π― What it does: Propose an end-to-end deep hashing framework CRH that dynamically reallocates hash centers during training, achieving joint optimization of hash centers and hash functions.
π― What it does: Propose a full-process machine vision-friendly underwater image compression framework based on vector quantization (VQ) codebooks, covering three stages: feature extraction, compression, and reconstruction, while maintaining excellent machine vision performance at low bitrates.
CoEvo: Continual Evolution of Symbolic Solutions Using Large Language Models
Ping Guo (City University of Hong Kong), Xi Lin (City University of Hong Kong)
CodeOptimizationTransformerLarge Language ModelBenchmarkPhysics Related
π― What it does: Propose the CoEvo framework by combining large language models with evolutionary search, achieving continuous evolution and knowledge management of symbolic solutions.
COGS: A Causal Representation Learning Framework for Out-of-Distribution Generalization in Time Series
Xinxin Song (Tsinghua University), Jinli Suo (Tsinghua University)
CodeDomain AdaptationRepresentation LearningAuto EncoderContrastive LearningTime Series
π― What it does: Propose the CO-GS framework, which leverages causal representation learning and unsupervised domain discovery to achieve out-of-distribution (OOD) generalization for time series
CodeRetrievalCompressionTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityRetrieval-Augmented Generation
π― What it does: Proposes the CogStream task and a corresponding semi-automatically created dataset, challenging models to identify and utilize the most relevant historical context in real-time videos for question answering.
COIN: Uncertainty-Guarding Selective Question Answering for Foundation Models with Provable Risk Guarantees
Zhiyuan Wang (University of Electronic Science and Technology of China), Kaidi Xu (Drexel University)
CodeTransformerLarge Language ModelVision Language ModelTextMultimodality
π― What it does: Proposes the COIN framework, which sets thresholds during answer generation by the base model and filters high-uncertainty samples to ensure the error rate does not exceed the user-specified FDR.
π― What it does: Proposes an end-to-end collaborative feature matching framework CFM, comprising a keypoint learning module and a correspondence learning module, which progressively enhance matching quality through mutual reinforcement.
CoMA-SLAM: Collaborative Multi-Agent Gaussian SLAM with Geometric Consistency
Lin Chen (Northwestern Polytechnical University), Guangming Wang (Northwestern Polytechnical University)
CodePose EstimationDepth EstimationOptimizationComputational EfficiencyGaussian SplattingSimultaneous Localization and MappingImagePoint CloudBenchmark
π― What it does: Proposes CoMA-SLAM, a distributed multi-agent Gaussian SLAM framework, achieving high-precision tracking, rendering, and globally consistent scene reconstruction.
CoMA: Compositional Human Motion Generation with Multi-modal Agents
Shanlin Sun (University of California, Irvine), Xiaohui Xie (Columbia University)
CodeGenerationTransformerLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelAuto EncoderVideoTextSequential
π― What it does: CoMA proposes a framework based on multimodal agents to generate high-quality, editable, and self-correcting 3D human motion from text.
ComoRAG: A Cognitive-Inspired Memory-Organized RAG for Stateful Long Narrative Reasoning
Juyuan Wang (South China University of Technology), Liyan Xu (Tencent)
CodeRetrievalRepresentation LearningLarge Language ModelAgentic AITextRetrieval-Augmented Generation
π― What it does: Introduces ComoRAG, a cognitive-inspired memory-organized retrieval-augmented generation framework for long-form narrative reasoning, capable of dynamically constructing and updating the global context across multiple rounds of retrieval;
π― What it does: Achieve sample-free incremental learning on pre-trained vision Transformers, proposing a compensation method to address distribution drift caused by gradual fine-tuning.
π― What it does: Designed and implemented an end-to-end complex-valued event-RGB fusion framework named CompEvent for low-light video deblurring and enhancement.
Complex Mathematical Expression Recognition: Benchmark, Large-Scale Dataset and Strong Baseline
Weikang Bai (Fudan University), Zhineng Chen (Fudan University)
CodeRecognitionConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelImageTextBenchmark
π― What it does: Studies complex mathematical expression recognition, proposes the CMER-Bench benchmark, the CMER-3M/MER-17M datasets, and introduces the CMERNet model.
