IJCAI 2024 Papers — Page 7
International Joint Conference on Artificial Intelligence · 790 papers
Quality-Diversity Algorithms Can Provably Be Helpful for Optimization
Chao Qian (Nanjing University), Ren-Jian Wang (Nanjing University)
Optimization
🎯 What it does: This paper proves through rigorous runtime analysis that the quality diversity (QD) algorithm MAP-Elites can achieve optimal or near-optimal approximation ratios in expected polynomial time for two classes of NP-hard problems (constrained monotonic submodular function maximization and set cover), whereas the (µ+1)-EA of the same scale requires exponential expected time to reach similar approximation ratios on certain instances.
Quantitative Claim-Centric Reasoning in Logic-Based Argumentation
Markus Hecher (Massachusetts Institute of Technology), Johannes Schmidt (Jonkping University)
🎯 What it does: This paper proposes counting the support set in logical foundational argumentation, using this to define the strength and possibility of propositions.
Quantitative Reasoning over Incomplete Abstract Argumentation Frameworks
Bettina Fazzinga (University of Calabria), Giuseppina Monterosso (University of Calabria)
🎯 What it does: Studied the problem of quantitative reasoning on incomplete abstract argumentation frameworks, proposed problems such as PERCVER, PERCACC, and CNTCOM, and analyzed their computational complexity.
R2V-MIF: Rule-to-Vector Contrastive Learning and Multi-channel Information Fusion for Therapy Recommendation
Nengjun Zhu (Shanghai University), Huanjing Gao (Shanghai University)
Recommendation SystemGraph Neural NetworkContrastive LearningGraphBiomedical Data
🎯 What it does: Proposed the R2V-MIF model, which achieves treatment recommendation through rule-to-vector contrastive learning and multi-channel information fusion.
Randomized Learning-Augmented Auctions with Revenue Guarantees
Ioannis Caragiannis (Aarhus University), Georgios Kalantzis (Aarhus University)
Optimization
🎯 What it does: This paper designs a class of randomized, prediction-augmented single-item auction mechanisms that guarantee the ratio of revenue to the highest valuation in the worst case, and provides the optimal trade-off between consistency and robustness;
Rank and Align: Towards Effective Source-free Graph Domain Adaptation
Junyu Luo (Peking University), Ming Zhang (Peking University)
Domain AdaptationGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Studied how to transfer pre-trained graph neural networks to a target graph domain when the source graph is unavailable, achieving source-free graph domain adaptation; proposed the Rank and Align (RNA) framework, which realizes semantic learning and domain alignment on the target domain through steps such as spectral seriation ranking, harmonic graph detection, subgraph extraction, and pseudo-label filtering.
Real-World Networks Are Low-Dimensional: Theoretical and Practical Assessment
Tobias Friedrich (University of Potsdam), Leon Schiller (ETH)
Graph
🎯 What it does: This paper proves through theoretical analysis and a linear-time algorithm that the hidden geometric dimension of real-world networks is extremely low (1~10 dimensions), and proposes a method to rapidly estimate this dimension on large-scale datasets.
RealDex: Towards Human-like Grasping for Robotic Dexterous Hand
Yumeng Liu (ShanghaiTech University), Yuexin Ma (ShanghaiTech University)
GenerationPose EstimationRobotic IntelligenceTransformerLarge Language ModelAuto EncoderImagePoint CloudSequential
🎯 What it does: Developed the RealDex dataset by synchronizing human hands with robot hands using a remote control system, achieving 52 objects, over 2,600 grasping sequences, 955k frames of multi-view RGB-D data, and proposed a grasping pose selection framework based on MLLM and a self-recursive MotionNet for grasping motion generation.
Recall, Retrieve and Reason: Towards Better In-Context Relation Extraction
Guozheng Li (Southeast University), Zijie Xu (Southeast University)
RetrievalKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a RE4 recall-retrieve-reason framework, which first generates entity pairs consistent with the input sentence as queries using open-source LLMs through ontological knowledge, then retrieves corresponding examples from training corpora, and finally performs in-context reasoning on the retrieved examples, significantly improving relation extraction performance.
Reconfigurability-Aware Selection for Contrastive Active Domain Adaptation
Zeyu Zhang (Jilin University), Shaojie Zhang (Jilin University)
ClassificationDomain AdaptationContrastive LearningImage
🎯 What it does: Proposes the RASC (Reconfigurability-Aware Selection for Contrastive Active Domain Adaptation) method, combining reconfigurability-aware sample selection with contrastive learning-based progressive domain adaptation, and introduces a one-annotation variant RASC-Ob
Reconstructing Missing Variables for Multivariate Time Series Forecasting via Conditional Generative Flows
Xuanming Hu (Arizona State University), Yanjie Fu (Arizona State University)
RestorationGenerationMeta LearningFlow-based ModelTime SeriesElectrocardiogram
🎯 What it does: This paper addresses the problem of time series prediction with missing variable subsets by proposing a missing variable reconstruction framework based on conditional generative flows, and enhances the model's generalization across different missing subsets through meta-learning.
