These 2140 AAAI 2026 papers come with a code repository. Each shows an AI one-line summary below — get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every AAAI 2026 paper, free trial on arXivSub.
“As Eastern Powers, I Will Veto.”: An Investigation of Nation-Level Bias of Large Language Models in International Relations
Jonghyeon Choi (Yonsei University), Beakcheol Jang (Yonsei University)
CodeTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper systematically evaluates country-level bias in large language models within the field of international relations, constructing a real-world dataset using United Nations Security Council resolutions and voting records, and designing three experiments (direct question answering, association testing, and voting simulation) to quantify bias.
🎯 What it does: This paper proposes the Gaussian Spatial Transport (GST) framework for point annotation density regression tasks. By first performing 2D Gaussian splatting on the input image to estimate the conditional probability between pixels and annotated points, a precomputable transport kernel is constructed; subsequently, during network training, the predicted density map is projected into the annotation space with a single matrix multiplication to compute the error, thereby avoiding the high computational cost of traditional OT iterative solutions for transport plans.
🎯 What it does: Proposed a 2D-compatible scanning framework called 2D-CrossScan, improving state space models such as Mamba to enable spatially consistent multi-path information propagation in 2D images.
3D-DRES: Detailed 3D Referring Expression Segmentation
Qi Chen (Xiamen University), Liujuan Cao (Xiamen University)
CodeSegmentationConvolutional Neural NetworkTransformerVision Language ModelTextPoint Cloud
🎯 What it does: This paper proposes the detailed 3D referring expression segmentation (3D-DRES) task, which requires the model to map each noun phrase in a sentence to its corresponding 3D instance and generate a segmentation mask.
🎯 What it does: Proposed a physics-based anisotropic three-dimensional diffusion model (3DDM) for recovering complete 3D models from partial point clouds.
🎯 What it does: Using the pre-trained SAM2 model, first render 3D dental meshes into multi-view 2D images, segment these images, and then project the 2D segmentation results back into 3D space through voting and Graph Cut, achieving automated 3D dental instance segmentation and semantic annotation.
🎯 What it does: Proposes a learning rate sensitivity-based adaptive warm-up strategy (SAWU) for the warm-up phase in large model fine-tuning, dynamically adjusting loss, learning rate, and phase switching;
🎯 What it does: This paper studies the challenges of discrete entities in zero-shot relation triplet extraction (ZSRTE), proposing a new framework based on the Boundary Marking Graph (BoG) that transforms triplet extraction into a graph path search problem.
A Domain-specific Heuristic for PDDL+-based Traffic Signal Optimisation
Francesco Doria (University of Calabria), Mauro Vallati (University of Huddersfield)
CodeOptimizationTabularTime Series
🎯 What it does: This paper proposes a domain-specific heuristic hCAFE for PDDL+ traffic signal optimization and combines it with Greedy Best-First Search to build a planning framework deployable on actual traffic control infrastructure.
🎯 What it does: This paper theoretically derives how label-preserving data augmentation enhances model robustness under distribution shift through the perspective of flat minima.
🎯 What it does: Proposed and implemented Turbo, a discrete constraint solver that runs entirely on the GPU, employing integer interval boundary propagation and backtracking search. The core technology involves converting the constraint network into a ternary constraint network (TCN) and performing parallel propagation and on-demand subproblem search on the GPU.
🎯 What it does: Proposes a dual-space framework that jointly optimizes image registration and fusion to address error alignment and information fusion in infrared and visible images.
A Multi-Agent LLM Framework for Multi-Domain Low-Resource In-Context NER via Knowledge Retrieval, Disambiguation and Reflective Analysis
Wenxuan Mu, Yijia Zhang (Dalian Minzu University)
CodeRecognitionLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a multi-agent large language model framework called KDR-Agent, specifically designed for context learning in named entity recognition in multi-domain low-resource scenarios.