Composition-Incremental Learning for Compositional Generalization
Zhen Li (Beijing Institute Of Technology), Yunde Jia (Shenzhen Msu-Bit University)
CodeKnowledge DistillationRepresentation LearningTransformerVision Language ModelAuto EncoderImageTextMultimodalityBenchmark
π― What it does: Proposes a Composition-Incremental Learning (CompIL) setting, enabling models to progressively learn new combinations over time while keeping the primitive set fixed, thereby continuously enhancing compositional generalization ability; simultaneously designs a pseudo-replay framework that generates visual representations of past tasks using a visual synthesizer and maintains consistent primitive representations through language primitive distillation.
π― What it does: Studied a multi-view joint learning and structure-aware contrastive learning framework (MVJC) for urban area embeddings, enhancing crime prediction and land use clustering performance.
Compression Artifacts Removal for VVC with Frequency Domain Mixture of Experts Network
Qijun Wang (Anhui University), Jun Wang (Anhui University)
CodeRestorationConvolutional Neural NetworkTransformerMixture of ExpertsImage
π― What it does: Propose a hybrid expert network named ARMoE that combines multi-frequency domain transformations with a sparse activation strategy to remove various artifacts from H.266/VVC compressed images.
Computing Syntax Tree-based Minimal Unsatisfiable Cores of LTLf Formulas
Valeria Fionda (University of Calabria), Francesco Ricca (University of Calabria)
CodeTextBenchmark
π― What it does: This paper proposes a method based on Answer Set Programming (ASP) for computing minimal unsatisfiable cores (MUC) and minimal correction sets (MCS) at the syntax tree level in finite-trace linear temporal logic (LTLf), supporting arbitrary formulas rather than only conjunctions.
Concepts from Representations: Post-hoc Concept Bottleneck Models via Sparse Decomposition of Visual Representations
Shizhan Gong (Chinese University of Hong Kong), Qi Dou (Chinese University of Hong Kong)
CodeClassificationExplainability and InterpretabilityRepresentation LearningLarge Language ModelAuto EncoderImageMultimodalityBenchmark
π― What it does: Propose the posterior concept bottleneck model PCBM-ReD, which automatically decomposes representations from pre-trained visual models into sparse interpretable concepts and uses a linear layer to predict categories.
CondDiff-AMO: Integrating Conditional Diffusion Mechanism for Unified Amodal Mask Generation
CaiJie Zhao (University of Macau), Bob Zhang (University of Macau)
CodeSegmentationTransformerDiffusion modelImage
π― What it does: Propose CondDiff-AMO, which treats the task of generating complete masks for blindly occluded objects as a step-by-step denoising problem using diffusion models, providing a single-stage, unsupervised panoramic mask prediction framework.
π― What it does: Proposed a semi-supervised graph classification method CDL, which learns graph representations by aligning conditional distributions between weakly and strongly augmented views.
π― What it does: Proposed a multimodal fusion framework MCIB based on conditional information bottleneck, achieving multimodal sarcasm detection on the MUStARD++ R dataset with shortcut learning signals removed.
Conditional Prompt Learning via Degradation Perception for Underwater Image Enhancement
Mingze Yao (Dalian Maritime University), Huibing Wang (Dalian Maritime University)
CodeRestorationConvolutional Neural NetworkTransformerPrompt EngineeringVision Language ModelContrastive LearningImage
π― What it does: Proposed a Conditional Prompt Learning via Degradation Perception (CPLDP) model, which enhances visual quality by achieving perception and enhancement of different underwater degradations through conditional prompt learning, dynamic learning network, and adaptive loss function.
Conformal Prediction for Multi-Source Detection on a Network
Xingchao Jian (Technical University of Munich), Felix Krahmer (Technical University of Munich)
CodeAnomaly DetectionGraph Neural NetworkGraph
π― What it does: Propose a network multi-source detection framework based on homomorphic prediction, which can estimate the set of infection sources without relying on diffusion model assumptions.
