IJCAI 2025 Papers — Page 9
International Joint Conference on Artificial Intelligence · 1014 papers
RetroMoE: A Mixture-of-Experts Latent Translation Framework for Single-step Retrosynthesis
Xinjie Li (Pennsylvania State University), Abhinav Verma (Pennsylvania State University)
Drug DiscoveryGraph Neural NetworkTransformerMixture of ExpertsAuto EncoderGraph
🎯 What it does: Propose RetroMoE, a generative model that treats single-step retrosynthesis tasks as potential translation;
Revealing Concept Shift in Spatio-Temporal Graphs via State Learning
Kuo Yang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
Graph Neural NetworkGraphTime Series
🎯 What it does: The paper proposes a dynamic graph learning framework named Samen based on state inference, aiming to address the concept drift problem in dynamic graphs;
Revisiting Proportional Allocation with Subsidy: Simplification and Improvements
Xiaowei Wu (IOTSC University of Macau), Shengwei Zhou (IOTSC University of Macau)
Optimization
🎯 What it does: This paper studies subsidized fair allocation problems and proposes an optimal subsidy upper bound scheme for any number of agents and any mix of goods/bads.
Rewarding Explainability in Drug Repurposing with Knowledge Graphs
Susana Nunes (University of Lisbon), Catia Pesquita (University of Lisbon)
Explainability and InterpretabilityDrug DiscoveryRecurrent Neural NetworkReinforcement LearningGraphBiomedical Data
🎯 What it does: Propose a reinforcement learning-based knowledge graph path search method called REx, designed to generate scientifically interpretable paths for drug repurposing prediction;
Richer Semantics, Better Alignment: Aligning Visual Features with Explicit and Enriched Semantics for Visible-Infrared Person Re-Identification
Neng Dong (Nanjing University of Science and Technology), Jinhui Tang (Nanjing University of Science and Technology)
RetrievalConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Propose the RSBA framework, which significantly enhances VIReID performance by aligning visual features of visible and infrared images into a semantic space through explicit semantic guidance, indirect alignment, and cross-view semantic compensation.
Riding the Wave: Multi-Scale Spatial-Temporal Graph Learning for Highway Traffic Flow Prediction Under Overload Scenarios
Xigang Sun (Southeast University), Jun Zhu (Zhejiang Mobile Digital Intelligence Technology Co., Ltd.)
Autonomous DrivingGraph Neural NetworkTransformerContrastive LearningGraphTime Series
🎯 What it does: Propose a traffic flow prediction framework HST-WAVE for highway overloading scenarios, combining multi-scale woven Transformer, coupled heterogeneous graph attention network, and adaptive temporal contrastive learning;
Risk-Aware Task Migration for Multiplex Unmanned Swarm Networks in Adversarial Environments
Kai Di (Zhejiang Normal University), Yichuan Jiang (Southeast University)
OptimizationRobotic IntelligenceGraph
🎯 What it does: Designed and implemented a dual-scale risk-aware task migration algorithm to balance task loads in multi-layer drone swarms under adversarial environments.
RLBCD: Residual-guided Latent Brownian-bridge Co-Diffusion for Anatomical-to-Metabolic Image Synthesis
Tianxu Lv (Jiangnan University), Xiang Pan (Jiangnan University)
Image TranslationDiffusion modelAuto EncoderImageBiomedical DataComputed TomographyPositron Emission Tomography
🎯 What it does: Developed a residual-guided latent Brownian bridge co-diffusion model called RLBCD for generating metabolic images from anatomical images (A2MIS).
RLMiniStyler: Light-weight RL Style Agent for Arbitrary Sequential Neural Style Generation
Jing Hu (Chengdu University of Information Technology, China), Xin Wang (University at Albany, SUNY, USA)
Image TranslationGenerationReinforcement LearningAuto EncoderContrastive LearningImage
🎯 What it does: Proposed a lightweight arbitrary style transfer framework RLMiniStyler based on reinforcement learning.
Robult: Leveraging Redundancy and Modality-Specific Features for Robust Multimodal Learning
Duy A. Nguyen (UIUC), Minh N. Do (UIUC)
Representation LearningContrastive LearningMultimodality
🎯 What it does: Designed the Robult framework to address missing modalities and semi-supervised learning issues in multi-modal learning.
Robust Finite-Memory Policy Gradients for Hidden-Model POMDPs
Maris F. L. Galesloot (Radboud University Nijmegen), Nils Jansen (Ruhr-University Bochum)
OptimizationReinforcement LearningTabularSequentialBenchmark
🎯 What it does: Investigated the hidden model POMDP (HM-POMDP) scenario, proposing the robust finite-memory policy gradient (RFPG) algorithm to achieve robust optimal policies across multiple model sets.
Robust Graph Contrastive Learning for Incomplete Multi-view Clustering
Deyin Zhuang, Zhenwen Ren (Southwest University of Science and Technology)
Representation LearningGraph Neural NetworkAuto EncoderContrastive LearningGraph
🎯 What it does: Proposed a robust graph contrastive learning framework RGCL to address the graph noise correspondence problem in incomplete multi-view clustering, completing missing data, constructing relational graphs, performing contrastive learning, and achieving graph-level alignment.
Robust Misinformation Detection by Visiting Potential Commonsense Conflict
Bing Wang (Jilin University), Shengsheng Wang (Jilin University)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Designed a plug-in augmentation method for misinformation detection called MD-PCC, which expresses potential common-sense conflicts as text increments to enhance detection effectiveness.