CodeClassificationTransformerMixture of ExpertsContrastive LearningMultimodalityBiomedical Data
🎯 What it does: This study proposes a multimodal EEG-eye movement model called E Mo, which automatically detects depression using EEG and eye movement signals.
A Network of Biologically Inspired Rectified Spectral Units (ReSUs) Learns Hierarchical Features Without Error Backpropagation
Shanshan Qin (Flatiron Institute), Dmitri Chklovskii (Flatiron Institute)
CodeImage TranslationRepresentation LearningTime SeriesBiomedical Data
🎯 What it does: Designed and trained a self-supervised neural network based on past-future Canonical Correlation Analysis (CCA) - Rectified Spectral Units (ReSUs) - achieving hierarchical features analogous to fly visual motion detection in a 1D natural image translation task.
🎯 What it does: Built an evaluation framework based on task approximate invariance, using this framework to systematically evaluate the reliability of existing evaluation metrics in multi-output structured prediction tasks (image colorization, machine translation, image captioning), and added invariance auxiliary loss during model training, comparing model performance across multiple metrics before and after invariance regularization.
A Novel Fine-Tuned CLIP-OOD Detection Method with Double Loss Constraint Through Optimal Transport Semantic Alignment
Hengyang Lu (Jiangnan University), Chenyou Fan (Anhui University)
CodeAnomaly DetectionRepresentation LearningTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper proposes a Fine-Tuned Out-of-Distribution (OOD) detection method called DOT-OOD for Vision-Language models (CLIP), addressing the low-focus attention problem.
🎯 What it does: Proposed the Retrieve-Read-Group three-phase paradigm, and implemented the normalization of entity phrases in open knowledge bases (OKB) using the self-supervised DUVK framework.
🎯 What it does: This paper proposes a lightweight similarity learning network called LITOFNET, which directly estimates depth from SPAD LiDAR histograms, avoiding traditional signal filtering methods.
🎯 What it does: Proposed a pseudo-label optimization method (PMPC) based on polar coordinate modeling and prior constraints to enhance the accuracy of medical image semi-supervised segmentation and reduce false positives.
CodeRecognitionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a Chain-of-Thought (CoT) based named entity recognition framework called ReasoningNER, which achieves explicit reasoning for entities through a three-phase process (CoT generation, CoT fine-tuning, and inference enhancement);
A Scalable and Exact Relaxation for Densest k-Subgraph via Error Bounds
Ya Liu (Chinese University of Hong Kong), Aritra Konar (Chinese University of Hong Kong)
CodeOptimizationGraph
🎯 What it does: Proposes an exact penalty continuous relaxation method based on the error bound principle for solving the dense k-subgraph (DkS) problem.
🎯 What it does: Proposes a Solution Space Transformation Guided Coevolutionary Algorithm (SSTCE) for the Energy-Saving Distributed Heterogeneous Flexible Job Shop Scheduling Problem (ES-DHFJSP), significantly improving scheduling quality and reducing energy consumption.
A Text-Routed Sparse Mixture-of-Experts Model with Explanation and Temporal Alignment for Multi-Modal Sentiment Analysis
Dongning Rao (Guangdong University of Technology), Jujian Lv (Guangdong Polytechnic Normal University)
CodeClassificationExplainability and InterpretabilityLarge Language ModelMixture of ExpertsMultimodality
🎯 What it does: This paper proposes a text routing sparse expert model that combines multi-modal large language models to generate explanations and time alignment for multi-modal sentiment analysis.
CodeTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Develop and evaluate an adaptive teaching agent, Inquizzitor, based on large language models, integrating Evidence-Centered Design, Social Cognitive Theory, and the Zone of Proximal Development framework, to implement immediate formative assessment and personalized feedback in middle school Earth science STEM+C courses.
🎯 What it does: To address the challenge of matching hand-drawn sketches with RGB surveillance images, the KTCAA framework is proposed, which achieves transfer from the data-rich RGB domain to the sketch domain with limited samples through local sketch-style enhancement and adversarial perturbations of RGB images under a meta-learning framework.