Reconstruction Weighting Principal Component Analysis with Fusion Contrastive Learning
Qianqian Wang (Xidian University), Quanxue Gao (Xidian University)
ClassificationRepresentation LearningContrastive LearningImageBiomedical Data
🎯 What it does: This paper proposes an unsupervised feature extraction method that combines reconstruction-weighted principal component analysis (Weighted PCA) with contrastive learning;
Redefining Contributions: Shapley-Driven Federated Learning
Nurbek Tastan (Mohamed bin Zayed University of Artificial Intelligence), Karthik Nandakumar (Mohamed bin Zayed University of Artificial Intelligence)
Federated LearningExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper studies the use of Shapley values in federated learning to evaluate the contributions of participants, and proposes the ShapFed algorithm based on class-level Shapley values, the weighted aggregation method ShapFed-WA, and contribution-based personalized updates.
Reframing Spatial Reasoning Evaluation in Language Models: A Real-World Simulation Benchmark for Qualitative Reasoning
Fangjun Li (University of Leeds), Anthony G. Cohn (University of Leeds)
Data SynthesisTransformerLarge Language ModelTextTabularBenchmark
🎯 What it does: This paper proposes RoomSpace, a spatial reasoning evaluation benchmark based on 3D simulated scenarios, providing diverse object layouts, topological, directional, and distance relationships, and introducing a logical consistency checking tool;
Regression Residual Reasoning with Pseudo-labeled Contrastive Learning for Uncovering Multiple Complex Compositional Relations
Chengtai Li (University of Nottingham Ningbo China), Xudong Jiang (Nanyang Technological University)
Anomaly DetectionConvolutional Neural NetworkContrastive LearningImageBenchmark
🎯 What it does: This paper proposes a new abstract visual reasoning benchmark called MC2R, which requires models to identify two anomalous samples among five images, and introduces a model named R3PCL to address this task.
Reinforcement Learning from Diverse Human Preferences
Wanqi Xue (Nanyang Technological University), Zhongwen Xu (Tencent AI Lab)
Reinforcement Learning from Human FeedbackReinforcement LearningBenchmark
🎯 What it does: Improved the robustness of preference-based reinforcement learning under diverse human feedback by imposing constraints in the latent space of the reward model and using confidence-weighted ensembles.
Reinforcement Nash Equilibrium Solver
Xinrun Wang (Nanyang Technological University), Bo An (Nanyang Technological University)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper proposes a game modification method based on reinforcement learning—REinforcement Nash Equilibrium Solver (RENES), which trains a general strategy to modify normal-form games of any size, then applies existing approximate solvers (such as α-rank, CE, FP, PRD) to the modified game, ultimately obtaining solutions closer to Nash equilibrium in the original game.
ReinforceNS: Reinforcement Learning-based Multi-start Neighborhood Search for Solving the Traveling Thief Problem
Tao Wu (Huazhong University of Science and Technology), Yan Jin (Huazhong University of Science and Technology)
OptimizationReinforcement LearningGraphBenchmark
🎯 What it does: Propose the ReinforceNS algorithm, integrating reinforcement learning-driven multi-start neighborhood search, preprocessing, dedicated initialization, and post-optimization to solve the Traveling Thief Problem (TTP).
Relevant Irrelevance: Generating Alterfactual Explanations for Image Classifiers
Silvan Mertes (University of Augsburg), Elisabeth André (University of Augsburg)
ClassificationExplainability and InterpretabilityGenerative Adversarial NetworkImage
🎯 What it does: This study proposes a GAN-based generative framework to produce Alterfactual (modifiable irrelevant features) and Counterfactual (decision-changing) explanations for black-box image classifiers, which can preserve the original decision while maximizing modifications to irrelevant features.
ReliaAvatar: A Robust Real-Time Avatar Animator with Integrated Motion Prediction
Bo Qian (Xi'an Jiaotong University), Xing Wei (Xi'an Jiaotong University)
Pose EstimationRecurrent Neural NetworkTransformerTime SeriesSequential
🎯 What it does: This paper proposes a real-time, robust full-body avatar animator called ReliaAvatar, which can continuously generate complete skeleton motions even when wearable device signals are of low quality or missing.
Reschedule Diffusion-based Bokeh Rendering
Shiyue Yan (Tsinghua University), Shaojun Liu
Image TranslationDiffusion modelImage
🎯 What it does: Propose a prior-based sampling and noise scheduling method to improve diffusion models in Bokeh rendering from small aperture images to large aperture images.
Resolving Word Vagueness with Scenario-guided Adapter for Natural Language Inference
Yonghao Liu (Jilin University), Renchu Guan (Jilin University)
ClassificationConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Propose the ScenaFuse model, introducing contextual images in natural language inference (NLI) tasks. The model integrates visual information with text features through an adapter, achieving more accurate judgment of sentence relationships.
Rethinking Centered Kernel Alignment in Knowledge Distillation
Zikai Zhou (East China Normal University), Shaohui Lin (East China Normal University)
ClassificationObject DetectionKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: Investigated the theoretical foundations of Centered Kernel Alignment (CKA) in knowledge distillation, and proposed two concise and efficient distillation frameworks: relation-based CKA (RCKA) and patch-based CKA (PCKA), to achieve effective alignment of features and logit layers.
Rethinking Correlation Learning via Label Prior for Open Set Domain Adaptation
Zi-Xian Huang (Sun Yat-Sen University), Chuan-Xian Ren (Sun Yat-Sen University)
Domain AdaptationImageBenchmark
🎯 What it does: Proposes a Balanced Correlation Learning (BCL) framework based on the Hilbert-Schmidt Independence Criterion (HSIC) to address the open-set domain adaptation (OSDA) problem;
Rethinking the Effectiveness of Graph Classification Datasets in Benchmarks for Assessing GNNs
Zhengdao Li (Guangzhou University), Kai Hwang (Chinese University of Hong Kong)
ClassificationData SynthesisGraph Neural NetworkGraphBenchmark
🎯 What it does: Designed and implemented new experimental protocols and metrics to evaluate whether graph classification benchmark datasets can effectively distinguish GNNs from baseline methods, and proposed a new effectiveness metric E and a synthetic data generation framework with controllable correlation.