RobustHAR: Multi-scale Spatial-temporal Masked Self-supervised Pre-training for Robust Human Activity Recognition
Xiao Liu (China University of Mining and Technology), Qiuyan Yan (China University of Mining and Technology)
RecognitionTransformerAuto EncoderTime Series
🎯 What it does: Proposes a multi-scale spatiotemporal masked self-supervised pre-training framework called RobustHAR for robust motion recognition in the presence of sensor missing data.
Robustness in Single-Audience Value-based Abstract Argumentation: Complexity Results
Bettina Fazzinga (University of Calabria), Filippo Furfaro (University of Calabria)
🎯 What it does: Defined the concept of 'robustness' within the single-audience value-based abstract argumentation framework (AVAF) and proposed the corresponding k-robustness decision problem to evaluate the stability of argumentation outcomes under changing audience preferences.
Robustness to Spurious Correlations via Dynamic Knowledge Transfer
Xiaoling Zhou (Peking University), Shikun Zhang (Peking University)
Domain AdaptationConvolutional Neural NetworkTransformerReinforcement LearningImageTextBenchmark
🎯 What it does: By transferring knowledge in the deep feature space, the method generates real and counterfactual enhanced features using class distribution and semantic vectors of pseudo-related classes, and employs reinforcement learning to dynamically determine the direction and magnitude of transfer, thereby improving the model's robustness to pseudo-correlations.
Rolling in Classical Planning with Conditional Effects and Constraints
Matteo Cardellini (University of Genova), Enrico Giunchiglia (University of Genova)
🎯 What it does: Studied introducing 'rolling' techniques into classical planning with conditional effects (CEs) and constraints, i.e., executing multiple consecutive repetitions of the same action within a single step;
RoLocMe: A Robust Multi-agent Source Localization System with Learning-based Map Estimation
Thanh Dat Le (University of North Texas), Yan Huang (University of North Texas)
Convolutional Neural NetworkReinforcement LearningSimultaneous Localization and MappingImage
🎯 What it does: Proposed a multi-agent reinforcement learning system called RoLocMe, which combines SkipNet's full-area RSS estimation with VDN for cooperative localization.
RotateKV: Accurate and Robust 2-Bit KV Cache Quantization for LLMs via Outlier-Aware Adaptive Rotations
Zunhai Su (Tsinghua University), Kehong Yuan (Tsinghua University)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes RotateKV, aiming to achieve memory compression and acceleration for LLM inference through 2-bit KV cache quantization, addressing the performance degradation issue of existing KV quantization methods at extremely low bit widths.
Rotation Invariant Spatial Networks for Single-View Point Cloud Classification
Feng Luan (Shanghai Research Institute for Intelligent Autonomous Systems), Bin He (Tongji University)
ClassificationConvolutional Neural NetworkPoint Cloud
🎯 What it does: Designed and implemented a rotation-invariant point cloud classification network called RISpaNet, which can achieve high-precision classification from a single view and under arbitrary poses.
RPMIL: Rethinking Uncertainty-Aware Probabilistic Multiple Instance Learning for Whole Slide Pathology Diagnosis
Zhikang Zhao (East China Normal University), Jing Zhao (East China Normal University)
ClassificationAuto EncoderImageBiomedical Data
🎯 What it does: This paper proposes an uncertainty-aware probabilistic MIL framework called RPMIL for multiple instance learning (MIL) tasks on whole slide images (WSI), replacing traditional point estimates with probability distributions to aggregate instance features and perform classification.
RRG-Mamba: Efficient Radiology Report Generation with State Space Model
Xiaodi Hou (Dalian Maritime University), Yijia Zhang (Dalian Maritime University)
GenerationComputational EfficiencyTransformerLarge Language ModelVision Language ModelMultimodalityBiomedical Data
🎯 What it does: Propose the RRG-Mamba framework, combining Mamba and RoPE to generate radiology reports, addressing the trade-off between global dependency modeling and computational efficiency.
RTdetector: Deep Transformer Networks for Time Series Anomaly Detection Based on Reconstruction Trend
Xinhong Liu (Central South University), Ming Zhao (Central South University)
Anomaly DetectionTransformerTime Series
🎯 What it does: Proposed an RTdetector, a time series anomaly detection model based on Transformer, which enhances anomaly discrimination by leveraging reconstruction trends.
Rule-Guided Reinforcement Learning Policy Evaluation and Improvement
Martin Tappler (TU Wien), Ezio Bartocci (University of Vienna)
Explainability and InterpretabilityReinforcement Learning
🎯 What it does: This paper proposes the LEGIBLE framework, which evaluates and improves policies by extracting rules from deep reinforcement learning strategies, generalizing these rules using domain knowledge, and enforcing them at runtime.
Run Like a Neural Network, Explain Like k-Nearest Neighbor
Xiaomeng Ye (Berry College), David Crandall (Indiana University Bloomington)
ClassificationRecognitionExplainability and InterpretabilityConvolutional Neural NetworkRecurrent Neural NetworkImageText
🎯 What it does: Studied and improved the neural network k-nearest neighbors (NN-kNN) model, proposing a scalable, high-dimensional processable, and parameter-efficient version, and verified its interpretability and performance on image and text tasks.