A TSP-Based Algorithm for Multi-League Traveling Tournament
Jingyang Zhao (University of Electronic Science and Technology of China), Ken-Ichi Kawarabayashi
CodeOptimization
🎯 What it does: This paper studies the p-partite Traveling Tournament Problem (p-partite TTP), proposes an efficient algorithm based on the Traveling Salesman Problem (TSP), and proves that the problem is NP-hard when p≥3.
🎯 What it does: This paper conducts a unified convergence analysis of two server-to-device communication primitives in semi-decentralized federated learning—Sampled-to-Sampled (S2S) and Sampled-to-All (S2A)—and provides both theoretical and experimental comparisons;
A Unified Shape-Aware Foundation Model for Time Series Classification
Zhen Liu (South China University of Technology), Qianli Ma (South China University of Technology)
CodeClassificationExplainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningTime Series
🎯 What it does: Proposes UniShape, a unified shape-aware foundation model for time series classification, which enhances cross-domain generalization capability through pre-training + transfer learning.
CodeLarge Language ModelAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: Designed and implemented the A²Flow framework, which automatically generates agentic workflows from expert examples without requiring manual definition of operators.
AbductiveMLLM: Boosting Visual Abductive Reasoning Within MLLMs
Boyu Chang, Tianfei Zhou (Beijing Institute of Technology)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningVideoTextMultimodality
🎯 What it does: Proposed a multimodal large language model called AbductiveMLLM that integrates language and visual imagination to enhance visual abductive reasoning (VAR) capabilities.
🎯 What it does: Proposed the Swin-VIB framework, which uses variational information bottleneck and sliding windows to adaptively filter retrieval contexts, alleviating knowledge conflicts in retrieval-enhanced LLMs and improving response reliability.
🎯 What it does: Propose the ACID-Style framework, which employs a lightweight content and style injection adapter to achieve efficient arbitrary style transfer on Stable Diffusion.
Action-and-object Aware Alignment for Partially Relevant Video Retrieval
Chuanshen Chen (South China University of Technology), Mingkui Tan (South China University of Technology)
CodeRetrievalConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: Propose a dual-branch framework A3PRVR for retrieving partially relevant video segments from untrimmed videos based on text queries, employing both action and object visual features, and incorporating query-specific deformable temporal attention (Q-DTA) and action/object-aware contrastive loss to achieve fine-grained video-text alignment.
Activation Manipulation Attack: Penetrating and Harmful Jailbreak Attack Against Large Vision-Language Models
Haojie Hao (Beihang University), Xianglong Liu (Beihang University)
CodeAdversarial AttackVision Language ModelImageTextMultimodality
🎯 What it does: Proposed ActMan, an activation manipulation attack framework that achieves stronger penetration and harmful destruction in large-scale vision-language models through fine-grained control of visual and language activations.
🎯 What it does: Propose Activation Layer Membrane Potential Propagation (AMP2), a spiking neural network (SNN) framework enabling training and inference within a single clock cycle.
Activations as Features: Probing LLMs for Generalizable Essay Scoring Representations
Jinwei Chi, Qiang Xu (Jinan University)
CodeClassificationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Studied and verified that intermediate layer activations of large language models can effectively distinguish essay quality through linear probes, achieving significant performance in cross-prompt automatic essay scoring tasks.
🎯 What it does: Designed a new Actor-Critic framework AC3 specifically for directly learning continuous action blocks to address long-horizon robotic manipulation tasks with sparse rewards
AdaField: Generalizable Surface Pressure Modeling with Physics-Informed Pre-training and Flow-Conditioned Adaptation
Junhong Zou (MAIS, Institute of Automation, Chinese Academy of Sciences), Xiangyu Zhu (MAIS, Institute of Automation, Chinese Academy of Sciences)
CodeAutonomous DrivingConvolutional Neural NetworkTransformerSupervised Fine-TuningPoint CloudPhysics Related
🎯 What it does: Propose the AdaField framework for predicting pressure fields on vehicle surfaces, achieving efficient transfer from pre-trained public car data to data-scarce train and aircraft scenarios.