Rethinking the Soft Conflict Pseudo Boolean Constraint on MaxSAT Local Search Solvers
Jiongzhi Zheng (Huazhong University of Science and Technology), Kun He (Huazhong University of Science and Technology)
OptimizationBenchmark
🎯 What it does: Proposes integrating Soft conflict Pseudo Boolean constraints into the clause weighting system of local search MaxSAT and designs an adaptive weight update strategy.
Revealing the Two Sides of Data Augmentation: An Asymmetric Distillation-based Win-Win Solution for Open-Set Recognition
Yunbing Jia (Jilin University), Yi Yang (Chinese Academy of Sciences)
ClassificationRecognitionKnowledge DistillationImageBiomedical Data
🎯 What it does: This paper reveals that multi-sample data augmentation (MSA) improves performance in closed-set recognition but has a significant negative impact on open-set recognition (OSR), and proposes an asynchronous distillation framework. By introducing cross mutual information loss and two-hot label smoothing, where the teacher model receives original samples and the student model receives mixed samples, it achieves a win-win effect for both closed-set and open-set recognition.
Revising Beliefs and Intentions in Stochastic Environments
Nima Motamed (Utrecht University), Dragan Doder (Utrecht University)
Reinforcement Learning
🎯 What it does: This paper constructs a theoretical framework for the joint revision of belief and intention in a stochastic environment, and provides representations and definitions of belief and intention revision on the decidable probabilistic temporal logic PLBP+.
Revisiting Causal Discovery from a Complexity-Theoretic Perspective
Robert Ganian (TU Wien), Stefan Szeider (TU Wien)
Graph
🎯 What it does: This paper conducts a complexity theory analysis of constraint-based causal discovery algorithms under a fine-grained Oracle model, provides a quantitative characterization of the relationship between the size of d-separating sets and graph structure, and proposes a new fixed-parameter algorithm.
Revisiting Neural Networks for Continual Learning: An Architectural Perspective
Aojun Lu (Sichuan University), Yanan Sun (Sichuan University)
ClassificationNeural Architecture SearchConvolutional Neural NetworkImage
🎯 What it does: Systematically experiment on the impact of network architecture (width, depth, skip connections, global pooling, downsampling, etc.) in continual learning (Task IL and Class IL), discovering that wider, shallower, skip-connection-equipped networks with removed GAP (Task IL) or retained GAP (Class IL) are more favorable. Based on these insights, we propose the ArchCraft NAS method, automatically searching for lightweight networks suitable for continual learning (AlexAC, ResAC).
Revitalizing Real Image Deraining via a Generic Paradigm towards Multiple Rainy Patterns
Xin Li (Sun Yat-sen University), Zhuo Su (Sun Yat-sen University)
RestorationData SynthesisConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposed a general paradigm for real image rain removal, including a multi-type rain synthesis pipeline and a Pattern-aware Rain Removal Network (PRRN), which can effectively handle various real rain textures and exhibits strong generalization capability.
RoboFusion: Towards Robust Multi-Modal 3D Object Detection via SAM
Ziying Song (Beijing Jiaotong University), Li Wang (Beijing Institute of Technology)
Object DetectionAutonomous DrivingConvolutional Neural NetworkTransformerImageMultimodalityPoint Cloud
🎯 What it does: Proposes the RoboFusion framework, combining the visual foundation model SAM with multi-modal information to enhance the robustness of 3D object detection in noisy scenarios.
Robust Contrastive Multi-view Kernel Clustering
Peng Su (Sichuan University), Jiancheng Lv (Sichuan University)
OptimizationRepresentation LearningContrastive LearningMultimodality
🎯 What it does: This paper proposes a robust contrastive learning multi-view kernel clustering framework named R-CMK, aiming to enhance kernel quality by processing pseudo-negative samples through reverse gradients, thereby improving multi-view clustering performance.
Robust Heterophilic Graph Learning against Label Noise for Anomaly Detection
Junhang Wu (Wuhan University), Yilong Zang (Wuhan University)
Anomaly DetectionGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes the NRGL framework, which addresses graph anomaly detection tasks with graph heterogeneity and label noise. It first splits the original graph into homophilic (high homophily) and heterophilic views using an edge discriminator. Then, it obtains robust node representations through unsupervised contrastive learning and enhances the structure via low-pass/high-pass graph filters. Subsequently, it divides the training nodes into clean, confident, and remaining groups based on the small loss criterion and designs a sampler based on degree, class frequency, and confidence for balanced training.
Robust Losses for Decision-Focused Learning
Noah Schutte (Delft University of Technology), Neil Yorke-Smith (Delft University of Technology)
OptimizationTabularTime SeriesFinance Related
🎯 What it does: Proposed and evaluated three robust loss functions for decision-focused learning to alleviate generalization and bias issues caused by empirical regret failure.