ConInstruct: Evaluating Large Language Models on Conflict Detection and Resolution in Instructions
Xingwei He, Siu-Ming Yiu (University of Hong Kong)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: Construct the ConInstruct dataset and evaluate the conflict detection and resolution capabilities of LLMs under instructions with conflicting constraints.
Connecting the Dots: Training-Free Visual Grounding via Agentic Reasoning
Liqin Luo (Peking University), Yonghong Tian (Peking University)
CodeRetrievalTransformerLarge Language ModelAgentic AIVision Language ModelImageTextMultimodalityChain-of-Thought
π― What it does: This paper proposes a training-agnostic visual alignment framework named GroundingAgent, which leverages open-vocabulary detectors, cross-modal large language models, and regular large language models for structured agent reasoning to achieve zero-shot visual localization.
π― What it does: Proposed a sparsification framework called Co-Sparsify, guided by graph connectivity, to enhance the efficiency of 2-FWL graph neural networks while maintaining full expressive power.
Consensus-Aligned Neuron Efficient Fine-Tuning Large Language Models for Multi-Domain Machine Translation
Shuting Jiang (Kunming University of Science and Technology), Zhengtao Yu (Kunming University of Science and Technology)
CodeDomain AdaptationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes the CANEFT framework, achieving efficient fine-tuning for multi-domain machine translation by identifying and updating only consistent aligned neurons;
π― What it does: Propose a molecular generation model based on context-free grammar (CFG) constraints using SMILES, incorporating multiple chemical and physicochemical property constraints.
Ramiz Gindullin (Uppsala University), MarΓa AndreΓna Francisco RodrΓguez (Uppsala University)
CodeOptimizationBiomedical Data
π― What it does: Propose a constrained optimization model (COMPD) to generate microplate layouts, aiming to reduce plate effects and improve experimental data quality.
Constraint-Guided Clustering for Identifying in-Vehicle Electronic Control Units from Voltage Data
Bogdan Groza (Politehnica University of Timisoara), Lucian Popa (Politehnica University of Timisoara)
CodeAutonomous DrivingTime Series
π― What it does: Studied unsupervised clustering based on voltage features in vehicle internal electronic control units (ECUs) and proposed a method that utilizes domain constraints to guide the selection of the number of clusters.
Constraints-Guided Diffusion Reasoner for Neuro-Symbolic Learning
Xuan Zhang (Fudan University), Yuan Qi (Fudan University)
CodeOptimizationReinforcement LearningDiffusion model
π― What it does: Propose a diffusion model named DDReasoner, serving as a symbolic reasoner. It first acquires basic reasoning capabilities through supervised learning with masked DDPM, then treats the denoising process as a Markov Decision Process (MDP), and fine-tunes the model using reinforcement learning (PPO-CLIP) to ensure outputs meet hard logical constraints.
π― What it does: This paper proposes the E-Bike Rider Search (EBRS) dataset specifically for urban electric bike riders and designs a Context-aware Dynamic Contrastive Learning (CDCL) framework combining context-aware mixed convolution and dynamic contrastive learning, aiming to address occlusion, pose variation, and scale differences in person search tasks.
Ningbo Huang (State Key Laboratory of Mathematical Engineering and Advanced Computing), Yi Xia (State Key Laboratory of Mathematical Engineering and Advanced Computing)
CodeClassificationRepresentation LearningMeta LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmark
π― What it does: Proposes CAGML, a context-aware meta-learning model that unifies cross-domain and cross-granularity graph few-shot tasks through structural-aware sequence modeling, achieving in-graph parameter learning without fine-tuning via pre-trained Structure-Aware Transformer (SAT).
Context-Sensitive Abstractions for Reinforcement Learning with Parameterized Actions
Rashmeet Kaur Nayyar (Arizona State University), Siddharth Srivastava (Arizona State University)
CodeReinforcement Learning
π― What it does: This paper proposes the PEARL algorithm, which can adaptively learn context-aware abstractions for states and parameterized actions in reinforcement learning, and uses TD(Ξ») for policy learning in the abstract space.
