Mixed-Integer Programming Problems
Junyang Cai (University of Southern California), Bistra Dilkina (University of Southern California)
CodeOptimizationReinforcement LearningTabular
π― What it does: Proposed an online meta-solver called BALANS, integrating Mixed Integer Programming (MIP), Adaptive Large Neighborhood Search (ALNS), and Multi-Armed Bandit (MAB), to dynamically select destroy-repair operators for solving MIP problems without relying on offline training.
π― What it does: This paper proposes BEVTrack, a 3D single-target tracking method based on bird's-eye view (BEV) that uses only a single regression loss, significantly simplifying the complex frameworks of traditional multi-task, multi-loss approaches;
π― What it does: Propose Counterfactual Reasoning Decision Transformer (CRDT), which enhances Decision Transformer with counterfactual reasoning capabilities by generating and leveraging counterfactual experiences to improve offline RL decision-making;
π― What it does: Propose a Bi-Directional Diffusion-Guided Collaborative Change Detection Model (Bi-DiffCD), which achieves modal alignment for remote sensing images of arbitrary modalities through conditional diffusion, and realizes more accurate change detection by leveraging multi-level difference features.
Bimodal Depth-First Search for Scalable GAC for AllDifferent
Sulian Le Bozec-Chiffoleau (IMT Atlantique), Xavier Lorca (IMT Mines Albi)
CodeOptimizationComputational EfficiencyGraph
π― What it does: Proposed a Bimodal DFS/BFS that can efficiently execute on both sparse and dense graphs, embedded into Regin's GAC algorithm to achieve efficient strong arc consistency inference for ALLDIFFERENT constraints.
π― What it does: Proposed the Binary Event-Driven Spiking Transformer (BESTformer), achieving efficient integration of Transformer and Spiking Neural Network;
π― What it does: Propose the BRCB algorithm to achieve robust compression of embodied AI models, fine-grained splitting model weights into anti-disturbance stable and sensitive branches, and adding a gating layer during compression training to enhance the robustness and efficiency of the compressed model compared to the original model.
BTPG: A Platform and Benchmark for Behavior Tree Planning in Everyday Service Robots
Xinglin Chen (National University of Defense Technology), Ji Wang (National University of Defense Technology)
CodeRobotic IntelligenceLarge Language ModelBenchmark
π― What it does: Designed and implemented Behavior Tree Planning Gym (BTPG), providing a unified predicate logic + STRIPS behavior tree planning platform, four real/simulated environments (RoboWaiter, VirtualHome, OmniGibson, RobotHow), as well as automatically generated task datasets and benchmark evaluation frameworks.
CABIN: Debiasing Vision-Language Models Using Backdoor Adjustments
Bo Pang (University of Auckland), Yun Sing Koh (University of Auckland)
CodeExplainability and InterpretabilityData-Centric LearningVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Utilize the backdoor adjustment method from causal inference to neutralize the image embeddings in the sensitive attribute subspace of vision-language models (VLM), achieving unbiased prediction.
CADP: Towards Better Centralized Learning for Decentralized Execution in MARL
Yihe Zhou (Zhejiang University), Mingli Song (Zhejiang University)
CodeTransformerReinforcement LearningBenchmark
π― What it does: Proposes the Centralized Advising and Decentralized Pruning (CADP) framework, which enhances training efficiency in CTDE through centralized advice exchange and achieves communication-free independent policies during execution via self-pruning.
Can Large Models Teach Student Models to Solve Mathematical Problems Like Human Beings? A Reasoning Distillation Method via Multi-LoRA Interaction
Xinhe Li (Southeast University), Peng Wang (Southeast University)
CodeKnowledge DistillationTransformerLarge Language ModelTextChain-of-Thought
π― What it does: Propose LoRID, a math reasoning distillation framework based on multi-LoRA interaction, which enhances the reasoning capabilities of small language models by explicitly augmenting knowledge and improving the interaction between System 1/System 2.
π― What it does: This paper constructs the R2PE benchmark to systematically assess the relationship between the reasoning chain and the correctness of the final answer in chain-of-thought (CoT) generated by large language models, and proposes the Process Discernibility Score (PDS) framework to determine the authenticity of answers.