Robust Reward Placement under Uncertainty
Petros Petsinis (Aarhus University), Panagiotis Karras (University of Copenhagen)
OptimizationGraph
🎯 What it does: Propose the Robust Reward Placement (RRP) problem under various Markov mobility models, and provide its NP-hardness and approximation upper bounds
ROCES: Robust Class Expression Synthesis in Description Logics via Iterative Sampling
N'Dah Jean Kouagou (Paderborn University), Axel-Cyrille Ngonga Ngomo (Paderborn University)
Data SynthesisTransformerGraph
🎯 What it does: Propose a new class expression learning problem GLP and introduce the ROCES algorithm, achieving efficient class expression reasoning for any number of samples through iterative sampling and neural synthesizers.
ROME: Robust Multi-Modal Density Estimator
Anna Mészáros (TU Delft), Jens Kober (TU Delft)
Representation LearningMultimodalityTabularSequential
🎯 What it does: Propose a robust multi-modal probability density estimation method called ROME, which is based on clustering, decorrelation, normalization, and then kernel density estimation.
RSAP-DFM: Regime-Shifting Adaptive Posterior Dynamic Factor Model for Stock Returns Prediction
Quanzhou Xiang (Fudan University), Rujun Jiang (Fudan University)
OptimizationRepresentation LearningRecurrent Neural NetworkTime SeriesFinance Related
🎯 What it does: Propose an adaptive posterior dynamic factor model RSAP-DFM based on continuous dual market state switching to predict returns of China's A-shares.
SACNN: Self Attention-based Convolutional Neural Network for Fraudulent Behaviour Detection in Sports
Maxx Richard Rahman (Saarland University), Wolfgang Maass (Saarland University)
Anomaly DetectionAdversarial AttackConvolutional Neural NetworkTime SeriesBiomedical Data
🎯 What it does: This study proposes a SACNN model based on self-attention and convolution to identify 'sample swapping' fraud in longitudinal steroid urine samples from athletes;
SAEIR: Sequentially Accumulated Entropy Intrinsic Reward for Cooperative Multi-Agent Reinforcement Learning with Sparse Reward
Xin He (Dalian University of Technology), Jincheng Yu (Dalian University of Technology)
Graph Neural NetworkReinforcement Learning
🎯 What it does: Proposed a novel recursive cumulative intrinsic reward SAEIR based on system entropy, combined with a multi-scale hypergraph Critic, to address the challenges of multi-agent cooperative learning in sparse reward environments.
Safety Constrained Multi-Agent Reinforcement Learning for Active Voltage Control
Yang Qu (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
OptimizationReinforcement LearningTime SeriesPhysics Related
🎯 What it does: Studied the active voltage control problem in distributed photovoltaic systems, and proposed a multi-agent reinforcement learning algorithm MA-DELC based on dual Lagrange constraints.
Sample Quality Heterogeneity-aware Federated Causal Discovery through Adaptive Variable Space Selection
Xianjie Guo (Hefei University of Technology), Xiaoxiao Li (The University of British Columbia)
Federated LearningGraphTabularBiomedical DataBenchmark
🎯 What it does: Propose the FedACD method to address the issue of sample quality heterogeneity across different variable spaces in federated causal discovery, using adaptive masking to transmit only the causal relationships learned from high-quality variable spaces.
Sampling Winners in Ranked Choice Voting
Matthew Iceland (University of Rochester), Joseph Saber (University of Rochester)
Tabular
🎯 What it does: The study investigates the feasibility of using ranking voting (RCV) with sample voting to predict complete election outcomes, and evaluates its prediction accuracy at both theoretical and empirical levels.
SaSDim:Self-Adaptive Noise Scaling Diffusion Model for Spatial Time Series Imputation
Shunyang Zhang (Central South University), Jianxin Wang (Central South University)
Data-Centric LearningConvolutional Neural NetworkGraph Neural NetworkTransformerDiffusion modelGraphTime SeriesStochastic Differential Equation
🎯 What it does: This paper proposes SaSDim—an adaptive noise scaling diffusion model for spatiotemporal sequence missing value imputation.
Scalable Federated Unlearning via Isolated and Coded Sharding
Yijing Lin (Beijing University of Posts and Telecommunications), Jinke Ren (Chinese University of Hong Kong (Shenzhen))
Federated LearningImageText
🎯 What it does: Proposing a scalable federated forgetting framework based on phased-based isolated partitioning and encoding computation
Scalable Landmark Hub Labeling for Optimal and Bounded Suboptimal Pathfinding
Sabine Storandt (Universitat Konstanz)
OptimizationGraph
🎯 What it does: This paper proposes a scalable Landmark Hub Labeling (LHL) algorithm that addresses the preprocessing and query bottlenecks of LHL on large-scale graphs.
Scalable Mechanism Design for Multi-Agent Path Finding
Paul Friedrich (ETH AI Center), Sven Seuken (ETH AI Center)
OptimizationGraphBenchmark
🎯 What it does: Propose a scalable mechanism design framework in multi-agent path finding (MAPF) to achieve strategyproofness and individual rationality (IR), and design three mechanisms: PCBS, EPBS, and MCPP.
Scalable Ultrafast Almost-optimal Euclidean Shortest Paths
Stefan Funke (University of Stuttgart), Felix Weitbrecht (University of Stuttgart)
OptimizationComputational Efficiency
🎯 What it does: In a 2D plane obstacle environment, a refined constrained Delaunay triangulation is constructed, combined with contraction hierarchies (CH) and local path optimization, achieving millisecond-level query times for approximate Euclidean shortest paths even with millions of obstacles.
