IJCAI 2025 Papers — Page 8
International Joint Conference on Artificial Intelligence · 1014 papers
On the Power of Optimism in Constrained Online Convex Optimization
Haobo Zhang (ShanghaiTech University), Xin Liu (ShanghaiTech University)
OptimizationTime Series
🎯 What it does: This paper proposes an adaptive, environment-agnostic gradient optimization algorithm called Optimistic-COCO for solving constrained online convex optimization (COCO) problems, achieving regret related to gradient variations and constant-level constraint violations without prior knowledge of the time horizon or gradient changes.
One-step Label Shift Adaptation via Robust Weight Estimation
Ruidong Fan, Chenping Hou (National University Of Defense Technology)
Domain AdaptationConvolutional Neural NetworkImage
🎯 What it does: Propose a one-stage label distribution drift adaptation method (OLSA), which simultaneously learns the target classifier and importance weights through bi-level optimization, achieving adaptive correction for label distribution shift.
Online 3D Gaussian Splatting Modeling with Novel View Selection
Byeonggwon Lee (Dongguk University), Soohwan Song (Dongguk University)
Depth EstimationOptimizationTransformerGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: This paper proposes an online 3D Gaussian Splatting (3DGS) modeling method for RGB-only image streams, introducing a novel view selection mechanism (Novel View Selection, NVS) based on uncertainty. It utilizes non-keyframes for incremental training, significantly enhancing model completeness and rendering quality.
Online Housing Market
Julien Lesca (Universit e Paris-Dauphine)
OptimizationFinance Related
🎯 What it does: Proposed and analyzed the resource allocation problem in online housing markets (where agents have different arrival and departure times), designing multiple online mechanisms based on sequential dictatorship and vertex trading cycles (TTC).
Online Planning in MDPs with Stochastic Durative Actions
Tal Berman, Erez Karpas (Technion)
Reinforcement Learning
🎯 What it does: Proposed an online planning algorithm called TP-MCTS that can handle continuous actions with random effects while meeting constraints such as deadlines and time windows, and supports concurrent execution.
Online Resource Sharing: Better Robust Guarantees via Randomized Strategies
David X. Lin (Cornell University), Éva Tardos (Cornell University)
OptimizationFinance Related
🎯 What it does: Propose a method to achieve higher robust resource allocation guarantees through randomized bidding in a repeated first-price auction mechanism without money, and prove that this method can ensure approximately 0.59 ideal utility under any opponent's strategy.
Open-Vocabulary Fine-Grained Hand Action Detection
Ting Zhe (Wuhan University), Jing Zhang (Wuhan University)
Object DetectionTransformerVision Language ModelMultimodality
🎯 What it does: Propose a new baseline for open-vocabulary fine-grained hand action detection called Open-FGHA, and design three modules (HiH-LoRA, BSF, CQG) to address recognition bias caused by data scarcity and class similarity.
Open-World Semi-Supervised Learning with Class Semantic Correlations
Yuxin Fan (Shanxi University), Jianqing Liang (Shanxi University)
ClassificationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Developed an open-world semi-supervised learning method called CSC-OWSSL, which enhances image representations using a frozen text encoder and semantic margin contrastive loss, while simultaneously identifying known and unknown classes.
Optical Flow Estimation for Tiny Objects: New Problem, Specialized Benchmark, and Bioinspired Scheme
Xueyao Ji, Yizheng Wang (Beijing Institute of Basic Medical Sciences)
Convolutional Neural NetworkOptical FlowVideoBenchmark
🎯 What it does: Proposed a lightweight bio-inspired optical flow estimation method named OTHR for accurately capturing the motion of extremely small objects (<100 pixels) and constructed a specialized optical flow dataset FlyingTO along with refined evaluation metrics tailored for small objects.
Optimal Capacity Modification for Stable Matchings with Ties
Keshav Ranjan (IIT Madras), Prajakta Nimbhorkar (Chennai Mathematical Institute)
Optimization
🎯 What it does: Studied the feasibility of ensuring strongly stable matching by increasing hospital quotas in the hospital/resident (HR-HT) problem, and provided algorithms and complexity results for minimizing total quota increments (MINSUM) and minimizing the maximum individual hospital increment (MINMAX).
Optimal Distributed Training With Co-Adaptive Data Parallelism in Heterogeneous Environments
Lifang Chen (Hangzhou Dianzi University), Pan Li (Hangzhou Dianzi University)
OptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: Developed the C-ADP framework to optimize distributed deep learning training in heterogeneous environments through adaptive data allocation and delayed parameter synchronization.
Optimal Metric Distortion for Matching on the Line
Aris Filos-Ratsikas (University of Edinburgh), Alexandros A. Voudouris (University of Essex)
Optimization
🎯 What it does: This paper studies the upper bounds of distortion achievable using ordinal information (i.e., preference rankings) for one-sided matching (agent–item) and two-sided matching (agent–agent) problems in linear metric spaces. It proposes a ranking-based algorithm and proves that it achieves the optimal distortion of 3 for one-sided matching (for all k-centrum costs) and the optimal distortion of 1 for two-sided matching.