π― What it does: Studied out-of-distribution (OOD) detection in the continual learning (CIL) scenario, proposing a training-free Analytic Neural Collapse (AnaNC) method that achieves continuous OOD detection and classification by constructing Neural Collapse geometry through analytical projection on frozen pre-trained models.
π― What it does: Propose a continuous degradation modeling framework called DegFlow that leverages potential flow matching, enabling the synthesis of realistic low-resolution images of arbitrary scales from a single high-resolution image, thereby generating large-scale realistic scene SR training data.
Continuum Dropout for Neural Differential Equations
Jonghun Lee (Ulsan National Institute of Science and Technology), Dong-Young Lim (Ulsan National Institute of Science and Technology)
CodeClassificationExplainability and InterpretabilityImageTime SeriesBiomedical DataElectronic Health RecordsStochastic Differential EquationOrdinary Differential EquationAudio
π― What it does: Proposed a continuous-time dropout mechanism called Continuum Dropout, and applied it to Neural Differential Equations (NDE) to enhance model generalization capability and uncertainty quantification.
Control Illusion: The Failure of Instruction Hierarchies in Large Language Models
Yilin Geng (The University of Melbourne), Lea Frermann (The Hebrew University of Jerusalem)
CodeTransformerLarge Language ModelText
π― What it does: Evaluate the hierarchical priority of large language models when facing conflicting instructions, propose a systematic assessment framework based on constraint prioritization and new metrics, and systematically test the performance of six mainstream LLMs;
Libo Zhao (Jilin University), Zeyu Wang (Dalian Minzu University)
CodeRestorationGraph Neural NetworkTransformerVision Language ModelContrastive LearningMultimodality
π― What it does: Proposes ControlFuse, an infrared-visible image fusion method controllable at global, semantic, and instance levels through user instructions.
π― What it does: Study the convergence of fast policy iteration algorithms for Markov games and robust MDPs, prove that the original Filar-Tolwinski (FT) algorithm may not converge, and propose a new Residual Conditioned Policy Iteration (RCPI) algorithm that ensures convergence while maintaining speed.
Conversational Learning Diagnosis via Reasoning Multi-Turn Interactive Learning
Fangzhou Yao (University of Science and Technology of China), Qi Liu (University of Science and Technology of China)
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextSequentialChain-of-Thought
π― What it does: Proposed the ParLD framework, which uses behavior preview, state analysis, and performance reasoning in multi-turn dialogues to track students' cognitive states in real-time and support teaching decisions.
ConvMix: A Mixed-Criteria Data Augmentation Framework for Conversational Dense Retrieval
Fengran Mo (University of Montreal), Jian-Yun Nie (University of Montreal)
CodeRetrievalData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes the ConvMix framework, which utilizes large language models to perform bidirectional relevance judgment augmentation on conversational retrieval data, and combines quality control with near-distribution supervision to significantly enrich the training set.
Cook and Clean Together: Teaching Embodied Agents for Parallel Task Execution
Dingkang Liang (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
CodeOptimizationRobotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelPoint Cloud
π― What it does: Proposes the 3D Oriented Task Scheduling (ORS3D) problem based on operations research knowledge, requiring agents to simultaneously generate efficient parallel task plans and locate target objects in real 3D scenarios.
π― What it does: Proposes the CoRe-Fed framework by simultaneously addressing representation fairness and collaborative fairness in federated learning;
π― What it does: Propose a test-time adaptation framework called TUNE to address distribution drift caused by the emergence of unknown normal patterns in graph anomaly detection.
CodeOptimizationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
π― What it does: This paper proposes the Chain-of-Scheduling (CoS) framework, which leverages the ordered reasoning of large language models across three stagesβexploration, verification, and integrationβto automatically generate event schedules that comply with time and location constraints;
π― What it does: This paper proposes a label ranking weight-based cost-sensitive conformal training algorithm (RWCE), which directly optimizes the weighted cross-entropy of true label rankings to approximate and upper bound the prediction set size;
Counterfactual eXplainable AI (XAI) Method for Deep Learning-Based Multivariate Time Series Classification
Alan Gabriel Paredes Cetina (SnT University of Luxembourg), Sylvain Kubler (SnT University of Luxembourg)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkTime Series
π― What it does: Proposed a multi-objective adversarial counterfactual explanation method for time series called CONFETTI, which can significantly reduce the modification of time series while maintaining confidence and enhancing proximity.