🎯 What it does: This paper proposes a framework called ACL that adaptively fine-tunes the pre-trained model (PTM) before each incremental task to enhance plasticity in continual learning while maintaining stability.
🎯 What it does: Propose the AdaptCLIP framework, which adds three simple adapters to CLIP, achieving zero/few-shot visual anomaly detection and segmentation in cross-domain scenarios without requiring target domain fine-tuning.
🎯 What it does: Proposed ADAR, a controllable negative sampling enhancement module based on diffusion models, which generates high-quality negative samples by progressively corroding positive samples during the diffusion process, thereby improving the performance of recommendation systems.
Adaptive Dynamic Dehazing via Instruction-Driven and Task-Feedback Closed-Loop Optimization for Diverse Downstream Task Adaptation
Yafei Zhang (Kunming University of Science and Technology), Yu Liu (Hefei University of Technology)
CodeRestorationObject DetectionSegmentationDepth EstimationTransformerLarge Language ModelVision Language ModelContrastive LearningImageText
🎯 What it does: Proposed a closed-loop optimization based adaptive dynamic dehazing framework that can adjust dehazing results in real-time without retraining during inference through bidirectional guidance from task feedback and text instructions;
Adaptive Evidential Learning for Temporal-Semantic Robustness in Moment Retrieval
Haojian Huang (Hong Kong University of Science and Technology), Zhongjiang He (China Telecom)
CodeRetrievalTransformerVision Language ModelVideoText
🎯 What it does: Propose the DEMR model, introducing deep evidence regression in temporal-semantic retrieval, and significantly improving uncertainty estimation and retrieval accuracy through cross-modal fusion, query reconstruction, and geometric regularization.
🎯 What it does: Propose an adaptive hierarchical fusion framework AEF-MVC based on unsupervised evolutionary algorithms, which uses a tree structure to fuse multi-view features to enhance clustering performance.
🎯 What it does: This paper proposes an adaptive hyperbolic kernel based on curvature-adjustable de Branges-Rovnyak spaces, utilizing isometric mapping to embed hyperbolic spaces with arbitrary curvature into RKHS, and dynamically modulating hyperbolic features through learnable multipliers, further constructing adaptive hyperbolic kernels applicable to linear, polynomial, RBF, Laplacian, and the novel AHRad kernel;
🎯 What it does: Propose Adaptive Initial Residual Connections (Adaptive IRC) to alleviate the over-smoothing problem in graph neural networks and enhance their expressiveness.
Adaptive Momentum and EMA-weighted Modeling for Imbalanced Label Distribution Learning
Yongbiao Gao (Qilu University of Technology), Guohua Lv (Qilu University of Technology)
CodeClassificationOptimizationImageBenchmark
🎯 What it does: Proposed the AMEMA framework, which addresses the issues of gradient vanishing and optimization imbalance in imbalanced label distribution learning by splitting the label distribution into dominant and non-dominant branches, and applying exponential moving average (EMA) dynamic reweighting and adaptive momentum allocation on each branch.
Adaptive Morph-Patch Transformer for Aortic Vessel Segmentation
Zhenxi Zhang (Institute of Scientific Instrumentation, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Shoujun Zhou (Institute of Scientific Instrumentation, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
CodeSegmentationTransformerBiomedical Data
🎯 What it does: Proposed an adaptive Morph-Patch Transformer (MPT) for aortic vessel segmentation, addressing the issues of vascular shape distortion and insufficient semantic hierarchy caused by the fixed-size rectangular patches in traditional Transformers.
Adaptive Theory of Mind for LLM-based Multi-Agent Coordination
Chunjiang Mu (Northwestern Polytechnical University), Shuyue Hu (Shanghai Artificial Intelligence Laboratory)
CodeTransformerLarge Language ModelAgentic AITabular
🎯 What it does: Designed and implemented an adaptive Theory of Mind (A-ToM) agent that dynamically estimates and matches the ToM levels of partners using large language models, thereby enhancing zero-shot multi-agent coordination performance.