Optimal Planning to Coordinate Science Data Collection and Downlink for a Constellation of Agile Satellites with Limited Storage
Richard Levinson (NASA Ames Research Center), Sreeja Roy-Singh (NASA Ames Research Center)
OptimizationTabularTime Series
🎯 What it does: Develop an optimal data acquisition and downlink planning for a group of maneuverable satellites to maximize scientific rewards.
Optimal Policy Adaptation Under Covariate Shift
Xueqing Liu, Peng Wu (ByteDance Research)
Domain AdaptationReinforcement LearningTabular
🎯 What it does: This paper proposes a method to construct a semi-parametric efficient estimator and learn the optimal intervention strategy under covariate shift conditions, utilizing complete information from the source domain and only covariate data from the target domain.
Optimal Transport on Categorical Data for Conterfactuals Using Compositional Data and Dirichlet Transport
Agathe Fernandes Machado (Université du Québec à Montréal), Arthur Charpentier (Université du Québec à Montréal)
Data SynthesisExplainability and InterpretabilityTabular
🎯 What it does: Convert discrete categorical features into compositional (probability) vectors and perform optimal transport on the probability simplex to generate interpretable counterfactual predictions.
Optimized View and Geometry Distillation from Multi-view Diffuser
Youjia Zhang (Huazhong University of Science and Technology), Wei Yang (Huazhong University of Science and Technology)
GenerationData SynthesisKnowledge DistillationSupervised Fine-TuningDiffusion modelScore-based ModelNeural Radiance FieldImageMesh
🎯 What it does: Extract and optimize 3D geometry and views from a single input image through a multi-view diffusion model (Zero-1-to-3), ultimately generating a 3D model that is consistent across multiple views and rich in details.
Optimizing Parameters of Quantum Circuits with Sparsity-Inducing Coordinate Descent
Rudy Raymond (JPMorganChase & Co.), Zichang He (JPMorganChase & Co.)
OptimizationGraphPhysics Related
🎯 What it does: Proposed a coordinate descent algorithm called Rotolasso based on L1 regularization to optimize parameters of parameterized quantum circuits (PQC), balancing accuracy and parameter count.
Optimizing Personalized Federated Learning Through Adaptive Layer-Wise Learning
Weihang Chen (Shaanxi Normal University), Zheng Wang (University of Leeds)
OptimizationFederated LearningConvolutional Neural NetworkImageText
🎯 What it does: Proposed FLAYER, a method in federated learning that achieves personalized model training through hierarchical aggregation, hierarchical adaptive learning rates, and hierarchical sparse masks.
OS-GCL: A One-Shot Learner in Graph Contrastive Learning
Cheng Ji (Beihang University), Jianxin Li (Beihang University)
Representation LearningMeta LearningContrastive LearningGraph
🎯 What it does: Proposed the OS-GCL framework, a graph contrastive learning method based on probability distribution estimation, to address the issue of insufficient self-supervised signals in traditional GCL under one-shot learning scenarios.
OT-DETECTOR: Delving into Optimal Transport for Zero-shot Out-of-Distribution Detection
Yu Liu (Nanjing University of Science and Technology), Zechao Li (Nanjing University of Science and Technology)
Domain AdaptationAnomaly DetectionImage
🎯 What it does: Propose the OT-DETECTOR framework to achieve out-of-distribution (OOD) detection under zero-shot scenarios.
Outstanding Orthodontist: No More Artifactual Teeth in Talking Face
Zibo Su (Xidian University), Cheng Deng (Xidian University)
GenerationData SynthesisConvolutional Neural NetworkRecurrent Neural NetworkGenerative Adversarial NetworkVideoAudio
🎯 What it does: This paper proposes a full-process framework called OrthoNet for audio-driven talking face synthesis, specifically addressing the issues of temporal inconsistency and artifacts in teeth during speech.
PALA: Class-imbalanced Graph Domain Adaptation via Prototype-anchored Learning and Alignment
Xin Ma (Sichuan University), Jiancheng Lv (Sichuan University)
Domain AdaptationAdversarial AttackGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Propose a prototype-anchored graph domain adaptation framework called PALA to address knowledge transfer bias caused by class imbalance in the source graph.
PAMol: Pocket-Aware Drug Design Method with Hypergraph Representation of Protein Pocket Structure and Feature Fusion
Xiaoli Lin (Wuhan University of Science and Technology), Xiaolong Zhang (Wuhan University of Science and Technology)
Drug DiscoveryGraph Neural NetworkTransformerGenerative Adversarial NetworkGraphBiomedical Data
🎯 What it does: Propose a drug design framework called PAMol based on protein pocket hypergraph representation and multimodal feature fusion, which can generate molecules with high affinity and favorable pharmacological properties under target protein pocket conditions.
PanComplex: Leveraging Complex-Valued Neural Networks for Enhanced Pansharpening
Chunhui Luo (University of Science and Technology of China), Xueyang Fu (University of Science and Technology of China)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: A dual-domain (spatial-frequency) framework called PanComplex based on complex convolution for fusing multispectral and panchromatic images, outputting high-resolution multispectral images.