Bo Zhang (Nanjing Normal University), Junsheng Zhou (Nanjing Normal University)
CodeGenerationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the GapProbe framework for generating adversarial follow-up questions (CFQ) by interacting Large Language Models (LLMs) with knowledge graphs (KGs) to reveal hidden contradictions in learners' answers and promote reflection.
π― What it does: Proposed Covariance Scattering Transforms (CST), a training-free deep network that achieves data representation through local wavelet filtering of covariance spectra.
Coverage-Constrained Human-AI Cooperation with Multiple Experts
Zheng Zhang (University of Surrey), Gustavo Carneiro (University of Surrey)
CodeClassificationMixture of ExpertsImageBiomedical Data
π― What it does: This paper proposes the CL2DC method, achieving a coverage-constrained human-AI collaborative classification system that integrates learning-based delay (L2D) and compensation (L2C) in a multi-expert environment.
CP-Search: A Chain Progressive Search Training Framework Incentivizing the Cognitive Behaviors for Searching in LLMs
Zehua Wang (Harbin Institue of Technology (Shenzhen)), Buzhou Tang (Harbin Institue of Technology (Shenzhen))
CodeRetrievalOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
π― What it does: This study proposes CP-Search, a two-phase training framework designed to enhance the retrieval capabilities of large language models in multi-hop reasoning scenarios;
π― What it does: Propose the CRΒ³ framework, which enhances the compositional reasoning ability of multimodal large language models (MLLMs) through rule-based reinforcement learning.
CRAF: A Clinical Reasoning-Adaptive Framework via Reinforcement Learning for Similar Case Retrieval
Jie Lin (Xiamen University), Liansheng Wang (Xiamen University)
CodeRetrievalTransformerLarge Language ModelReinforcement LearningBiomedical Data
π― What it does: Proposed the CRAF framework for clinical similar case retrieval, achieving query rewriting through reinforcement learning to generate diagnostic reasoning paths.
Creating Blank Canvas Against AI-enabled Image Forgery
Qi Song (Hong Kong Baptist University), Renjie Wan (Hong Kong Baptist University)
CodeAnomaly DetectionSafty and PrivacyAdversarial AttackTransformerImage
π― What it does: Proposed an active image forgery localization method that converts the original image into a 'blank canvas' through frequency domain adaptive adversarial perturbation, making the SAM model unable to recognize image content in a seamless state; when the image is maliciously tampered, SAM can detect anomalous regions.
Cross Modal Fine-grained Alignment via Granularity-aware and Region-uncertain Modeling
Jiale Liu (South China Normal University), Yuncheng Jiang (Harbin Institute of Technology)
CodeRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Propose the GRM framework, employing significance-aware and granularity-aware adapters, region prompts, and Gaussian Mixture Uncertainty Modeling to achieve cross-modal fine-grained image-text alignment.
Cross-domain Joint Learning with Prototype-guided Mixture-of-Experts for Infrared Moving Small Target Detection
Weiwei Duan (University of Electronic Science and Technology of China), Mao Ye (University of Electronic Science and Technology of China)
CodeObject DetectionConvolutional Neural NetworkMixture of ExpertsContrastive LearningImage
π― What it does: Propose a cross-domain joint learning framework named CoMoE to address the limitations of single-domain learning in infrared small target detection, constructing a detector that is generalizable across multiple datasets.
Cross-Field Interface-Aware Neural Operators for Multiphase Flow Simulation
Zhenzhong Wang (Xiamen University), Min Jiang (Xiamen University)
CodeConvolutional Neural NetworkTransformerPhysics Related
π― What it does: Propose an Interface Information-Aware Neural Operator (IANO) specifically designed for high-precision, low-cost simulation of multiphase flows.