Causal View of Time Series Imputation: Some Identification Results on Missing Mechanism
Ruichu Cai (Guangdong University of Technology), Zhifeng Hao (Shantou University)
CodeRestorationFlow-based ModelTime Series
π― What it does: Studied the problem of missing value imputation in time series, proposing a DMM framework based on causal mechanisms that can model both MAR and MNAR missing mechanisms.
Causality-Inspired Disentanglement for Fair Graph Neural Networks
Guixian Zhang (China University of Mining and Technology), Yanmei Zhang (China University of Mining and Technology)
CodeClassificationExplainability and InterpretabilityGraph Neural NetworkGenerative Adversarial NetworkGraphTabularBenchmark
π― What it does: Proposes a causality-inspired separable framework CDFG for disentangling causal factors and sensitive factors from graph neural networks to achieve fair node classification.
CFDONEval: A Comprehensive Evaluation of Operator-Learning Neural Network Models for Computational Fluid Dynamics
Menghan Liu (Sun Yat-sen University), Qingsong Zou (Sun Yat-sen University)
CodeGraph Neural NetworkTransformerAuto EncoderMeshBenchmarkPhysics Related
π― What it does: Proposed the CFDONEval evaluation framework, systematically assessing 12 operator learning neural network models across 22 datasets, covering 7 classical CFD benchmark problems, and exploring five major challenges: multi-scale, convection-dominated, long-term prediction, multiphase flow, and complex geometry with non-uniform grids.
Vincent Derkinderen (KU Leuven), Jean-Marie Lagniez (CRIL)
CodeOptimizationComputational Efficiency
π― What it does: They improved the d4 compiler, preserving the original Boolean circuit and implementing dynamic elimination of don't-care variables (irrelevant gates).
Co-Learning of Strategy and Structure Achieves Full Cooperation in Complex Networks with Dynamical Linking
Xiaoqing Fan (University of Warwick), Paolo Turrini (University of Warwick)
CodeReinforcement LearningGraph
π― What it does: Enabling reinforcement learning agents to achieve full cooperation through co-learning strategies and dynamic links on complex networks.
CoderAgent: Simulating Student Behavior for Personalized Programming Learning with Large Language Models
Yi Zhan (University of Science and Technology of China), Zhenya Huang (University of Science and Technology of China)
CodeAI Code AssistantLarge Language ModelAgentic AITextSequentialChain-of-Thought
π― What it does: This paper proposes a CoderAgent based on a large language model to fine-grainedly simulate the student programming process and generate interpretable learning trajectories.
COGRASP: Co-Occurrence Graph Based Stock Price Forecasting
Zhengze Li (University of Gttingen), Xiaoming Fu (University of Gttingen)
CodeRecurrent Neural NetworkGraph Neural NetworkTextGraphTime SeriesFinance Related
π― What it does: Construct a social media co-occurrence relationship graph and combine it with a multi-scale attention LSTM model to predict stock prices.
Collaborative Multi-LoRA Experts with Achievement-based Multi-Tasks Loss for Unified Multimodal Information Extraction
Li Yuan (South China University of Technology), Tao Wang (King's College London)
CodeTransformerMixture of ExpertsVision Language ModelMultimodalityBenchmark
π― What it does: Propose a collaborative multi-LoRA expert model (C-LoRAE), utilizing a two-layer structure combining general experts and task-specific experts, along with mutual information maximization, a gating router, and achievement-based multi-task loss, to enable knowledge sharing and mitigate gradient conflicts across multi-modal information extraction tasks (MNER, MRE, MEE).
CodeClassificationGraph Neural NetworkTransformerGraphBiomedical Data
π― What it does: Propose Community-Aware Graph Transformer (CAGT), which identifies brain disorders by integrating the community structure and global topological information of brain functional networks.