Scale and Direction Guided GAN for Inertial Sensor Signal Enhancement
Yifeng Wang (Harbin Institute of Technology), Yi Zhao (Harbin Institute of Technology)
RestorationPose EstimationGenerative Adversarial NetworkTime Series
🎯 What it does: This paper proposes an unpaired data inertial sensor signal enhancement method called SDG-GAN, and designs scale-guided (SG-GAN) and direction-guided (DG-GAN) mechanisms in different experimental scenarios to achieve high-quality reconstruction and denoising of low-cost sensor signals.
Scaling Up Unbiased Search-based Symbolic Regression
Paul Kahlmeyer (Friedrich Schiller University Jena), Henrik Voigt (Friedrich Schiller University Jena)
OptimizationData-Centric LearningBenchmarkPhysics Related
🎯 What it does: A symbolic regression method based on unbiased expression DAG search is studied, and a variable expansion technique is proposed to reduce the search space
SCAT: A Time Series Forecasting with Spectral Central Alternating Transformers
Chengjie Zhou (Dalian University of Technology), Qiang Zhang (Dalian University of Technology)
TransformerTime Series
🎯 What it does: Propose a Transformer-based time series forecasting model named SCAT, which enhances multi-variable prediction accuracy by leveraging spectral clustering centers and alternating attention mechanisms.
Scene-Adaptive Person Search via Bilateral Modulations
Yimin Jiang (Dalian Maritime University), Yang Wang (Hefei University of Technology)
RetrievalDomain AdaptationConvolutional Neural NetworkTransformerContrastive LearningImageBenchmark
🎯 What it does: Proposed a scene-adaptive person search framework named SEAS, which eliminates scene noise through bilateral modulation and maintains feature consistency of the same person across different scenes;
SceneDiff: Generative Scene-Level Image Retrieval with Text and Sketch Using Diffusion Models
Ran Zuo (Beijing Key Laboratory of Human-Computer Interaction, Institute of Software, Chinese Academy of Sciences), Hongan Wang (Beijing Key Laboratory of Human-Computer Interaction, Institute of Software, Chinese Academy of Sciences)
RetrievalTransformerVision Language ModelDiffusion modelContrastive LearningImageText
🎯 What it does: Propose SceneDiff, a scene-level text+sketch retrieval framework based on pre-trained diffusion models, which first encodes text, sketches, and images, projects them into the diffusion latent space, generates latent fused features, and achieves retrieval through dual-level contrastive learning.
Score-CDM: Score-Weighted Convolutional Diffusion Model for Multivariate Time Series Imputation
Shunyang Zhang (Central South University), Jian Zhang (Central South University)
Data SynthesisConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelTime Series
🎯 What it does: Proposed a diffusion model called Score-CDM based on score-weighted convolution for imputing missing values in multivariate time series.
ScreenAgent: A Vision Language Model-driven Computer Control Agent
Runliang Niu (Jilin University), Qi Wang (Jilin University)
TransformerSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelMultimodality
🎯 What it does: Built a computer control environment and agent based on a vision-language model (VLM), enabling mouse and keyboard operations on real computer screens, and trained the ScreenAgent model.
ScreenAI: A Vision-Language Model for UI and Infographics Understanding
Gilles Baechler (Google DeepMind), Abhanshu Sharma (Google DeepMind)
TransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: Propose ScreenAI, a vision-language model for UI and infographics, achieving cross-domain understanding through self-supervised pre-training and data generated by large-scale LLMs.
SCTrans: Multi-scale scRNA-seq Sub-vector Completion Transformer for Gene-selective Cell Type Annotation
Lu Lin (South China University of Technology), Hau San Wong (City University of Hong Kong)
ClassificationExplainability and InterpretabilityRepresentation LearningTransformerContrastive LearningBiomedical DataBenchmark
🎯 What it does: Proposed a multi-scale scRNA-seq subvector completion Transformer (SCTrans) to achieve cell type annotation based on gene subsets, and implemented interpretable gene subset selection through an attention mechanism.
SDformer: Transformer with Spectral Filter and Dynamic Attention for Multivariate Time Series Long-term Forecasting
Ziyu Zhou (Beijing University of Technology), Zhen Yang (Beijing University of Technology)
TransformerTime Series
🎯 What it does: Propose the SDformer architecture, addressing the issue of smooth attention distribution in long-range forecasting of multivariate time series through Spectral-Filter-Transform (SFT) and Dynamic-Directional-Attention (DDA).
Searching for Programmatic Policies in Semantic Spaces
Rubens O. Moraes (Universidade Federal de Vicosa), Levi H. S. Lelis (University of Alberta)
AI Code AssistantReinforcement LearningBenchmark
🎯 What it does: Proposed a library-induced semantic search space (LISS), improving programmatic strategy search by constructing neighborhoods with semantically diverse programs;
Seed Selection in the Heterogeneous Moran Process
Petros Petsinis (Aarhus University), Panagiotis Karras (University of Copenhagen)
OptimizationGraph
🎯 What it does: Studied the seed selection problem in the heterogeneous Moran process, i.e., selecting k nodes given a budget k to maximize a fixed probability.
SeeDRec: Sememe-based Diffusion for Sequential Recommendation
Haokai Ma (Shandong University), Zhanhui Kang (Tencent)
Recommendation SystemPrompt EngineeringDiffusion modelTextSequential
🎯 What it does: Proposed SeeDRec, a diffusion framework based on recommendation semantic units (sememe), to enhance the performance of sequential recommendations.