Cross-Granularity Hypergraph Retrieval-Augmented Generation for Multi-hop Question Answering
Changjian Wang (Mashang Consumer Finance Co Ltd), Ning Jiang (Mashang Consumer Finance Co Ltd)
CodeGenerationRetrievalGraph Neural NetworkLarge Language ModelTextGraphRetrieval-Augmented Generation
π― What it does: Proposed a cross-grained hypergraph retrieval augmented generation (HGRAG) framework, utilizing entity hypergraphs, hypergraph diffusion retrieval, and retrieval augmented modules to jointly exploit structural and semantic information from fine-grained entities and coarse-grained paragraphs for multi-hop question answering retrieval and answer generation.
Cross-modal Prompting for Balanced Incomplete Multi-modal Emotion Recognition
Wen-Jue He (Harbin Institute of Technology), Zheng Zhang (Harbin Institute of Technology)
CodeRecognitionPrompt EngineeringMultimodality
π― What it does: This paper proposes a cross-modal prompting (ComP) framework that enhances the robustness of incomplete multimodal sentiment recognition by leveraging cross-modal knowledge propagation and prompt generation.
Cross-Modal Unlearning via Influential Neuron Path Editing in Multimodal Large Language Models
Kunhao Li (South China University Of Technology), Jason Xue (Southeast University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelMultimodalityBenchmark
π― What it does: Proposed a cross-modal machine unlearning method called MIP-Editor, which can achieve targeted forgetting of specific knowledge in multimodal large language models by locating and editing path neurons that influence the model, while maintaining the model's overall capabilities.
Cross-Sample Augmented Test-Time Adaptation for Personalized Intraoperative Hypotension Prediction
Kanxue Li, Baosheng Yu (Yunnan United Vision Technology Company Limited)
CodeRetrievalDomain AdaptationAnomaly DetectionTransformerAuto EncoderContrastive LearningTime SeriesBiomedical DataElectronic Health Records
π― What it does: Propose the CSA-TTA framework, achieving test-time personalized prediction of intraoperative hypotension (IOH) through cross-sample augmentation, coarse-to-fine retrieval, and multi-task optimization.
CodeRecommendation SystemTransformerLarge Language ModelTextSequential
π― What it does: Propose the CoderRec framework, which combines semantic IDs generated by large language models (LLMs) with a lightweight sequential recommender through cross-scale model collaboration, and introduces domain-specific latent reasoning to enhance sequential recommendation performance.
π― What it does: Proposes a cross-view anchor graph learning and decomposition (AGLF) method for handling incomplete multi-view clustering problems, directly performing subgraph learning and generating soft labels through the anchor graph without requiring post-processing.
π― What it does: Propose a cross-perspective progressive feature filtering multi-view graph clustering framework, CF-MVGC, for unsupervised clustering of remote sensing data.
CodePrompt EngineeringVision Language ModelMultimodalityBenchmark
π― What it does: Designed and constructed the CrossCheck-Bench benchmark, which includes three cognitive levels (perception, integration, reasoning) and seven atomic capabilities, generating 15,000 conflicting question-answer pairs derived from real e-commerce listings.
CrossVid: A Comprehensive Benchmark for Evaluating Cross-Video Reasoning in Multimodal Large Language Models
Jingyao Li (Xiaohongshu Inc), Yao Hu (Xiaohongshu Inc)
CodeLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
π― What it does: Proposed the CrossVid benchmark to evaluate the capabilities of multimodal large language models in cross-video reasoning (CVR) tasks.
CSP4SDG: Constraint and Information-Theory Based Role Identification in Social Deduction Games with LLM-Enhanced Inference
Kaijie Xu (McGill University), Simon Mark Lucas (Queen Mary University of London)
CodeExplainability and InterpretabilityLarge Language ModelText
π― What it does: Proposed a training-free, constraint-based probabilistic framework called CSP4SDG for character inference in social deduction games, which integrates LLM to automatically extract language-agnostic constraints and update character posterior distributions in real-time.