Parallel Belief Contraction via Order Aggregation
Jake Chandler (La Trobe University), Richard Booth (Cardiff University)
🎯 What it does: This paper proposes a parallel belief contraction method achieved through sequential aggregation (TeamQueue aggregation), extending traditional single-item contraction to parallel and iterative parallel contraction;
Parallel Belief Revision via Order Aggregation
Jake Chandler (La Trobe University), Richard Booth (Cardiff University)
🎯 What it does: Propose an iterative parallel belief revision framework based on TeamQueue aggregation;
Parameterized Approximation Algorithm for Doubly Constrained Fair Clustering
Xiaoliang Wu (Central South University), Jianxin Wang (Central South University)
Optimization
🎯 What it does: Study the k-median clustering problem with dual fairness constraints and propose an FPT(k)-time approximation algorithm.
Partial Label Clustering
Yutong Xie (Southeast University), Yuheng Jia (Southeast University)
OptimizationRepresentation LearningImageTabular
🎯 What it does: Propose a PLC method that utilizes partial label information for clustering. First, construct a neighborhood weight matrix and perform label disambiguation, then generate must-link/cannot-link constraints and optimize them through dual propagation, ultimately obtaining clustering results in spectral clustering.
Partially Observable Reference Policy Programming
Edward Kim (Australian National University), Hanna Kurniawati (Australian National University)
Reinforcement LearningMesh
🎯 What it does: Proposed the PORPP algorithm, an online approximate POMDP solver capable of deeply sampling future history and incrementally updating policies.
Participatory Budgeting Project Strength via Candidate Control
Piotr Faliszewski (AGH University of Krakow), Krzysztof Sornat (AGH University of Krakow)
OptimizationTabular
🎯 What it does: This paper studies the candidate control problem in participatory budget elections, where candidates can manipulate the outcome by deleting or adding projects, and provides theoretical analysis and experimental validation of its computational complexity.
PatternCIR Benchmark and TisCIR: Advancing Zero-Shot Composed Image Retrieval in Remote Sensing
Zhechun Liang (Xidian University), Guangming Shi (Xidian University)
RetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose a zero-shot remote sensing composite image retrieval method, construct the PatternCIR dataset, and design two core modules: Zero-Shot Query Text Generator (ZS-QTG) and Text-Image Sequential Training of Composed Image Retrieval (TisCIR);
PDDFormer: Pairwise Distance Distribution Graph Transformer for Crystal Material Property Prediction
Xiangxiang Shen (East China Normal University), Xian Wei (East China Normal University)
Graph Neural NetworkTransformerGraphPhysics Related
🎯 What it does: Constructed a multi-edge crystal graph based on atomic weighting and unit cell distance distribution, and proposed PDDFormer (WPDDFormer and UPDDFormer) to predict crystal material properties.
Performance Guaranteed Poisoning Attacks in Federated Learning: A Sliding Mode Approach
Huazi Pan (Deakin University), Suiyang Khoo (Deakin University)
Federated LearningAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Proposed FedSA, a controllable backdoor attack based on sliding mode control, which can precisely control the global model's performance in federated learning with predetermined accuracy.
PerfSeer: An Efficient and Accurate Deep Learning Models Performance Predictor
Xinlong Zhao (Shandong Normal University), Ke Liu (Shandong Normal University)
Computational EfficiencyGraph Neural NetworkGraph
🎯 What it does: This paper proposes PerfSeer, a performance predictor capable of efficiently and accurately forecasting the execution time, memory usage, and SM utilization of deep learning models during both training and inference phases.
PeSANet: Physics-encoded Spectral Attention Network for Simulating PDE-Governed Complex Systems
Han Wan (Renmin University of China), Hao Sun (Renmin University of China)
Convolutional Neural NetworkTime SeriesPhysics Related
🎯 What it does: Proposed a novel Physical Encoding Spectral Attention Network (PeSANet), achieving long-term prediction for two-dimensional complex systems governed by partial differential equations (PDEs) by combining hard-constrained local differential operator learning with frequency-domain spectral attention mechanisms.
Phenotypic Profile-Informed Generation of Drug-Like Molecules via Dual-Channel Variational Autoencoders
Hui Liu (Nanjing Tech University), Xuejun Liu (Nanjing Tech University)
Drug DiscoveryAuto EncoderTextBiomedical Data
🎯 What it does: Propose SmilesGEN, a dual-channel VAE that integrates drug SMILES and gene expression profiles, explicitly modeling drug-induced perturbations to cellular phenotypes in the latent space, thereby generating drug molecules with potential efficacy under the condition of a desired expression profile.
Physical Adversarial Camouflage Through Gradient Calibration and Regularization
Jiawei Liang (Sun Yat-sen University), Xiaochun Cao (Sun Yat-sen University)
Object DetectionAutonomous DrivingOptimizationAdversarial AttackImagePhysics Related
🎯 What it does: This paper proposes a physics-based adversarial camouflage framework based on gradient optimization, which optimizes vehicle surface textures to mislead multi-angle and multi-distance object detectors by utilizing gradient calibration and gradient decorrelation techniques.
Physics-Assisted and Topology-Informed Deep Learning for Weather Prediction
Jiaqi Zheng (Sun Yat-Sen University), Yerong Feng (Shenzhen Institute of Meteorological Innovation)
Graph Neural NetworkTime SeriesPhysics Related
🎯 What it does: Developed a weather prediction model called PASSAT, which integrates physical equations with Earth's surface topology, enabling numerical solutions of transport and Navier-Stokes equations on a sphere, and employs spherical graph neural networks to estimate initial velocity fields and interactions between the atmosphere and the surface.