Selecting the Most Conflicting Pair of Candidates
Théo Delemazure (CNRS, LAMSADE, Université Paris Dauphine - PSL), Stanisław Szufa (CNRS, LAMSADE, Université Paris Dauphine - PSL)
TextTabular
🎯 What it does: Studied how to identify and select the most conflicting candidate pairs in multi-choice voting rules, proposed new axioms such as conflict consistency and reverse stability, and designed a conflict voting rule based on these axioms.
Self-adaptive Extreme Penalized Loss for Imbalanced Time Series Prediction
Yiyang Wang (Dalian Maritime University), Yuhan Guo (Dalian Maritime University)
Anomaly DetectionRecurrent Neural NetworkTime Series
🎯 What it does: This paper proposes an adaptive extreme penalty loss function (EPL), combining it with Empirical Wavelet Transform decomposition and an LSTM with attention mechanism to form a framework capable of accurately predicting both extreme and normal events in imbalanced time series.
Self-adaptive PSRO: Towards an Automatic Population-based Game Solver
Pengdeng Li (Nanyang Technological University), Bo An (Nanyang Technological University)
Hyperparameter SearchTransformerReinforcement LearningGraphTabular
🎯 What it does: Propose an adaptive PSRO framework that enables PSRO to automatically adjust hyperparameters to solve two-player zero-sum games
Self-Promoted Clustering-based Contrastive Learning for Brain Networks Pretraining
Junbo Ma (Hangzhou Dianzi University), Kai Zhao (Chinese PLA General Hospital)
Representation LearningRecurrent Neural NetworkGraph Neural NetworkContrastive LearningGraphBiomedical Data
🎯 What it does: Propose a self-promoting clustering contrastive learning framework (SPCCL) without requiring graph structure perturbation for brain network pre-training;
Self-Supervised Learning for Enhancing Spatial Awareness in Free-Hand Sketches
Xin Wang (Shanghai Jiao Tong University), Lei Xu (Shanghai Jiao Tong University)
RestorationRecurrent Neural NetworkGraph Neural NetworkAuto EncoderImage
🎯 What it does: Propose a self-supervised task called 'Sketch Reorganization' to correct stroke position errors in hand-drawn sketches, and design the SketchGloc model to utilize multi-graph structures to learn spatial layout information, enhancing the robustness of sketch representations.
Self-Supervised Monocular Depth Estimation in the Dark: Towards Data Distribution Compensation
Haolin Yang (East China University of Science and Technology), Yang Tang (East China University of Science and Technology)
Depth EstimationAutonomous DrivingConvolutional Neural NetworkTransformerImage
🎯 What it does: Propose a self-supervised nighttime monocular depth estimation framework that uses only daytime images during training, compensating for day-night distribution differences through physical priors without requiring nighttime images.
Self-Supervised Pre-training with Symmetric Superimposition Modeling for Scene Text Recognition
Zuan Gao (University of Science and Technology of China), Hongtao Xie (University of Science and Technology of China)
RecognitionRepresentation LearningTransformerPrompt EngineeringAuto EncoderContrastive LearningImage
🎯 What it does: Propose a self-supervised pre-training framework called Symmetric Superimposition Modeling (SSM), which constructs symmetric superimposed inputs by overlapping the original text image with its horizontally/vertically flipped or 180° rotated inverted image, and learns character shape and latent linguistic information through reconstruction at both pixel and feature levels.
Self-supervised Weighted Information Bottleneck for Multi-view Clustering
Zhengzheng Lou (Zhengzhou University), Shizhe Hu (Zhengzhou University)
OptimizationRepresentation LearningImageVideoTextMultimodality
🎯 What it does: Proposed a self-supervised weighted information bottleneck (SWIB) method for multi-view clustering.
SemanticMask: A Contrastive View Design for Anomaly Detection in Tabular Data
Shuting Tao (Zhejiang University), Xiangming Meng (Zhejiang University)
Anomaly DetectionLarge Language ModelContrastive LearningTabular
🎯 What it does: Proposes a contrastive learning data augmentation method called SemanticMask based on semantic information for unsupervised anomaly detection in tabular data.
Semantics for Non-Flat Assumption-Based Argumentation, Revisited
Jesse Heyninck (Open Universiteit), Ofer Arieli (Tel-Aviv Academic College)
🎯 What it does: This paper proposes a six-valued (eight-valued) marked semantics to address the issues of unavailability and incomplete extensions in traditional three-valued semantics when assumptions exist in rule heads within the non-flat assumption reasoning framework (ABA). It achieves finer-grained argument analysis through refined inference and acceptance distinctions.
SEMANTIFY: Unveiling Memes with Robust Interpretability beyond Input Attribution
Dibyanayan Bandyopadhyay (Indian Institute of Technology Patna), Asif Ekbal (Indian Institute of Technology Jodhpur)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Propose the SEMANTIFY framework, which employs a multi-step filtering mechanism combining GPT-2 and CLIP to generate interpretable keywords for determining the offensiveness of online memes.
SEMv3: A Fast and Robust Approach to Table Separation Line Detection
Chunxia Qin (University of Science and Technology of China), Jun Du (University of Science and Technology of China)
RecognitionSegmentationConvolutional Neural NetworkTransformerTabular
🎯 What it does: This paper proposes the SEMv3 framework, achieving fast and robust table structure recognition under the split-merge paradigm.
Sentence-Level or Token-Level? A Comprehensive Study on Knowledge Distillation
Jingxuan Wei (Shenyang Institute of Computing Technology Chinese Academy of Sciences), Ruifeng Guo (Shenyang Institute of Computing Technology Chinese Academy of Sciences)
Knowledge DistillationTransformerLarge Language ModelText
🎯 What it does: This paper systematically compares the performance of sentence-level and word-level knowledge distillation in neural machine translation, and proposes a dynamic gated hybrid distillation method.