π― What it does: Proposed a controllable infrared-visible image fusion framework called CtrlFuse based on mask prompts, which can dynamically control the fusion process according to user-specified masks while simultaneously enhancing downstream semantic segmentation performance.
CTX-Coder: Cross-Attention Architectures Empower LLMs for Long-Context Vulnerability Detection
Jujie Wang (Beijing University of Posts and Telecommunications), Minjiao Yang (Beijing University of Posts and Telecommunications)
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposed the CTX-Coder framework, which leverages cross-attention to integrate context function embedding vectors with target functions, thereby enhancing vulnerability detection performance in long contexts.
CodeAnomaly DetectionSupervised Fine-TuningReinforcement LearningVision Language ModelVideoBenchmark
π― What it does: Proposed the CUEBENCH benchmark, constructed a five-layer hierarchical dictionary containing absolute and conditional anomalous events, and provided a unified evaluation framework with multi-task (recognition, detection, localization, prediction) implementations for video anomaly understanding; simultaneously developed the CUE-R1 model, utilizing SFT+RFT RL fine-tuning to achieve unified generative anomaly understanding.
π― What it does: This paper proposes a forward compression framework named D-FCGS for scene-agnostic compression of dynamic 3D Gaussian Splatting, to support free-viewpoint video (FVV) applications.
D2 Prune: Sparsifying Large Language Models via Dual Taylor Expansion and Attention Distribution Awareness
Lang Xiong (Chongqing University), Duo Liu (Chongqing University)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelImageText
π― What it does: For post-training sparsification of large language models, Dβ―Pruneβ―2 is proposed to achieve precise pruning and weight updates through dual Taylor expansion and attention distribution awareness.
π― What it does: Designed a lightweight visual localization method D2-VPR based on a vision foundation model, achieving efficient feature learning through knowledge distillation and deformable aggregation.
D2MoRA: Diversity-Regulated Asymmetric MoE-LoRA Decomposition for Efficient Multi-Task Adaptation
Jianhui Zuo (Shandong University), Liqiang Nie (Harbin Institute of Technology)
CodeComputational EfficiencyRepresentation LearningMixture of ExpertsText
π― What it does: Propose D MoRA, a multi-task adaptation multi-expert LoRA framework that achieves flexible knowledge sharing through asynchronous expert splitting.
π― What it does: Proposed a density-agnostic, class-aware point cloud completion framework called DANCE, which can complete missing regions under any sparse input while preserving existing geometry;
DanceHA: A Multi-Agent Framework for Document-Level Aspect-Based Sentiment Analysis
Lei Wang (Temple University), Eduard Dragut (Independent Researcher)
CodeClassificationKnowledge DistillationTransformerLarge Language ModelAgentic AITextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: This paper proposes a multi-agent framework called DanceHA for handling the document-level aspect-based sentiment intensity analysis (ABSIA) task, with a particular focus on informal writing styles;
π― What it does: Propose a unified DETR framework DAPE, which achieves co-optimization of content and position encoding through Shifted Query Sampler and Low-Rank Position Encoder, applicable to object detection, instance segmentation, and few-shot detection.
π― What it does: Propose a dual-alignment prompt framework called DAPrompt for few-shot graph learning in heterophilic graph environments, addressing issues of graph structural noise and semantic inconsistency.
DarkFarseer: Robust Spatio-Temporal Kriging Under Graph Sparsity and Noise
Zhuoxuan Liang, Moustafa Youssef (American University in Cairo)
CodeRepresentation LearningGraph Neural NetworkContrastive LearningGraphTime Series
π― What it does: This paper proposes the DarkFarseer framework to improve the inductive spatiotemporal kriging (ISK) method based on graph neural networks, addressing issues such as insufficient spatiotemporal feature capture, graph sparsity, and noise in virtual node inference.