Addressing Polarization and Unfairness in Performative Prediction
Kun Jin (University of Michigan), Xueru Zhang (University of California, Santa Cruz)
CodeOptimizationData-Centric LearningImageTabular
🎯 What it does: This paper studies the impact of data distribution shifts caused by model deployment under the Performative Prediction framework on fairness, and proposes a Fair-PS solution.
Advanced Black-Box Tuning of Large Language Models with Limited API Calls
Zhikang Xie (Fudan University), Cheng Jin (Fudan University)
CodeComputational EfficiencyKnowledge DistillationLarge Language ModelText
🎯 What it does: This paper proposes a method that uses a Gaussian Process (GP) surrogate model to approximate the outputs of large language models under the premise of limited API calls, and uses this approximation to guide the training of the surrogate model, thereby achieving efficient black-box fine-tuning.
🎯 What it does: This paper proposes a training-agnostic, non-intrusive AI image source attribution method called AEDR, which uses dual autoencoder reconstruction and employs the ratio of reconstruction loss for attribution judgment.
AerialFusion: Co-Motion-Driven Unified Registration and Fusion on Multi-modal Data Streams from Aerial View
Junhui Qiu (Huazhong University of Science and Technology), Jiaqi Gui (Huazhong University of Science and Technology)
CodeConvolutional Neural NetworkGenerative Adversarial NetworkSimultaneous Localization and MappingMultimodality
🎯 What it does: Proposed a unified registration and fusion framework for multi-modal aerial visual streams based on collaborative motion, named AerialFusion, and constructed the EUM3D dataset.
AerialMind: Towards Referring Multi-Object Tracking in UAV Scenarios
Chenglizhao Chen (China University of Petroleum (East China)), Qing-Long Han (Shanghai Jiao Tong University)
CodeObject TrackingTransformerLarge Language ModelVideoTextBenchmark
🎯 What it does: Proposed the AerialMind large-scale UAV scenario referential multi-object tracking (RMOT) dataset, and developed a semi-automated collaborative annotation framework COALA and a novel tracking method HETrack
Xuesong Zhang (Hefei University of Technology), Zhenzhen Hu (Shanghai Jiao Tong University)
CodeRepresentation LearningGraph Neural NetworkTransformerReinforcement LearningVision Language ModelVision-Language-Action ModelContrastive LearningImageTextPoint Cloud
🎯 What it does: Propose a hierarchical semantic-spatial representation enhancement framework (SUSA), which improves environmental perception and instruction alignment in vision-language navigation through a text semantic view and a deep exploration graph.
AgentSense: Virtual Sensor Data Generation Using LLM Agents in Simulated Home Environments
Zikang Leng (Georgia Institute of Technology), Thomas Plötz
CodeData SynthesisLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTime SeriesRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes AgentSense, which generates structured, privacy-preserving environmental sensor data by having LLM-driven virtual agents perform diversified daily activities in an extended VirtualHome simulation environment, used for training and pretraining human activity recognition (HAR) models.
AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search
Yu Li (Tsinghua University), Fengli Xu (Tsinghua University)
CodeOptimizationNeural Architecture SearchLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Propose the AgentSwift framework to achieve automated design of LLM agents, jointly optimizing workflows and pluggable functional modules (memory, tool usage, planning), and employing a value model and uncertainty-guided hierarchical MCTS for efficient search.
AgriEval: A Comprehensive Chinese Agricultural Benchmark for Large Language Models
Lian Yan (Harbin Institute of Technology), Jingchi Jiang (Harbin Institute of Technology)
CodeTransformerLarge Language ModelTextBenchmarkAgriculture RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose AgriEval, a large-scale Chinese agricultural domain evaluation benchmark covering 6 major categories and 29 subcategories, with a total of 14,697 multiple-choice questions and 2,167 open-ended questions.