Picturized and Recited with Dialects: A Multimodal Chinese Representation Framework for Sentiment Analysis of Classical Chinese Poetry
Xiaocong Du (ShanghaiTech University), Haipeng Zhang (ShanghaiTech University)
ClassificationRepresentation LearningConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityAudio
🎯 What it does: This paper proposes a tri-modal Chinese representation framework that uses audio, visual, and textual modalities together to perform emotion analysis on classical Chinese poetry;
Pixel-wise Divide and Conquer for Federated Vessel Segmentation
Tian Chen (Wuhan University), Yongchao Xu (Wuhan University)
SegmentationFederated LearningConvolutional Neural NetworkBiomedical Data
🎯 What it does: Propose a pixel-level 'divide and conquer' method for federated vessel segmentation (Federated Vessel-Aware Calibration, FVAC), which guides local models to focus on hard-to-discriminate pixels through global uncertainty, and separates and aligns foreground and background features to improve segmentation accuracy.
PNAct: Crafting Backdoor Attacks in Safe Reinforcement Learning
Weiran Guo (Tongji University), Ling Wang (Tongji University)
Adversarial AttackReinforcement LearningBenchmark
🎯 What it does: Proposed the PNAct framework, which achieves precise control over safety constraints in Safe Reinforcement Learning (Safe RL) through the design of positive and negative action samples and a loss function, enabling agents to execute dangerous actions under triggering conditions without affecting rewards.
Point Cloud Mixture-of-Domain-Experts Model for 3D Self-supervised Learning
Yaohua Zha (Tsinghua University), Shu-Tao Xia (Shenzhen University)
ClassificationObject DetectionSegmentationRepresentation LearningTransformerMixture of ExpertsPoint Cloud
🎯 What it does: Propose a hybrid domain expert model called Point-MoDE and its block-to-scene pre-training strategy, integrating knowledge from the scene domain and object domain to learn comprehensive 3D representations.
Poisoning-based Backdoor Attacks for Arbitrary Target Label with Positive Triggers
Binxiao Huang (University of Hong Kong), Ngai Wong (University of Hong Kong)
Adversarial AttackImage
🎯 What it does: Proposed a poisoning-based backdoor attack that uses a positive trigger to implant multi-label, multi-payload backdoors into training data, enabling the model to steer any input to any target class during inference.
POLO: An LLM-Powered Project-Level Code Performance Optimization Framework
Jiameng Bai (Zhejiang University), Gang Chen (Zhejiang University)
OptimizationAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: Propose the POLO framework, which identifies hotspots through dynamic analysis, constructs a program structure graph via static structural analysis, and achieves project-level performance optimization by iteratively rewriting code using dual-agent LLMs.
Polynomial-Time Relational Probabilistic Inference in Open Universes
Luise Ge (Washington University in St. Louis), Kris Nilsson (Washington University in St. Louis)
Computational Efficiency
🎯 What it does: Proposed a first-order relational probabilistic reasoning framework that operates in polynomial time within an open world (unbounded object set), achieving inference under bounded degree constraints using the sum-of-squares (SOS) method.
POMP: Pathology-omics Multimodal Pre-training Framework for Cancer Survival Prediction
Suixue Wang (Hainan University), Qingchen Zhang (Hainan University)
Representation LearningTransformerContrastive LearningImageMultimodalityTabular
🎯 What it does: Proposed the POMP framework, which pre-trains and fuses pathological images with multi-omics data for cancer survival prediction.
Pre-defined Keypoints Promote Category-level Articulation Pose Estimation via Multi-Modal Alignment
Wenbo Xu (Hefei University of Technology), Rujing Wang (Chinese Academy of Sciences)
Pose EstimationGraph Neural NetworkSimultaneous Localization and MappingMultimodality
🎯 What it does: Proposed the PAGE framework, which achieves category-level 6D pose estimation of articulated objects by leveraging predefined keypoints, multimodal geometric-color alignment, and voting integration methods.
Predicting Spectral Information for Self-Supervised Signal Classification
Yi Xu (Xidian University), Luyang Mei (Xidian University)
ClassificationTransformerSupervised Fine-TuningAuto EncoderTime Series
🎯 What it does: Developed a self-supervised signal classification method called SGSSC, which leverages spectral information to predict the frequency domain spectrum of time-domain signals, thereby enhancing the effectiveness of communication signal classification in low-label environments.
Preference Elicitation for Multi-objective Combinatorial Optimization with Active Learning and Maximum Likelihood Estimation
Marianne Defresne (KU Leuven), Tias Guns (KU Leuven)
Optimization
🎯 What it does: This paper proposes a constructive preference inference method based on active learning and maximum likelihood estimation for multi-objective combinatorial optimization problems.
Preference Identification by Interaction Overlap for Bundle Recommendation
Fei-Yao Liang (Sun Yat-sen University), Hui-Yu Zhou (Guangxi Zhuang Autonomous Region Information Center)
Recommendation SystemAuto Encoder
🎯 What it does: Proposed the PIIO model, which improves the effectiveness of bundled product recommendations through two modules: data augmentation and preference aggregation.