Conditional Causal Representation Learning for Heterogeneous Single-cell RNA Data Integration and Prediction
Jiayi Dong (Fudan University), Fei Wang (Fudan University)
CodeRepresentation LearningData-Centric LearningAuto EncoderBiomedical Data
π― What it does: Propose the scConCRL framework, which integrates and predicts heterogeneous single-cell RNA-seq data using conditional causal representation learning.
Conditional Denoising Meets Polynomial Modeling: A Flexible Decoupled Framework for Time Series Forecasting
Jintao Zhang (University of Science and Technology of China), Daoyu Wang (University of Science and Technology of China)
CodeConvolutional Neural NetworkDiffusion modelTime Series
π― What it does: This paper proposes a Conditional Denoising Polynomial Modeling (CDPM) framework that can be end-to-end trained, decomposing time series into trend and seasonal components, where the seasonal fluctuations are captured by conditional denoising diffusion models and the trend is modeled using a polynomial trend module.
Consensus-Guided Incomplete Multi-view Clustering via Cross-view Affinities Learning
Qian Liu, Yang Wang (Hebei University)
CodeRepresentation LearningMultimodality
π― What it does: Proposed the CAL method, based on high-order interactions and consensus-guided cross-view affinity learning, to achieve incomplete multi-view clustering.
CodeGenerationComputational EfficiencyTransformerLarge Language ModelSequential
π― What it does: Designed and implemented the GeAI-BLAnC framework, which combines the probability distribution output by pre-trained sequence generation models (such as GPT-2, CMT) with the marginal probabilities calculated by constraint programming (CP), applying long-range structural constraints to generated sequences during inference.
Continuous Diffusive Prediction Network for Multi-Station Weather Prediction
Chujie Xu (Beihang University), Xianglong Liu (Beihang University)
CodeConvolutional Neural NetworkRecurrent Neural NetworkDiffusion modelOptical FlowTime SeriesPhysics Related
π― What it does: For multi-site weather forecasting, this paper proposes the Continuous Diffusion Prediction Network (CDPNet), which models spatial and temporal continuous weather changes through Continuous Calibration Initialization (CCI) and Diffusion Difference Estimation (DDE).
Contrastive Cross-Course Knowledge Tracing via Concept Graph Guided Knowledge Transfer
Wenkang Han (Zhejiang University), Jingyuan Chen (Zhejiang University)
CodeRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelContrastive LearningTextGraphSequential
π― What it does: This paper proposes a cross-course knowledge tracking model called TransKT, which constructs a cross-course concept graph using zero-shot large language models, achieves knowledge transfer through LLM-to-LM semantic enhancement and graph convolutional networks, and enhances knowledge state representation by incorporating cross-course contrastive learning.
π― What it does: Proposed the SSR model, enabling sequential recommendation on edge devices by converting dense embeddings into sparse spike representations to reduce memory and energy consumption.
Counterfactual Knowledge Maintenance for Unsupervised Domain Adaptation
Yao Li (China University of Mining and Technology), Bing Liu (China University of Mining and Technology)
CodeDomain AdaptationTransformerVision Language ModelImage
π― What it does: Propose an unsupervised domain adaptation framework based on counterfactual causal disentanglement and discriminative knowledge preservation
CRAFT: Time Series Forecasting with Cross-Future Behavior Awareness
Yingwei Zhang (Chinese Academy of Sciences), Detao Lv (Alibaba Group)
CodeTime Series
π― What it does: Proposes the CRAFT time series prediction framework, which leverages cross-future behavior (CFB) features to mine and predict future trends of the target sequence by utilizing trend information from CFB.
DANCE: Resource-Efficient Neural Architecture Search with Data-Aware and Continuous Adaptation
Maolin Wang (City University of Hong Kong), Xiangyu Zhao
CodeNeural Architecture SearchImage
π― What it does: The DANCE framework proposes a NAS method based on continuous distribution learning, achieving dynamic adaptation to different computational constraints by learning a continuously sampleable architecture distribution.
π― What it does: Construct a dual-conditional dual-stream network (DcDsDiff) based on diffusion models, simultaneously generating mask maps and detail maps to achieve generative image tampering localization.