Separate in the Speech Chain: Cross-Modal Conditional Audio-Visual Target Speech Extraction
Zhaoxi Mu (Xi'an Jiaotong University), Xinyu Yang (Xi'an Jiaotong University)
RestorationTransformerContrastive LearningMultimodality
🎯 What it does: Propose AVSepChain to achieve audio-visual target speech extraction through a two-stage speech chain simulation of perception and generation, introducing a contrastive semantic matching loss.
SGDCL: Semantic-Guided Dynamic Correlation Learning for Explainable Autonomous Driving
Chengtai Cao (City University of Hong Kong), Yung-Hui Li (Foxconn Research)
Autonomous DrivingExplainability and InterpretabilityConvolutional Neural NetworkGraph Neural NetworkVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a model called SGDCL, which combines semantic-guided category feature learning and dynamic association learning to achieve interpretable autonomous driving behavior and explanation prediction.
Shadow-Free Membership Inference Attacks: Recommender Systems Are More Vulnerable Than You Thought
Xiaoxiao Chi (Macquarie University), Hongsheng Hu (CSIRO Data61)
Recommendation SystemAdversarial AttackTabularBenchmark
🎯 What it does: Propose a shadow-free membership inference attack that directly utilizes personalized recommendations from a recommendation system based on users' historical interactions to determine membership status.
Shap-Mix: Shapley Value Guided Mixing for Long-Tailed Skeleton Based Action Recognition
Jiahang Zhang (Peking University), Jiaying Liu (Peking University)
RecognitionExplainability and InterpretabilityData-Centric LearningGraph
🎯 What it does: Propose a skeleton action hybrid enhancement method guided by Shapley values (Shap-Mix), aiming to address the performance degradation of skeleton action recognition under long-tailed distributions.
Simple Contrastive Multi-View Clustering with Data-Level Fusion
Caixuan Luo (Guangxi Normal University), Xiaofeng Zhu (Guangxi Normal University)
Representation LearningContrastive LearningMultimodality
🎯 What it does: Proposed a simple contrastive multi-view clustering framework SCM based on data-level fusion;
SketchEdit: Editing Freehand Sketches at the Stroke-Level
Tengjie Li (Shanghai Jiao Tong University), Lei Xu (Shanghai Jiao Tong University)
GenerationTransformerDiffusion modelAuto EncoderImage
🎯 What it does: Proposes SketchEdit, a stroke-level editing framework for freehand sketches based on diffusion models, which can decouple strokes and generate edited results in reasonable positions.
Skip-Timeformer: Skip-Time Interaction Transformer for Long Sequence Time-Series Forecasting
Wenchang Zhang (Shandong Technology and Business University), Fan Zhang (Shandong Technology and Business University)
TransformerTime Series
🎯 What it does: Propose Skip-Timeformer, which improves the performance of Transformer in long sequence time series forecasting by utilizing multi-hop subsequence embeddings and skip-time interaction mechanisms.
Solving Long-run Average Reward Robust MDPs via Stochastic Games
Krishnendu Chatterjee (Institute of Science and Technology Austria), Đorđe Žikelić (Singapore Management University)
OptimizationComputational EfficiencyReinforcement LearningBenchmark
🎯 What it does: This paper proposes a method to convert the long-term average reward robust MDP (RMDP) with polynomial uncertainty sets into a turn-based stochastic game (TBSG) with finite states and actions, and designs a robust polynomial policy iteration (RPPI) algorithm based on this, which can solve long-term average reward RMDPs without being constrained by single-chain or periodic conditions;
Solving Quantified Boolean Formulas with Few Existential Variables
Leif Eriksson (Linkoping University), Mateusz Rychlicki (University of Leeds)
OptimizationComputational Efficiency
🎯 What it does: This paper studies the fixed-parameter tractability (FPT) problem in quantified Boolean formulas (QBF) parameterized by the number of existential variables, presents an FPT algorithm under clause length restrictions, and proves that the problem is W[1]-hard when clause length is unbounded; further, under the Strong Exponential Time Hypothesis (SETH), it provides an exponential lower bound for the case of 3-clauses.
Span-based Unified Named Entity Recognition Framework via Contrastive Learning
Hongli Mao (Beijing Institute of Technology), Heyan Huang (Beijing Institute of Technology)
RecognitionTransformerContrastive LearningText
🎯 What it does: Proposed a span-based unified named entity recognition framework called SUNER, which maps entity spans and natural language type descriptions into the same semantic space using contrastive learning, achieving parallel entity extraction.
Sparse Multi-Relational Graph Convolutional Network for Multi-type Object Trajectory Prediction
Jianhui Zhang (Hangzhou Dianzi University), Zheng Wang (Wuhan University)
Object TrackingAutonomous DrivingGraph Neural NetworkTime SeriesSequential
🎯 What it does: Propose a multi-type object trajectory prediction framework SMGCN, combining sparse multi-relational graph convolutional networks with multi-round global temporal aggregation;
Spatial-Temporal Perceiving: Deciphering User Hierarchical Intent in Session-Based Recommendation
Xiao Wang (University of Electronic Science and Technology of China), Shuang Liang (University of Electronic Science and Technology of China)
Recommendation SystemGraph Neural NetworkContrastive LearningSequential
🎯 What it does: This paper proposes a hierarchical intent-based conversational recommendation framework called HearInt, aiming to simultaneously decouple and model user behavior from both the temporal aspect (long-term and short-term intents) and the spatial aspect (small-scale and large-scale intents).