Preference-based Deep Reinforcement Learning for Historical Route Estimation
Boshen Pan (Dalian University of Technology), Qiang Zhang (Dalian University of Technology)
Autonomous DrivingOptimizationData-Centric LearningReinforcement Learning from Human FeedbackReinforcement LearningSequentialBenchmark
🎯 What it does: Studied vehicle route planning based on preference-based deep reinforcement learning, leveraging historical routes to learn driver preferences and generate routes more aligned with human preferences.
Preventing Latent Diffusion Model-Based Image Mimicry via Angle Shifting and Ensemble Learning
Minghao Li (Chinese Academy of Sciences), Lihua Jing (University of Chinese Academy of Sciences)
Adversarial AttackDiffusion modelAuto EncoderImage
🎯 What it does: Perform untargeted adversarial attacks on LDMs by proposing an alternating iterative framework, angle offset (cosine similarity) attack, gradient integration, and fixed noise strategy to reduce VRAM usage while enhancing attack effectiveness.
Priority Guided Explanation for Knowledge Tracing with Dual Ranking and Similarity Consistency
Fan Li (Northeastern University), Ge Yu (Northeastern University)
Explainability and InterpretabilitySequential
🎯 What it does: Proposed a model-agnostic, priority-guided knowledge tracing explanation method;
Privacy Preserving Solution of DCOPs by Local Search
Shmuel Goldklang (Open University of Israel), Tamir Tassa (Open University of Israel)
OptimizationSafty and PrivacyGraph
🎯 What it does: Proposed P-DSA, a privacy-preserving implementation for distributed constraint optimization problems based on local search.
Proactive Data-driven Scheduling of Business Processes
Francesca Meneghello (Fondazione Bruno Kessler), Chiara Ghidini (Free University of Bozen-Bolzano)
OptimizationTabularBiomedical Data
🎯 What it does: This paper proposes an active data-driven business process scheduling method aimed at optimizing resource utilization and minimizing job flow time, particularly in addressing uncertainty challenges in service systems.
Probabilistic Analysis of Stable Matching in Large Markets with Siblings
Zhaohong Sun (Kyushu University), Makoto Yokoo (Kyushu University)
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper studies the existence of stable matching in a multi-child daycare matching market and proposes an improved Sorted Deferred Acceptance algorithm (ESDA) to satisfy a stricter concept of stability;
Probabilistic Multimodal Learning with von Mises-Fisher Distributions
Peng Hu (Sichuan University), Xi Peng (Sichuan University)
ClassificationContrastive LearningMultimodality
🎯 What it does: Proposed a probability-based multi-modal learning framework PML based on the von Mises-Fisher distribution, modeling each modal sample as a directional distribution to capture uncertainty, enhancing representations through vMF-prototype contrastive learning, and achieving robust classification via reliable multi-modal fusion.
Problem-dependent Regret for Lexicographic Multi-Armed Bandits with Adversarial Corruptions
Bo Xue (City University of Hong Kong), Qingfu Zhang (City University of Hong Kong)
OptimizationTabularBiomedical Data
🎯 What it does: Proposes two algorithms for lexicographic multi-objective multi-armed bandits (MOMAB) under adversarial corruption (known budget CMOB-KB and unknown budget CMOB-UB), and provides regret upper bounds that match problem-related lower bounds;
Progressive Modality-Adaptive Interactive Network for Multi-Modality Image Fusion
Chaowei Huang (Fuzhou University), Xiao Ke (Fuzhou University)
RestorationTransformerImageMultimodalityBiomedical Data
🎯 What it does: Proposed a Progressive Modality-Adaptive Interactive Network (PoMAI) framework for multi-modal fusion of infrared and visible images, adopting a two-stage training approach: the first stage separately extracts modal features using NGMM (adapted to infrared sparsity) and CAMN (extract visible details); the second stage freezes the first-stage network and introduces MICM to achieve cross-modal dynamic compensation through gated attention, ultimately generating the fused image.
Progressive Prefix-Memory Tuning for Complex Logical Query Answering on Knowledge Graphs
Xingrui Zhuo (Hefei University of Technology), Xindong Wu (Shandong Inspur Science Research Institute)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraphBenchmark
🎯 What it does: Propose the Progressive Prefix‑Memory Tuning (PPMT) framework, targeting serialized text of complex logical queries (EPFO) in knowledge graphs, utilizing a prefix‑memory rectification mechanism and progressive fine-tuning methods to enhance the reasoning performance of pre-trained language models (PLMs) on incomplete knowledge graphs.
Projection, Interaction and Fusion: A Progressive Difference Fusion Network for Salient Object Detection
Xiao Ke (Engineering Research Center Of Big Data Intelligence Ministry Of Education), Yuzhen Niu (Engineering Research Center Of Big Data Intelligence Ministry Of Education)
Object DetectionComputational EfficiencyTransformerImage
🎯 What it does: Propose a novel salient object detection network PDFNet based on Transformer, aiming to address the bottleneck of scale, shape, and confusion between global-detail information.