DARLING: Dual Hypergraph-Enhanced Curriculum-Guided Graph Structure Learning for Node Classification
Guangkai Wu, Zhongying Zhao (Shandong University of Science and Technology)
CodeClassificationGraph Neural NetworkGraph
π― What it does: Proposed a graph structure learning framework named DARLING, aiming to enhance node classification effectiveness through adaptive curriculum learning, dual hypergraph similarity learning, and multi-view graph structure enhancement.
Data Heterogeneity and Forgotten Labels in Split Federated Learning
Joana Tirana (University College Dublin), Nicolas Kourtellis (TelefΓ³nica Scientific Research)
CodeFederated LearningConvolutional Neural NetworkMixture of ExpertsImage
π― What it does: Studied the catastrophic forgetting (CF) phenomenon caused by data heterogeneity in split federated learning (SFL), and proposed the Hydra method to alleviate this issue.
Data-Centric Sequential Recommendation with Relation-Augmented Generation
Yichen Li (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
CodeRecommendation SystemData-Centric LearningTransformerLarge Language ModelGraphSequential
π― What it does: Propose the RaSR framework, which utilizes multi-relational graphs (co-occurrence, temporal, semantic) to augment and regenerate sequential recommendation data, thereby improving data quality and enhancing the performance of subsequent models;
Yifeng Gu (South China University of Technology), Xiangmin Xu (Foshan University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Proposed DAVID, a two-stage adaptive KV cache eviction strategy for multimodal large language models, dynamically distinguishing the separation and fusion phases of visual and text modalities, significantly reducing KV cache occupancy and improving inference speed.
DBGroup: Dual-Branch Point Grouping for Weakly Supervised 3D Semantic Instance Segmentation
Xuexun Liu (Shenzhen University), Xu Wang (Meituan)
CodeSegmentationConvolutional Neural NetworkVision Language ModelPoint Cloud
π― What it does: Propose DBGroup, a weakly supervised 3D instance segmentation framework based on scene-level labels, which generates pseudo labels through dual-branch point grouping and achieves segmentation via multi-round self-training.
Yanqi Wu (Sun Yat-sen University), Ruixuan Wang (Sun Yat-sen University)
CodeAnomaly DetectionImageVideoBenchmark
π― What it does: Propose a training-free, dynamic class-aware cache (DCAC) during testing, which enhances out-of-distribution (OOD) detection by collecting high-entropy samples and leveraging their visual features to calibrate model outputs.
π― What it does: Propose an unsupervised multi-shape matching framework DcMatch, which achieves more accurate correspondences by leveraging dual-layer consistency.
De-collapsing User Intent: Adaptive Diffusion Augmentation with Mixture-of-Experts for Sequential Recommendation
Xiaoxi Cui (Beijing Institute of Technology), Xiangmin Zhou (RMIT University)
CodeRecommendation SystemRecurrent Neural NetworkTransformerMixture of ExpertsDiffusion modelContrastive LearningSequential
π― What it does: This paper proposes the ADARec framework, which adaptively generates user intent hierarchies from extremely sparse user behavior sequences by leveraging the denoising trajectory of diffusion models, and enhances the accuracy of sequence recommendation through a hybrid expert network for decoding the hierarchy.
Debate over Mixed-knowledge: A Robust Multi-Agent Reasoning Framework for Incomplete Knowledge Graph Question Answering
Jilong Liu (Hefei University of Technology), Richang Hong (Hefei University of Technology)
CodeComputational EfficiencyData-Centric LearningTransformerLarge Language ModelAgentic AITextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the DoM (Debate over Mixed-knowledge) framework, which dynamically integrates structured knowledge graphs and unstructured text through multi-agent debate (KG Agent, RAG Agent, and Judge Agent), enabling question-answering reasoning on incomplete knowledge graphs; and constructs a more realistic IKGWQ dataset.
Debiased Cognitive Diagnosis: A Contrastive Counterfactual Modeling Method via Variational Autoencoder
Shangshang Yang (Anhui University), Xingyi Zhang (Anhui University)
CodeAuto EncoderContrastive LearningTabular
π― What it does: Propose a debiasing framework DBCD in cognitive diagnosis, eliminating bias caused by students' selective responses by aligning the predictive distributions of factual and counterfactual data.