S-EPOA: Overcoming the Indistinguishability of Segments with Skill-Driven Preference-Based Reinforcement Learning
Ni Mu (Tsinghua University), Qing-Shan Jia (Tsinghua University)
Reinforcement Learning from Human FeedbackReinforcement LearningSequentialBenchmark
🎯 What it does: Proposes a skill-based preference reinforcement learning framework, S-EPOA, to address the issue of segmental indistinguishability in PbRL;
SALE-MLP: Structure Aware Latent Embeddings for GNN to Graph-free MLP Distillation
Harsh Pal (Mastercard), Aakarsh Malhotra (Mastercard)
Computational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Designed a graph structure-aware graph-agnostic MLP distillation method called SALE-MLP, which maps node features into a latent space consistent with graph topology during training using unsupervised structural loss, followed by training a lightweight MLP for node classification and link prediction through soft label distillation.
Sample-Efficient Behavior Cloning Using General Domain Knowledge
Feiyu Zhu (Carnegie Mellon University), Reid Simmons (Carnegie Mellon University)
Large Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Leverage large language models to convert domain knowledge described in natural language by experts into structured strategy models, and enhance sample efficiency through behavioral cloning with a small number of demonstrations.
Sanitizing Backdoored Graph Neural Networks: A Multidimensional Approach
Rong Zhao (Jinan University), Jian Weng (Jinan University)
Anomaly DetectionSafty and PrivacyAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: This paper proposes a multidimensional anomaly detection framework called MAD to identify and remove backdoor triggers in graph neural networks, thereby improving model security.
SAP: Privacy-Preserving Fine-Tuning on Language Models with Split-and-Privatize Framework
Xicong Shen (Bytedance), Sheng Zhong (Nanjing University)
Federated LearningSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningTextFinance Related
🎯 What it does: Proposes the Split-and-Privatize (SAP) framework, combining Split Learning with dχ-Differential Privacy for privacy-preserving Parameter-Efficient Fine-Tuning of pre-trained language models (PLM);
Scalable Multi-Stage Influence Function for Large Language Models via Eigenvalue-Corrected Kronecker-Factored Parameterization
Yuntai Bao (Zhejiang University), Jianwei Yin (Zhejiang University)
Explainability and InterpretabilityComputational EfficiencyTransformerText
🎯 What it does: This paper proposes a Multi-Stage Influence Function to attribute predictions of fine-tuned large language models back to their pre-training data, achieving scalable computation through EK-FAC approximation and semantic similarity-based candidate screening;
Scalable Speed-ups for the SMS-EMOA from a Simple Aging Strategy
Mingfeng Li (Harbin Institute of Technology), Benjamin Doerr (Ecole Polytechnique)
OptimizationComputational EfficiencyBenchmark
🎯 What it does: Proposed a non-elite selection mechanism based on a simple aging strategy to accelerate the performance of the SMS-EMOA algorithm in multi-objective optimization.
SCNNs: Spike-based Coupling Neural Networks for Understanding Structural-Functional Relationships in the Human Brain
Shaolong Wei (Nantong University), Jiashuang Huang (Nantong University)
ClassificationConvolutional Neural NetworkSpiking Neural NetworkBiomedical DataMagnetic Resonance ImagingDiffusion Tensor Imaging
🎯 What it does: Proposed a SCNNs model based on spiking neural networks, using spiking neurons to simulate structure-function coupling to more realistically capture interbrain information transmission and dynamic behaviors.
SCOUT: Semi-supervised Camouflaged Object Detection by Utilizing Text and Adaptive Data Selection
Weiqi Yan (Xiamen University), Liujuan Cao (Xiamen University)
Object DetectionTransformerVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose a semi-supervised covert object detection framework named SCOUT, combining adaptive data augmentation and selection (ADAS) with a text fusion module (TFM), and construct the RefTextCOD dataset, providing image-level reference text.
Screening, Rectifying, and Re-Screening: A Unified Framework for Tuning Vision-Language Models with Noisy Labels
Chaowei Fang (Xidian University), Guanbin Li (Sun Yat-Sen University)
TransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningMultimodality
🎯 What it does: A unified three-step framework is proposed for fine-tuning pre-trained vision-language models under noisy labels: first, samples are 'filtered' through dual-layer semantic matching, dividing them into clean, ambiguous, and noisy samples; second, 'correction' is applied to noisy and ambiguous samples by generating pseudo-labels and fusing them with original labels; finally, 're-filtering' is performed using BLIP, eliminating self-verification bias via cross-validation and enhancing low-quality samples with Mixup.
scSiameseClu: A Siamese Clustering Framework for Interpreting Single-cell RNA Sequencing Data
Ping Xu (Chinese Academy of Sciences), Pengfei Wang (Chinese Academy of Sciences)
Explainability and InterpretabilityRepresentation LearningAuto EncoderContrastive LearningBiomedical Data
🎯 What it does: Propose a Siamese network-based clustering framework for single-cell RNA sequencing, named scSiameseClu, which achieves more robust and discriminative cell embeddings through dual data augmentation, a Siamese fusion module, and optimal transport clustering;
SCVBench: A Benchmark with Multi-turn Dialogues for Story-Centric Video Understanding
Sisi You (Nanjing University of Posts and Telecommunications), Bing-Kun Bao (Nanjing University of Posts and Telecommunications)
TransformerLarge Language ModelAgentic AIVision Language ModelVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose SCVBench, which evaluates video story-level understanding through multi-turn dialogue and event ordering tasks.