ProMEA: Prompt-driven Expansion and Alignment for Single Domain Generalization
Yunyun Wang (University of Posts and Telecommunications), Songcan Chen (Nanjing University of Aeronautics and Astronautics)
Domain AdaptationPrompt EngineeringVision Language ModelContrastive LearningImageBenchmark
🎯 What it does: Enhancing single-source domain generalization performance by leveraging learnable text prompt-driven frequency domain feature expansion and target domain alignment during inference.
Prompt-Aware Controllable Shadow Removal
Kerui Chen (Zhejiang University), Yi Yang (Zhejiang University)
RestorationConvolutional Neural NetworkTransformerPrompt EngineeringImage
🎯 What it does: Propose a controllable shadow removal method based on user prompts (PACSRNet), which can freely specify the subject to remove shadows through various prompts such as points, lines, or subject masks.
Prompt-Free Conditional Diffusion for Multi-object Image Augmentation
Haoyu Wang (Xi'an University of Posts & Telecommunications), Yanning Zhang (Xi'an University of Posts & Telecommunications)
GenerationDiffusion modelImage
🎯 What it does: This paper proposes a prompt-free conditional diffusion framework for multi-object image augmentation;
Prototype-based Optimal Transport for Out-of-Distribution Detection
Ao Ke (University of Science and Technology of China), Lei Feng (Southeast University)
ClassificationAnomaly DetectionConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposes a prototype-based optimal transport (POT) method for detecting out-of-distribution (OOD) samples in test inputs.
Prototype-guided Knowledge Propagation with Adaptive Learning for Lifelong Person Re-identification
Zhijie Lu (Wuhan University), Mang Ye (Wuhan University)
RetrievalMeta LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Developed an example-free lifelong person re-identification method called PKA, combining prototype-guided knowledge propagation and adaptive parameter evolution to prevent catastrophic forgetting.
Proven Approximation Guarantees in Multi-Objective Optimization: SPEA2 Beats NSGA-II
Yasser Alghouass (Ecole Polytechnique), Mohammed Lagmah (Ecole Polytechnique)
OptimizationBenchmark
🎯 What it does: This paper proves through mathematical analysis that Steady-state SPEA2 can achieve optimal distribution (optimal approximation) for the OneMinMax benchmark in polynomial time, and compares its approximation efficiency with NSGA-II on the same problem;
Pseudo-Label Reconstruction for Partial Multi-Label Learning
Yu Chen (Guangdong University of Technology), Xiaozhao Fang (Guangdong University of Technology)
ClassificationOptimizationRepresentation LearningImageBiomedical DataAudio
🎯 What it does: Propose a partially multi-label learning method called PML-PLR based on pseudo-label reconstruction, which utilizes candidate labels and features to co-learn instance associations and reconstruct high-quality pseudo-labels.
Public Signaling in Markets with Information Asymmetry Using a Limited Number of Signals
Xu Zhao (Renmin University of China), Weiran Shen (Renmin University of China)
OptimizationTabularFinance Related
🎯 What it does: Studies how to maximize the buyer's purchase probability through signal schemes in information-asymmetric markets under the scenario where third-party public signals are limited in quantity.
Q-Detection: A Quantum-Classical Hybrid Poisoning Attack Detection Method
Haoqi He (Sun Yat-sen University), Yan Xiao (Sun Yat-sen University)
Anomaly DetectionAdversarial AttackConvolutional Neural NetworkImagePhysics Related
🎯 What it does: Propose a quantum-classical hybrid data poisoning detection method called Q-Detection, which utilizes Q-WAN to train a weight allocation network and filter a clean subset from training data.
Q-MiniSAM2: A Quantization-based Benchmark for Resource-Efficient Video Segmentation
Xuanxuan Ren (Xidian University), Yanhua Yang (Xidian University)
SegmentationComputational EfficiencyContrastive LearningImageVideoBenchmark
🎯 What it does: Propose Q-MiniSAM2, significantly reducing SAM2's memory consumption and computational overhead in video segmentation tasks through post-training quantization, hierarchical video quantization, and adaptive mutual KV.
QiMeng-TensorOp: One-Line Prompt is Enough for High-Performance Tensor Operator Generation with Hardware Primitives
Xuzhi Zhang, Ling Li (Institute of Software Chinese Academy of Sciences)
Computational EfficiencyAI Code AssistantLarge Language ModelPrompt EngineeringChain-of-Thought
🎯 What it does: Proposes a framework called QiMengTensorOp that can automatically generate high-performance tensor operators at the hardware primitive level (e.g., GEMM and convolution) with just a single prompt.
Quantifying the Self-Interest Level of Markov Social Dilemmas
Richard Willis (King's College London), Michael Luck (University of Sussex)
Reinforcement Learning
🎯 What it does: This paper proposes an empirical method to quantify the self-interest level in Markov social dilemmas and evaluates agent cooperation through reward exchange.