Spatial-Temporal-Decoupled Masked Pre-training for Spatiotemporal Forecasting
Haotian Gao (University of Tokyo), Xuan Song (Southern University of Science and Technology)
Representation LearningTransformerAuto EncoderTime Series
🎯 What it does: Designed and implemented a space-time decoupled mask self-supervised pre-training framework called STD-MAE to improve the accuracy and robustness of spatiotemporal prediction models on long sequences.
Spear: Evaluate the Adversarial Robustness of Compressed Neural Models
Chong Yu (Fudan University), Jiayuan Fan (Fudan University)
ClassificationObject DetectionCompressionKnowledge DistillationAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: Designed and implemented a specialized adversarial attack method called Spear for sparse and quantized compressed models to evaluate the adversarial robustness of compressed models.
SpecAR-Net: Spectrogram Analysis and Representation Network for Time Series
Yi Dong (Intelligent Science & Technology Academy of CASIC), Zhe Ma (Intelligent Science & Technology Academy of CASIC)
ClassificationAnomaly DetectionConvolutional Neural NetworkRecurrent Neural NetworkTransformerTime Series
🎯 What it does: Propose SpecAR-Net, which utilizes time-frequency joint analysis and representation networks to extract global trends, periodicity, and mutation features from time series, achieving a unified multi-task model.
Speech-Forensics: Towards Comprehensive Synthetic Speech Dataset Establishment and Analysis
Zhoulin Ji (Xi'an Jiaotong University), Chao Shen (Xi'an Jiaotong University)
RecognitionAnomaly DetectionRecurrent Neural NetworkTransformerBenchmarkAudio
🎯 What it does: Propose the Speech-Forensics dataset and design the TEST model to achieve authenticity determination of synthetic speech, multiple forged segment localization, and synthesis algorithm identification;
SPGNet: A Shape-prior Guided Network for Medical Image Segmentation
Zhengxuan Song (Sun Yat-sen University), Kun Zeng (Sun Yat-sen University)
SegmentationConvolutional Neural NetworkTransformerImageBiomedical DataUltrasound
🎯 What it does: Developed SPGNet, a dual-path collaborative network integrating statistical shape models to guide medical image segmentation.
STAR: Spatio-Temporal State Compression for Multi-Agent Tasks with Rich Observations
Chao Li (Nanjing University), Yang Gao (Nanjing University)
Representation LearningReinforcement Learning
🎯 What it does: Learning compressed state representations in multi-agent tasks to improve decision-making efficiency based on scarce information
Stochastic Neural Simulator for Generalizing Dynamical Systems across Environments
Liu Jiaqi (Jilin University), Bo Yang (Jilin University)
Domain AdaptationTime SeriesPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes a stochastic neural ODE process model called CoNDP for learning and predicting dynamic systems from sparse observations under different environments, capable of capturing environmental commonalities and differences while rapidly adapting to new environments.
Strengthening Layer Interaction via Dynamic Layer Attention
Kaishen Wang (Sichuan University), Tao He (Sichuan University)
ClassificationObject DetectionConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: Investigate the static limitations of hierarchical attention, propose a dynamic hierarchical attention (DLA) framework, and design a DSU plugin;
Structure-Aware Spatial-Temporal Interaction Network for Video Shadow Detection
Housheng Wei (Sichuan University), Yanli Liu (Sichuan University)
SegmentationConvolutional Neural NetworkTransformerVideo
🎯 What it does: Propose a video shadow detection network SSTINet based on Transformer and lightweight CNN, achieving precise and consistent shadow detection in consecutive video frames through spatial-temporal feature interaction and structure-aware prediction module.
Structure-Preserving Physics-Informed Neural Networks with Energy or Lyapunov Structure
Haoyu Chu (Beijing Jiaotong University), Daisuke Furihata (Osaka University)
ClassificationImagePhysics Related
🎯 What it does: Propose structure-preserving physics-informed neural networks (SP-PINN), enhancing PDE solving and image classification performance by incorporating energy or Lyapunov structure loss terms.
Structured d-DNNF Is Not Closed under Negation
Harry Vinall-Smeeth (Technische Universitat Ilmenau)
🎯 What it does: Prove that structured d-DNNF is not closed under negation, disjunction, and existential quantification, and use this result to demonstrate the succinctness gap between structured d-DNNF and SDD as well as positive arithmetic circuits (PSDD).
Sub-Adjacent Transformer: Improving Time Series Anomaly Detection with Reconstruction Error from Sub-Adjacent Neighborhoods
Wenzhen Yue (Peking University), Taiyan Chen (Peking University)
Anomaly DetectionTransformerAuto EncoderTime Series
🎯 What it does: Proposes Sub-Adjacent Transformer, an unsupervised time series anomaly detection framework that combines sub-adjacent neighborhood attention with reconstruction error.
Subgraph Pooling: Tackling Negative Transfer on Graphs
Zehong Wang (University of Notre Dame), Yanfang Ye (University of Notre Dame)
Domain AdaptationGraph Neural NetworkSupervised Fine-TuningGraph
🎯 What it does: This paper proposes Subgraph Pooling (SP) and its improved version SP++, which reduces the distribution discrepancy between source and target graphs by aggregating subgraph information around nodes, thereby alleviating negative transfer caused by graph structures.