SDDiff: Boosting Radar Perception via Spatial-Doppler Diffusion
Shengpeng Wang (Huazhong University of Science and Technology), Wei Wang (Wuhan University)
Autonomous DrivingConvolutional Neural NetworkDiffusion modelPoint Cloud
🎯 What it does: Proposed the SDDiff model, which achieves radar point cloud extraction (PCE) and ego-vehicle velocity estimation (EVE) simultaneously through spatial-Doppler diffusion, significantly improving radar perception quality.
SE(3)-Equivariant Diffusion Models for 3D Object Analysis
Xie Min (Ningbo University), Chen Kangxin (Ningbo University)
RestorationGenerationPose EstimationDiffusion modelScore-based ModelAuto EncoderPoint CloudMesh
🎯 What it does: Proposed an SE(3)-equivariant diffusion model for pose denoising and reconstruction of fragmented 3D objects.
Secure and Efficient Watermarking for Latent Diffusion Models in Model Distribution Scenarios
Liangqi Lei (Beijing Institute of Technology), Qi Wu (University of Adelaide)
Safty and PrivacyAdversarial AttackSupervised Fine-TuningDiffusion modelAuto EncoderImageText
🎯 What it does: In large-scale model distribution scenarios, a secure and efficient watermarking scheme called DistriMark is proposed, achieving strong binding and non-bypassability of watermarks through secure controllers with random seed injection and VAE fine-tuning;
SecV: LLM-based Secure Verilog Generation with Clue-Guided Exploration on Hardware-CWE Knowledge Graph
Fanghao Fan (Hangzhou Dianzi University), Li Kuang (Central South University)
Safty and PrivacyAI Code AssistantTransformerLarge Language ModelTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the SecV framework, which leverages hardware CWE knowledge graphs and large language models to generate functionally correct and secure Verilog code.
Seeing the Unseen: Composing Outliers for Compositional Zero-Shot Learning
Chenchen Jing (Zhejiang University of Technology), Chunhua Shen (Zhejiang University of Technology)
ClassificationAnomaly DetectionPrompt EngineeringVision Language ModelImage
🎯 What it does: Propose a method called COMO, which synthesizes representations of unseen attribute-object combinations as anomaly samples during training. It utilizes two modules, compositional classification and original classification, to identify and distinguish between seen and unseen combinations in test images. Subsequently, different fusion weights are applied for different categories to achieve precise identification.
Seeking Proxy Point via Stable Feature Space for Noisy Correspondence Learning
Yucheng Xie (University of Electronic Science and Technology of China), Xiaofeng Zhu (University of Electronic Science and Technology of China)
RetrievalRepresentation LearningData-Centric LearningContrastive LearningMultimodality
🎯 What it does: Propose a robust learning framework called Seeking Proxy Point via Stable Feature Space (SPS) to address the problem of noisy correspondence in cross-modal retrieval.
Self-calibration Enhanced Whole Slide Pathology Image Analysis
Haoming Luo (Zhejiang University), Zunlei Feng (Zhejiang University)
ClassificationGraph Neural NetworkTransformerImageBiomedical Data
🎯 What it does: Proposed a self-calibrated enhanced whole-slide pathological image analysis framework called SEW, which integrates global thumbnails and local magnified regions' superpixel maps. Global structural and detailed features are jointly extracted through Transformer and GCN, achieving fast and accurate grading and prognosis prediction.
Self-Classification Enhancement and Correction for Weakly Supervised Object Detection
Yufei Yin (Hangzhou Dianzi University), Houqiang Li (University of Science and Technology of China)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: Proposes a self-classification enhancement and correction framework, which eliminates ambiguity in multi-stage multi-class classification by leveraging class-internal binary classification (ICBC) within the self-classification enhancement module (SCE), and improves weakly supervised object detection performance by correcting misclassifications during inference through the self-classification correction (SCC) algorithm.
Self-Consistent Model-based Adaptation for Visual Reinforcement Learning
Xinning Zhou (Tsinghua University), Jun Zhu (Tsinghua University)
Domain AdaptationReinforcement LearningWorld ModelImageVideo
🎯 What it does: Proposes Self-Consistent Model-based Adaptation (SCMA), a method that achieves adaptive enhancement for visual RL in visually disturbed environments through a denoising model based on unsupervised distribution matching.
Self-supervised End-to-end ToF Imaging Based on RGB-D Cross-modal Dependency
Weihang Wang (Soochow University), Fei Wen (Shanghai Jiao Tong University)
Depth EstimationRepresentation LearningConvolutional Neural NetworkAuto EncoderGenerative Adversarial NetworkContrastive LearningMultimodality
🎯 What it does: Propose a self-supervised end-to-end ToF imaging framework that does not require noise-clean depth pairs, leveraging cross-modal dependencies between RGB and depth as implicit supervision to achieve noise suppression and improved image fidelity.
Semantic-Guided Diffusion Model for Single-Step Image Super-Resolution
Zihang Liu (Beijing Institute of Technology), Hao Tang (Peking University)
Super ResolutionKnowledge DistillationDiffusion modelImage
🎯 What it does: Propose a semantic-guided single-step diffusion model called SAMSR for image super-resolution.
Semantic-Space-Intervened Diffusive Alignment for Visual Classification
Zixuan Li (Shandong University), Xiangxu Meng (Shandong University)
ClassificationDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes a cross-modal alignment method called SeDA based on diffusion models, which utilizes a semantic space bridge to gradually map visual features into the text feature space to enhance visual classification performance.