QuantileFormer: Probabilistic Time Series Forecasting with a Pattern-Mixture Decomposed VAE Transformer
Yimiao Shao (Nanjing University), Sanglu Lu (Nanjing University)
TransformerAuto EncoderTime Series
🎯 What it does: Proposed the QuantileFormer, a probabilistic time series forecasting framework based on pattern hybrid decomposition
Query-Based and Unnoticeable Graph Injection Attack from Neighborhood Perspective
Chang Liu (Beijing University of Posts and Telecommunications), Xingquan Zuo (Beijing University of Posts and Telecommunications)
Adversarial AttackGraph
🎯 What it does: Designed a model-agnostic, black-box graph injection attack method called QUGIA, which generates injected node features using a Bayesian framework and constructs attack edges from the perspective of victim node neighbors, maintaining graph homophily to achieve stealthy attacks.
R2DQG: A Quality Meets Diversity Framework for Question Generation over Knowledge Bases
Yimeng Ren (Beijing University of Posts and Telecommunications), Mingliang Yan (Beijing University of Posts and Telecommunications)
GenerationLarge Language ModelPrompt EngineeringTextGraph
🎯 What it does: Proposed a two-phase template-guided and error-correction framework named R DQG for generating diverse and semantically accurate knowledge graph question answering problems.
RAMer: Reconstruction-based Adversarial Model for Multi-party Multi-modal Multi-label Emotion Recognition
Xudong Yang (Hong Kong University of Science and Technology (Guangzhou)), Yuyu Luo (Hong Kong University of Science and Technology (Guangzhou))
RecognitionTransformerGenerative Adversarial NetworkContrastive LearningMultimodality
🎯 What it does: Propose the RAMer framework to address the issues of missing modalities, modality heterogeneity, and label correlations in multi-party multi-modal multi-label emotion recognition (MMER).
Randomised Optimism via Competitive Co-Evolution for Matrix Games with Bandit Feedback
Shishen Lin (University of Birmingham)
OptimizationTabular
🎯 What it does: Propose an algorithm named COEBL that combines evolutionary algorithms with multi-armed bandit learning, using a randomized optimistic approach to learn in two-player zero-sum matrix games with unknown payoff matrices.
RDPA: Real-Time Distributed-Concentrated Penetration Attack for Point Cloud Learning
Youtong Shi (Xiamen University), Cheng Wang (Xiamen University)
Adversarial AttackPoint Cloud
🎯 What it does: Proposes a real-time distributed concentrated penetration attack (RDPA) framework for performing minimal point perturbation attacks on 3D point cloud learning models;
Reasoning About Causal Knowledge in Nondeterministic Domains
Shakil M. Khan (University of Regina), Maryam Rostamigiv (University of Regina)
Explainability and InterpretabilityWorld Model
🎯 What it does: This paper proposes a framework based on non-deterministic situation calculus to reason about agents' causal knowledge in non-deterministic domains, integrating knowledge update with causal reasoning.
Recalling The Forgotten Class Memberships: Unlearned Models Can Be Noisy Labelers to Leak Privacy
Zhihao Sui (Shanghai Jiao Tong University), Qi Zhang (Tongji University)
ClassificationSafty and PrivacyKnowledge DistillationAdversarial AttackImage
🎯 What it does: Propose a Membership Recall Attack (MRA) framework based on teacher-student knowledge distillation, which utilizes an unlearned model (ULM) as a noise annotator to recover the class labels of forgotten data without accessing the original training model;
RegionMatch: Pixel-Region Collaboration for Semi-Supervised Semantic Segmentation in Remote Sensing Images
Xiaoqian Zhu (Xidian University), Licheng Jiao (Xidian University)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: Propose the RegionMatch method, which performs semi-supervised semantic segmentation from an object-level (region) perspective, and generates high-quality pseudo-labels through collaborative learning between pixels and regions.
Reinforced In-Context Black-Box Optimization
Lei Song (Nanjing University), Chao Qian (Nanjing University)
OptimizationRobotic IntelligenceTransformerSupervised Fine-TuningSequential
🎯 What it does: Learned an end-to-end black-box optimization algorithm called RIBBO, which uses offline optimization history to train a causal Transformer model to generate query points.
Relation-Augmented Dueling Bayesian Optimization via Preference Propagation
Xiang Xia (East China Normal University), Hong Qian (East China Normal University)
OptimizationContrastive Learning
🎯 What it does: Proposes Relation-Enhanced Dueling Bayesian Optimization (RADBO), which more fully utilizes existing pairwise preference information in dueling black-box optimization through preference propagation techniques.
Relational Decomposition for Program Synthesis
Céline Hocquette (University of Southampton), Andrew Cropper (University of Oxford)
AI Code AssistantImageTextSequential
🎯 What it does: This paper proposes an approach to improve program synthesis by decomposing the program synthesis task into relational subtasks;
Reliable and Calibrated Semantic Occupancy Prediction by Hybrid Uncertainty Learning
Song Wang (Zhejiang University), Jianke Zhu (Zhejiang University)
SegmentationAutonomous DrivingPoint Cloud
🎯 What it does: This paper evaluates the reliability of semantic occupancy prediction models and proposes the RELIOCC method, which combines absolute and relative uncertainty learning with uncertainty-aware temperature scaling calibration to enhance model reliability and calibration;
Reliable and Diverse Hierarchical Adapter for Zero-shot Video Classification
Wenxuan Ge (Nanjing University of Science and Technology), Xiangbo Shu (Nanjing University of Science and Technology)
ClassificationVision Language ModelVideo
🎯 What it does: Propose a training-free hierarchical adapter for zero-shot video classification
Reliable Disentanglement Multi-view Learning Against View Adversarial Attacks
Xuyang Wang (Sichuan University), Dezhong Peng (Sichuan University)
Representation LearningAdversarial AttackImage
🎯 What it does: Propose a reliable decoupled multi-view learning framework that combines evidence decoupling, feature re-calibration, and view-level evidence attention mechanisms to enhance the robustness and reliability of multi-view models against adversarial attacks.