π― What it does: This paper proposes the Decision-Aware Preference Modeling (DAPM) framework for multi-behavior recommendation, which constructs a behavior-agnostic graph to complement behavior-specific representations, and enhances the modeling of multi-behavior user preferences through decision thresholds and improved contrastive learning.
Deduction with Induction: Combining Knowledge Discovery and Reasoning for Interpretable Deep Reinforcement Learning
Haodi Zhang (Shenzhen University), Fangzhen Lin (Hong Kong University of Science and Technology)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerReinforcement LearningVideo
π― What it does: Propose a neural-symbolic framework HRL-ID that integrates automatic rule induction with logical reasoning into hierarchical reinforcement learning, enhancing the explainability and training efficiency of DRL.
π― What it does: This paper proposes a unified evaluation framework for assessing pedestrian crowd simulators based on deep learning and traditional knowledge-driven models, and systematically experiments on their performance under limited real data.
π― What it does: Designed a fully human-score-free deep BIQA model called DUBMA, which first learns through pseudo labels generated by synthetic image pairs from multiple FR-IQA models, and then enhances the model's adaptability to real images using unsupervised domain adaptation.
DERI: Cross-Modal ECG Representation Learning with Deep ECG-Report Interaction
Jian Chen (Shenzhen MSU-BIT University), Xiping Hu (Shenzhen MSU-BIT University)
CodeRepresentation LearningTransformerLarge Language ModelAuto EncoderContrastive LearningMultimodalityElectronic Health RecordsElectrocardiogram
π― What it does: Learn deep interactive representations across modalities of ECG and clinical reports, generating ECG representations for zero-shot classification and report generation.
Detecting Hallucination in Large Language Models Through Deep Internal Representation Analysis
Luan Zhang (Beijing Institute of Technology), Shuhao Zhang (Huazhong University of Science and Technology)
CodeAnomaly DetectionExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: This paper proposes the MHAD (Model Hallucination Awareness for Hallucination Detection) method, which detects hallucinations by analyzing internal representations of LLMs during the generation process (including attention outputs, feed-forward network outputs, and layer outputs), and constructs the SOQHD (Sustainable Open-Domain QA Hallucination Detection) benchmark dataset, providing internal representations and hallucination labels across multiple LLMs to ensure timeliness consistency.
DGCPL: Dual Graph Distillation for Concept Prerequisite Relation Learning
Miao Zhang (Hubei University), Shihui Wang (Hubei University)
CodeKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelGraphBenchmark
π― What it does: Construct and train a dual graph distillation model (DGCPL) to learn concept prerequisite relationships through a concept-resource hypergraph and a learning behavior graph.
DGExplainer: Explaining Dynamic Graph Neural Networks via Relevance Back-propagation
Yezi Liu (University of California Irvine), Yanning Shen (University of California Irvine)
CodeExplainability and InterpretabilityRecurrent Neural NetworkGraph Neural NetworkGraphTime Series
π― What it does: Propose DGExplainer, a method utilizing spatiotemporal hierarchical relevance propagation (LRP) to explain predictions of dynamic graph neural networks (Dynamic GNN) in link prediction and node regression tasks.
π― What it does: Propose the DGL model, which solves vehicle routing problems (VRP) by dynamically aggregating global and local information, and enhances robustness through replacement-based self-improvement learning (SIL).
DGraFormer: Dynamic Graph Learning Guided Multi-Scale Transformer for Multivariate Time Series Forecasting
Han Yan (Ocean University of China), Yanwei Yu (Ocean University of China)
CodeGraph Neural NetworkTransformerTime Series
π― What it does: Propose a multi-variable time series forecasting model called DGraFormer, which combines dynamic graph learning with multi-scale Transformer.
π― What it does: This paper proposes a hierarchical alignment-based deep graph kernel called DHTAGK, which generates transferable embeddings across graphs using an autoencoder.