Semi-Clairvoyant Scheduling of Speculative Decoding Requests to Minimize LLM Inference Latency
Ruixiao Li (Xi'an Jiaotong University), Peng Li (Xi'an Jiaotong University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a semi-clairvoyant scheduling algorithm, LAPS-SD, to minimize inference latency in LLM inference services.
Sentiment-enhanced Multi-hop Connected Graph Attention Network for Multimodal Aspect-Based Sentiment Analysis
Linlin Zhu (Xi'an Jiaotong University), Liang He (Xi'an Jiaotong University)
ClassificationGraph Neural NetworkTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose Sentiment-enhanced Multi-hop Connected Graph Attention Network (MCG) for multimodal aspect-based sentiment analysis tasks, achieving more fine-grained sentiment and aspect identification through sentiment contrastive pre-training and multi-hop syntactic dependency graph attention.
SEP: A General Lossless Compression Framework with Semantics Enhancement and Multi-Stream Pipelines
Meng Wan (Chinese Academy of Sciences), Chong Li (Chinese Academy of Sciences)
CompressionTransformerImageTextAudio
🎯 What it does: Proposed a general-purpose lossless compression framework called SEP based on semantic enhancement and a multi-stream pipeline
SepALM: Audio Language Models Are Error Correctors for Robust Speech Separation
Zhaoxi Mu (Xi'an Jiaotong University), Gang Wang (Xi'an Jiaotong University)
RestorationKnowledge DistillationTransformerLarge Language ModelMultimodalityChain-of-ThoughtAudio
🎯 What it does: Proposed the SepALM framework, integrating four modules: speech separation, text-domain error correction, resynthesis, and alignment, achieving more robust speech separation in complex noise environments.
SeqPose: An End-to-End Framework to Unify Single-frame and Video-based RGB Category-Level Pose Estimation
Yuzhu Ji (Guangdong University of Technology), Haijun Zhang (Harbin Institute of Technology)
Pose EstimationDepth EstimationTransformerImageVideo
🎯 What it does: Proposed SeqPose, an end-to-end framework for unified single-frame and video RGB category-level object pose and size estimation.
Set-Based Retrograde Analysis: Precomputing the Solution to 28-card Bridge Double Dummy Deals
Isaac Stone (University of Alberta), Jonathan Schaeffer (University of Alberta)
OptimizationComputational EfficiencyTabular
🎯 What it does: Proposed a set-based backward search algorithm called Setrograde Analysis to perfectly solve the 28-card (7-trick) bridge double dummy problem.
SetKE: Knowledge Editing for Knowledge Elements Overlap
Yifan Wei (Beihang University), Angsheng Li (Beihang University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a knowledge editing framework called SetKE based on set editing, which can handle multiple fact triplets (i.e., knowledge element overlaps) sharing the same prefix at once, and construct a new evaluation dataset called EDITSET.
Settling the Complexity of Popularity in Additively Separable and Fractional Hedonic Games
Martin Bullinger (University of Oxford), Matan Gilboa (University of Oxford)
🎯 What it does: Studied the computational complexity of the existence problem of popular partitions in additively separable and fractional utility games;
Shaping a Stabilized Video by Mitigating Unintended Changes for Concept-Augmented Video Editing
Mingce Guo (Zhejiang University), Lechao Cheng (Hefei University of Technology)
GenerationConvolutional Neural NetworkTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelVideoText
🎯 What it does: Propose a text-driven video editing framework based on Concept-Enhanced Text Inversion (CATI) and Dual Prior Supervision (DPS), enabling plug-and-play editing using user-provided concept videos while maintaining spatial and temporal consistency.
Sharpness-aware Zeroth-order Optimization for Graph Transformers
Yang Liu (Academy of Mathematics and Systems Science Chinese Academy of Science), Shirui Pan (Griffith University)
OptimizationDrug DiscoveryGraph Neural NetworkTransformerGraph
🎯 What it does: Proposed and implemented a sharpness-aware zeroth-order optimizer (SZO) for graph Transformers, addressing the challenge of gradient estimation caused by non-differentiable operations.
ShortcutProbe: Probing Prediction Shortcuts for Learning Robust Models
Guangtao Zheng (University of Virginia), Aidong Zhang (University of Virginia)
Domain AdaptationExplainability and InterpretabilityData-Centric LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: Proposes a post-hoc group-free bias mitigation framework called ShortcutProbe, which can automatically detect and eliminate prediction shortcuts after model training to enhance robustness.
SIFAR: A Simple Faster Accelerated Variance-Reduced Gradient Method
Zhize Li (Singapore Management University)
OptimizationTabular
🎯 What it does: Propose a loopless accelerated variance-reduced gradient method named SIFAR for solving finite-sum convex and strongly convex optimization problems.
Simulate, Refine and Integrate: Strategy Synthesis for Efficient SMT Solving
Bingzhe Zhou (Nanjing University), Xiaoxing Ma (Nanjing University)
OptimizationComputational EfficiencyGraph Neural NetworkReinforcement LearningGraphBenchmark
🎯 What it does: SIRISMT automatically synthesizes SMT solving strategies using reinforcement learning and graph neural networks, significantly improving the efficiency of Z3.