ReplayCAD: Generative Diffusion Replay for Continual Anomaly Detection
Lei Hu (South China University of Technology), Tianshui Chen (Guangdong University of Technology)
Anomaly DetectionDiffusion modelImage
🎯 What it does: Proposed a ReplayCAD framework based on pre-trained diffusion models to compress and replay historical data in continual learning anomaly detection tasks, thereby mitigating catastrophic forgetting and improving pixel-level defect segmentation performance.
RepObE: Representation Learning-Enhanced Obfuscation Encryption Modular Semantic Task Framework
Limei Lin (Fujian Normal University), Jie Wu (China Telecom Cloud Computing Research Institute)
ClassificationSafty and PrivacyAdversarial AttackConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposed the RepObE framework to achieve dynamic encryption during semantic extraction and feature transmission, aiming to defend against model inversion attacks and adversarial attacks.
Representation Learning with Mutual Influence of Modalities for Node Classification in Multi-Modal Heterogeneous Networks
Jiafan Li (Institute of Software, Chinese Academy of Sciences), Hongan Wang (Institute of Software, Chinese Academy of Sciences)
ClassificationRepresentation LearningGraph Neural NetworkTransformerMultimodalityGraph
🎯 What it does: This paper proposes a model named HGNN-IMA for node classification in multi-modal heterogeneous networks, achieving mutual influence and fusion of node representations by embedding cross-modal attention within the heterogeneous graph transformer framework.
RePST: Language Model Empowered Spatio-Temporal Forecasting via Semantic-Oriented Reprogramming
Hao Wang (Hong Kong University of Science and Technology), Hao Liu (Hong Kong University of Science and Technology)
TransformerLarge Language ModelTime Series
🎯 What it does: Propose the REPST framework, which utilizes pre-trained language models (PLM) for spatiotemporal prediction. It maps spatiotemporal sequences into the text vocabulary space through semantic-oriented decomposition and selective reprogramming, with predictions generated by a frozen GPT-2.
Requirement Patterns for Engineering Multiagent Interaction Protocols
Amit K. Chopra (Lancaster University), Munindar P. Singh (North Carolina State University)
Benchmark
🎯 What it does: Proposed the Mambo method for demand-driven path query verification of information protocols based on BSPL, achieving efficient checking through incremental table construction and pruning.
Resistance is Futile: Gradually Declining Immunity Retains the Exponential Duration of Immunity-Free Diffusion
Andreas Göbel, Marcus Pappik (Hasso Plattner Institute, University of Potsdam)
GraphStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose a continuous SIRS (cSIRS) diffusion model, analyze its survival time on star graphs and random graphs, and compare it with traditional SIRS and SIS models.
Responsibility Anticipation and Attribution in LTLf
Giuseppe De Giacomo (University of Oxford), Gianmarco Parretti (University of Rome La Sapienza)
🎯 What it does: Proposed and formalized the concepts of active and passive responsibility for LTLf outcomes, and provided their complexity analysis and algorithms.
Responsibility Gap in Collective Decision Making
Pavel Naumov (University of Southampton), Jia Tao (Lafayette College)
🎯 What it does: Investigated the responsibility gap in collective decision-making mechanisms and proposed the concept of 'electoral dictatorship' to eliminate the responsibility gap.
Rethinking Contrastive Learning in Graph Anomaly Detection: A Clean-View Perspective
Di Jin (Tianjin University), Jianwu Dang (Chinese Academy of Sciences)
Anomaly DetectionGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Proposed a graph anomaly detection framework based on clean views (CVGAD), which enhances the detection of abnormal nodes by identifying and iteratively cleaning noisy edges.
Rethinking Federated Graph Learning: A Data Condensation Perspective
Hao Zhang (Chinese Academy of Sciences), Lianglin Hu (Chinese Academy of Sciences)
Federated LearningKnowledge DistillationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Proposed the FedGM framework, which achieves one-time communication in federated graph learning by utilizing subgraph aggregation techniques from graph data distillation;
Rethinking Graph Contrastive Learning Through Relative Similarity Preservation
Zhiyuan Ning (Computer Network Information Center Chinese Academy Of Sciences), Yuanchun Zhou (Computer Network Information Center Chinese Academy Of Sciences)
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes a contrastive learning framework called RELGCL, based on the decay patterns of graph structural distance and label consistency with distance, aiming to preserve relative similarity rather than traditional absolute similarity.
Rethinking Removal Attack and Fingerprinting Defense for Model Intellectual Property Protection: A Frequency Perspective
Cheng Zhang (Hunan University), Zixing Zhang (Hunan University)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Propose a frequency domain-based model ownership removal attack and the corresponding frequency domain fingerprint defense.