π― What it does: Proposed a self-supervised learning framework called DiffECG based on diffusion models for efficient detection of arrhythmias and achieved personalized diagnosis
π― What it does: DiffFERV proposes a diffusion model-based facial video editing framework that achieves precise preservation and consistent editing of facial identity, motion, and background;
π― What it does: Propose a next-location recommendation framework called DPRL, which first decouples POI from its spatial/temporal context and separately learns sequence features, then integrates user preferences for POI and regions through a space-temporal aggregation mechanism, and finally enhances prediction using time queries.
Distribution-Aware Online Learning for Urban Spatiotemporal Forecasting on Streaming Data
Chengxin Wang (National University of Singapore), Beng Chin Ooi
CodeDomain AdaptationComputational EfficiencyGraph Neural NetworkTransformerTime Series
π― What it does: Propose a distribution-aware online learning framework named DOL to address progressive distribution shift and location-specific shift in urban spatiotemporal flows, providing an Awake-Hibernate update strategy based on the Streaming Update Mechanism and location-specific learners within the Adaptive ST network.
π― What it does: Proposes the DPMamba framework, achieving remote sensing image classification under missing modalities through knowledge distillation and learnable multimodal prompts (MMAP).
π― What it does: By leveraging rough masks and anomaly class information, using Stable Diffusion combined with attention mapping to automatically generate anomaly images and their precise mask labels, thus alleviating the problem of scarce real anomaly samples.
π― What it does: The paper proposes a training-free dynamic token manipulation method called DToMA, aiming to improve efficiency and understanding in long video comprehension tasks.
π― What it does: Propose a dual-agent reinforcement learning framework, DARL, for automatically generating and selecting features, while enhancing state representation through self-attention.
Dual-level Fuzzy Learning with Patch Guidance for Image Ordinal Regression
Chunlai Dong (Zhejiang University), Jian Wu (University Of Notre Dame)
CodeClassificationTransformerImageBiomedical Data
π― What it does: This work proposes a dual-layer fuzzy learning and patch-guided framework (DFPG), which achieves precise image ordinal regression grading by leveraging patch-level pseudo labels and fuzzy logic;
Dual-Perspective United Transformer for Object Segmentation in Optical Remote Sensing Images
Yanguang Sun (Nanjing University of Science and Technology), Lei Luo (Nanjing University of Science and Technology)
CodeSegmentationTransformerImage
π― What it does: This paper proposes a Dual-Perspective Unified Transformer (DPU-Former) for target segmentation in optical remote sensing images, and presents a complete encoder-decoder architecture.
DualCast: A Model to Disentangle Aperiodic Events from Traffic Series
Xinyu Su (University of Melbourne), Jianzhong Qi (University of Melbourne)
CodeGraph Neural NetworkTransformerTime Series
π― What it does: This paper proposes the DualCast framework, which decomposes traffic sequences into periodic and environmental (non-periodic) signals using a dual-branch structure, thereby improving prediction accuracy.
DUQ: Dual Uncertainty Quantification for Text-Video Retrieval
Xin Liu (Southwestern University of Finance and Economics), Yee-Hong Yang (University of Alberta)
CodeRetrievalTransformerVision Language ModelMultimodality
π― What it does: Studied the problem of uncertainty in text-video retrieval, proposing the Dual Uncertainty Quantification (DUQ) framework to improve retrieval performance.
Dynamic Anchor-based Ensemble Clustering via Hypergraph Reconstruction
Jiaxuan Xu (Sichuan University), Liang Du (Shanxi University)
CodeComputational EfficiencyImageTabular
π― What it does: This paper proposes an ensemble clustering method called YACHT, which combines dynamic anchor learning with hypergraph reconstruction to improve clustering accuracy without relying on original features.
Ben Bals (Centrum Wiskunde & Informatica), George Skretas (Hasso Plattner Institute)
CodeOptimizationExplainability and InterpretabilityGraph
π― What it does: Proposed and studied a model for recovering temporal graphs from infection trajectories, and designed the DiscoveryFollow algorithm based on Ξ΄-edge-connected components;
π― What it does: Proposes the ECC-SNN framework, which integrates SNN with cloud-based ANN to collaborate, leveraging the low energy consumption of edge devices and the high precision of the cloud, supporting incremental learning and dynamic environmental adaptation.