Simulating Misinformation Diffusion on Social Media Through CoNVaI: A Textual- and Agent-Based Diffusion Model
Raquel Rodríguez-García (UNED), Álvaro Rodrigo (UNED)
Agentic AIText
🎯 What it does: This paper proposes a misinformation diffusion simulation model called CoNVaI, based on text content and agents, to more realistically reproduce the information spreading process on social media.
Single-Node Trigger Backdoor Attacks in Graph-Based Recommendation Systems
Runze Li (Tianjin University), Zhen Wang (Northwestern Polytechnical University)
Recommendation SystemAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: This paper proposes a backdoor attack method for graph recommendation systems based on a single-node trigger (SNT-Ba), which inserts a pseudo user node to enable attackers to significantly increase the exposure rate of target items to target users while maintaining the overall recommendation quality of the system.
Situational-Constrained Sequential Resources Allocation via Reinforcement Learning
Libo Zhang (University of Electronic Science and Technology of China), Jiamou Liu (University of Auckland)
OptimizationReinforcement LearningSequentialAgriculture Related
🎯 What it does: Propose a situational constraint-aware sequential resource allocation framework SCRL, achieving a balance between dynamic constraint satisfaction and resource efficiency through reinforcement learning.
Sketch Decompositions for Classical Planning via Deep Reinforcement Learning
Michael Aichmüller, Hector Geffner (RWTH Aachen University)
OptimizationGraph Neural NetworkReinforcement LearningBenchmark
🎯 What it does: By transforming the subgoal decomposition of classical planning problems (Sketch Decomposition) into a deep reinforcement learning task, using IW(k) search to generate successor state sets for training graph neural network policies, thereby achieving width-constrained general subgoal decomposition.
SketchAgent: Generating Structured Diagrams from Hand-Drawn Sketches
Cheng Tan (Westlake University), Stan Z. Li (Westlake University)
Image TranslationGenerationTransformerLarge Language ModelSupervised Fine-TuningAgentic AIVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes SketchAgent, a multi-agent system capable of automatically converting hand-drawn sketches into structured, executable diagram code; and constructs the Sketch2Diagram Benchmark, containing over 6,000 samples across 8 chart types;
Smart Contracts for Trustless Sampling of Correlated Equilibria
Togzhan Barakbayeva (Hong Kong University of Science and Technology), Karaneh Keypoor (University of Oxford)
Safty and PrivacyTabularFinance Related
🎯 What it does: Propose two protocols based on smart contracts, utilizing one-time optional multi-party OT and zkSNARK to achieve trustless, decentralized correlated equilibrium sampling.
Smoothed Online Convex Optimization with Delayed Feedback
Sifan Yang (Nanjing University), Lijun Zhang (Nanjing University)
OptimizationComputational EfficiencyTabular
🎯 What it does: Proposes Smelt-DOGD and its efficient version to address the dynamic regret problem in smooth online convex optimization (SOCO) with switching costs under arbitrary delayed feedback.
SocialMP: Learning Social Aware Motion Patterns via Additive Fusion for Pedestrian Trajectory Prediction
Tianci Gao (Zhengzhou University), Pei Lv (Zhengzhou University)
Autonomous DrivingExplainability and InterpretabilityGraph Neural NetworkTransformerVideo
🎯 What it does: Designed SocialMP, a new pedestrian trajectory prediction representation, combining interpretable motion patterns with social interactions under visual field rules via attention additive fusion; simultaneously introduced map loss to suppress trajectories entering infeasible regions.
Soft Reasoning Paths for Knowledge Graph Completion
Yanning Hou (National University of Defense Technology), Jian Huang (National University of Defense Technology)
Computational EfficiencyRepresentation LearningTransformerContrastive LearningGraph
🎯 What it does: Proposes the SRP-KGC method with Soft Reasoning Paths (SRP) and hierarchical ranking, using learnable embeddings to fill missing paths and improve the stability and accuracy of knowledge graph completion.
Solving Copyright Infringement on Short Video Platforms: Novel Datasets and an Audio Restoration Deep Learning Pipeline
Minwoo Oh (Sungkyunkwan University), Eunil Park (Sungkyunkwan University)
RestorationRetrievalData-Centric LearningConvolutional Neural NetworkRecurrent Neural NetworkTransformerSupervised Fine-TuningContrastive LearningVideoAudio
🎯 What it does: Propose a complete pipeline that integrates music source separation with cross-modal video music retrieval for removing arbitrary background music and recovering original audio tracks from short videos.
Solving MDPs with LTLf+ and PPLTL+ Temporal Objectives
Giuseppe De Giacomo (University of Oxford), Pian Yu (University College London)
OptimizationReinforcement Learning
🎯 What it does: Propose an automaton construction method based on LTLf+ and PPLTL+ to solve optimal strategies in Markov Decision Processes (MDP) that satisfy infinite trace objectives.
Solving QNP and FOND+ with Generating, Testing and Forbidding
Zheyuan Shi (Sun Yat-sen University), Yongmei Liu (Sun Yat-sen University)
OptimizationBenchmark
🎯 What it does: Propose a Generate-Test-and-Forbid framework based on PRP to solve QNP and FOND+ planning problems, implementing SIEVE* termination testing and various optimization techniques.
SOTA: Spike-Navigated Optimal TrAnsport Saliency Region Detection in Composite-bias Videos
Wenxuan Liu (Peking University), Tiejun Huang (Peking University)
SegmentationSpiking Neural NetworkGenerative Adversarial NetworkVideo
🎯 What it does: Propose a visual saliency detection framework named SOTA based on spiking cameras, utilizing two modules: micro-debias (Spike-based Micro-Debias) and macro-debias (Spike-based Global-Debias), to balance temporal details and spatial global consistency.
SourceDetMamba: A Graph-aware State Space Model for Source Detection in Sequential Hypergraphs
Le Cheng (Northwestern Polytechnical University), Keke Tang
Anomaly DetectionGraph Neural NetworkGraphSequential
🎯 What it does: This paper proposes SourceDetMamba, which models reverse-ordered network snapshots using a hypergraph structure and the Mamba state space model, integrating a graph-aware state update mechanism to achieve efficient detection of rumor source nodes in social networks.
Spatial-Spectral Similarity-Guided Fusion Network for Pansharpening
Jiazhuang Xiong (China University of Geosciences), Lefei Zhang (Wuhan University)
Super ResolutionConvolutional Neural NetworkTransformerImage
🎯 What it does: Designed and implemented a multi-branch and similarity-constrained spatial-spectral fusion network (S3FNet) to fuse low-resolution multispectral images with high-resolution panchromatic images to generate high-resolution multispectral images;
Spatially Resolved Transcriptomics Data Clustering with Tailored Spatial-scale Modulation
Yuang Xiao (China University of Geosciences), Xinwang Liu (National University of Defense Technology)
Graph Neural NetworkBiomedical Data
🎯 What it does: Propose the TSstc network, utilizing adaptive multi-scale spatial graph construction and spatially aware sampling to achieve clustering of spatial transcriptomic data.
Spatio-temporal Prototype-based Hierarchical Learning for OD Demand Prediction
Shilu Yuan (Shandong University), Yongshun Gong (Shandong University)
Graph Neural NetworkTime Series
🎯 What it does: Propose the STPro dual-branch hierarchical model, combining micro-level node features with macro-level spatiotemporal prototypes for OD demand prediction.
SpectralGap: Graph-Level Out-of-Distribution Detection via Laplacian Eigenvalue Gaps
Jiawei Gu (Nanjing University of Science and Technology), Zechao Li (Nanjing University of Science and Technology)
Anomaly DetectionGraph Neural NetworkGraph
🎯 What it does: This paper proposes the SpecGap method, which utilizes the difference between the largest and second-largest eigenvalues (spectral gap) of the graph Laplacian matrix to perform post-processing detection of graph-level outliers/distributional out-of-distribution (OOD) samples;
Speeding Up Hyper-Heuristics With Markov-Chain Operator Selection and the Only-Worsening Acceptance Operator
Abderrahim Bendahi (Ecole Polytechnique), Johannes F. Lutzeyer (Institut Polytechnique De Paris)
OptimizationBenchmark
🎯 What it does: This paper proposes two improvements: using a two-state Markov chain to dynamically switch acceptance operators, and introducing an only-worsening operator that accepts only deteriorating solutions to help escape local optima;
SPoRt - Safe Policy Ratio: Certified Training and Deployment of Task Policies in Model-Free RL
Jacques Cloete (University of Oxford), Alessandro Abate (University of Oxford)
Safty and PrivacyReinforcement LearningSequential
🎯 What it does: This paper proposes the SPoRt framework, which provides an upper bound on the probability of safety property violations in model-free RL environments through the maximum policy ratio, and trains task-specific policies based on this upper bound.
Spotlighting Partially Visible Cinematic Language for Video-to-Audio Generation via Self-distillation
Feizhen Huang (Wuhan University), Bo Du (Wuhan University)
GenerationKnowledge DistillationDiffusion modelVideoAudio
🎯 What it does: Train a video-audio generation model using self-distillation, generating training pairs through simulated cinematic language (e.g., close-ups and camera movements) to address audio generation challenges when Foley targets are only partially visible.
Squeezing Context into Patches: Towards Memory-Efficient Ultra-High Resolution Semantic Segmentation
Wang Liu (Hunan University), Shutao Li (Hunan University)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a single-branch UHR semantic segmentation method called SCPSeg, which utilizes a Context Compression Module (CSM) to compress global information into local patches, while incorporating an auxiliary super-resolution decoder and a Local Feature Alignment loss (LFA) to achieve a balance between high accuracy and low memory consumption;
SRA-MCTS: Self-driven Reasoning Augmentation with Monte Carlo Tree Search for Code Generation
Bin Xu (Beijing Institute of Technology), Yang Gao (Beijing Institute of Technology)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Utilizes MCTS to autonomously generate intermediate reasoning steps (thinking) and perform self-evaluation to improve code generation.
SSTrack: Sample-interval Scheduling for Lightweight Visual Object Tracking
Yutong Kou (State Key Laboratory of Multimodal Artificial Intelligence Systems, Chinese Academy of Sciences), Jin Gao (People AI, Inc)
Object TrackingTransformerVideo
🎯 What it does: Designed a lightweight visual object tracker training framework called STrack, which enhances model performance through sample interval scheduling and gradient scaling.
ST-TAR: An Efficient Spatio-Temporal Learning Framework for Traffic Accident Risk Forecasting
Hongyu Wang (University of Electronic Science and Technology of China), Christian S. Jensen (Aalborg University)
Anomaly DetectionComputational EfficiencyGraph Neural NetworkTransformerContrastive LearningTabularTime SeriesSequential
🎯 What it does: Propose an efficient spatiotemporal learning framework named ST-TAR for traffic accident risk prediction.
ST-USleepNet: A Spatial-Temporal Coupling Prominence Network for Multi-Channel Sleep Staging
Jingying Ma (Chinese Academy of Sciences), Mengling Feng (Chinese Academy of Sciences)
ClassificationConvolutional Neural NetworkGraph Neural NetworkTime SeriesBiomedical Data
🎯 What it does: Proposed ST-USleepNet, a method that constructs multi-channel raw sleep signals into a spatiotemporal graph and utilizes a U-shaped network to simultaneously extract significant temporal features and spatial brain networks for sleep staging.
Stability and Generalization for Stochastic (Compositional) Optimizations
Xiaokang Pan (Central South University), Zhe Qu (Central South University)
Optimization
🎯 What it does: Explore the generalization performance of estimators in stochastic (combinatorial) optimization, propose a general framework to analyze the stability and generalization of various algorithms, and verify the impact of estimators on convergence rates and generalization;
Stabilizing Holistic Semantics in Diffusion Bridge for Image Inpainting
Jinjia Peng (Hebei University), Huibing Wang (Dalian Maritime University)
RestorationDiffusion modelScore-based ModelImage
🎯 What it does: Proposes a global structure guidance-based diffusion bridge framework called GSGDiff for image inpainting.
Stackelberg vs. Nash in the Lottery Colonel Blotto Game
Yan Liu (Renmin University of China), Jie Zhang (University of Bath)
Optimization
🎯 What it does: Propose modeling the Lottery Colonel Blotto game as a Stackelberg game, solving for the leader's optimal commitment strategy and comparing it with Nash equilibrium.
STAMImputer: Spatio-Temporal Attention MoE for Traffic Data Imputation
Yiming Wang (Beihang University), Guanlin Wu (National University of Defense Technology)
RestorationTransformerMixture of ExpertsTime Series
🎯 What it does: Designed a spatiotemporal attention hybrid model called STAMImputer based on Mixture of Experts for traffic data missing value imputation, and constructed real-time dynamic graphs through Low-rank Guided Sampling Graph Attention (LrSGAT) to dynamically balance spatiotemporal feature extraction.
StarFT: Robust Fine-tuning of Zero-shot Models via Spuriosity Alignment
Younghyun Kim (Samsung), Jinwoo Shin (KAIST)
ClassificationDomain AdaptationRepresentation LearningLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Introduce 'Spurious Textual Alignment Regularization (StarFT)' during fine-tuning of large zero-shot vision-language models (e.g., CLIP) to suppress the learning of spurious features, thereby enhancing out-of-distribution (OOD) robustness.
State Feedback Enhanced Graph Differential Equations for Multivariate Time Series Forecasting
Jiaxu Cui (Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education), Bo Yang (Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education)
Graph Neural NetworkTime SeriesOrdinary Differential Equation
🎯 What it does: This paper proposes a state feedback-based graph differential equation (SF-GDE) for multivariate time series prediction, aiming to alleviate the over-smoothing problem in traditional graph networks.
State Revisit and Re-explore: Bridging Sim-to-Real Gaps in Offline-and-Online Reinforcement Learning with An Imperfect Simulator
Xingyu Chen (Xi'an Jiaotong University), Xuguang Lan (Xi'an Jiaotong University)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: Propose a State Revisit and Re-explore (SR²) framework based on meta-policy and sub-policy, which utilizes the meta-policy to select high-quality states from offline trajectories and re-explore them in an imperfect simulator, thereby narrowing the sim-to-real gap in hybrid offline-online reinforcement learning.
Steady-State Strategy Synthesis for Swarms of Autonomous Agents
Martin Jonáš (Masaryk University), Jan Mačák (Masaryk University)
OptimizationGraph
🎯 What it does: This paper studies the steady-state strategy synthesis problem under multi-agent settings, proposing an incremental fully memoryless strategy synthesis algorithm and experimentally evaluating it on randomly generated strongly connected graphs.
STLSP: Integrating Structure and Text with Large Language Models for Link Sign Prediction of Networks
Lijia Ma (Shenzhen University), Jianqiang Li (Shenzhen University)
ClassificationGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraph
🎯 What it does: Propose the STLSP framework, integrating structural balance theory, node embeddings, and edge text, leveraging LLMs to accomplish link sign prediction.
Stochasticity-aware No-Reference Point Cloud Quality Assessment
Songlin Fan (Peking University), Qicheng Wang (Youiia Innov Tech Co Ltd)
Convolutional Neural NetworkAuto EncoderPoint Cloud
🎯 What it does: Propose a no-reference point cloud quality assessment framework based on Conditional Variational AutoEncoder (CVAE), which can generate multiple quality evaluation results and obtain the final quality score through averaging.
Strategies, Credences, and Shannon Entropy: Reasoning about Strategic Uncertainty in Stochastic Environments
Wojciech Jamroga (Polish Academy of Sciences), Aniello Murano (University of Naples Federico II)
🎯 What it does: Proposed the PATLH logic, combining Shannon entropy with PATL to model strategy uncertainty in multi-agent stochastic environments.
Strategy-Architecture Synergy: A Multi-View Graph Contrastive Paradigm for Consistent Representations
Shuman Zhuang (Fuzhou University), Ximeng Liu (Fuzhou University)
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Propose an adaptive edge deletion preprocessing and contrastive learning framework (CAMEL) for multi-view graph data, enhancing view consistency through edge deletion and designing a neighborhood consistency multi-view contrastive loss;