🎯 What it does: Proposes the AIR-DR framework, treating image aspect ratio adaptation as instance-level relayout to avoid pixel-level deletion, and combines adaptive decision-making with bidirectional guided background repainting to achieve high-quality, detail-preserving retargeting.
🎯 What it does: This paper proposes AirDDE, which is based on a neural delay differential equation framework, integrating multi-factor memory-enhanced attention and physics-guided delay evolution functions to achieve continuous-time air quality prediction.
🎯 What it does: This paper proposes AirWino, a high-performance Winograd convolution implementation targeting ARMv8-A processors, covering 2D and 3D convolutions, FP32/FP16 precision, 3×3 and 5×5 kernels, and achieving significant acceleration through multiple microkernels, double buffering, and custom data layouts.
🎯 What it does: Propose a human-centric adaptive AI integration framework that can switch between alignment models and complementary models based on context in human-AI collaboration to achieve a balance between alignment and complementarity;
🎯 What it does: Propose the AlignCVC framework, which jointly post-trains multi-view generation and reconstruction models, enhances cross-view consistency through distribution alignment, and compresses the 3D-aware sampling steps to 4 steps.
ALTER: Asymmetric LoRA for Token-Entropy-Guided Unlearning of LLMs
Xunlei Chen (University of Electronic Science and Technology of China), Wenhong Tian (University of Electronic Science and Technology of China)
CodeSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Propose ALTER, a lightweight LLM unlearning framework based on heterogeneous LoRA and token-entropy guided selective forgetting, achieving knowledge forgetting without modifying the base model's weights.
🎯 What it does: Proposed the Ambiguity-aware Truncated Flow Matching (ATFM) framework for simultaneously improving prediction accuracy and diversity in medical image segmentation.
🎯 What it does: This paper studies the partially labeled scenario in cross-modal retrieval, proposing a fuzzy-tolerant cross-modal hashing method called ATCH to address label ambiguity and modality alignment issues.
AMS-IO-Bench and AMS-IO-Agent: Benchmarking and Structured Reasoning for Analog and Mixed-Signal Integrated Circuit Input/Output Design
Zhishuai Zhang (Tsinghua University), Nan Sun (Princeton University)
CodeTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: This paper proposes AMS-IO-Agent, a domain-specific agent based on large language models (LLMs), designed to convert natural language design intent into structured I/O subsystem generation, while simultaneously providing the AMS-IO-Bench evaluation benchmark.
An Adaptive Configuration-Aware Simulated Annealing for the Maximally Diverse Grouping Problem
Baiyu Chen (Huazhong University of Science and Technology), Zhipeng Lü (Huazhong University of Science and Technology)
CodeOptimizationGraphBenchmark
🎯 What it does: Propose an Adaptive Configuration-Aware Simulated Annealing (ACSA) algorithm to solve the Maximum Diversity Grouping Problem (MDGP), achieving search through relaxed insertion, memory-based exchange, and vertex-level sequence coordination.
🎯 What it does: This study proposes a dataset distillation framework based on diffusion models, which generates both faithful and diverse synthetic datasets by leveraging adaptive sampling (Bayesian optimization to select CFG scale and denoising strength) and repulsion regularization, without requiring fine-tuning of the diffusion model.
🎯 What it does: Fine-tune Stable Diffusion using LoRA and achieve single-image style transfer through cross-model feature and attention injection, supporting mask-based local transfer and multi-style combinations
An Improved Privacy and Utility Analysis of Differentially Private SGD with Bounded Domain and Smooth Losses
Hao Liang (Hong Kong University of Science and Technology), Hong Xing (Hong Kong University of Science and Technology)
CodeOptimizationSafty and Privacy
🎯 What it does: This paper presents an improved analysis of privacy and utility for differential privacy stochastic gradient descent (DPSGD), particularly under bounded domains and smooth loss functions, revealing the convergence of privacy loss over iterations.
An LLM-based Quantitative Framework for Evaluating High-Stealthy Backdoor Risks in OSS Supply Chains
Zihe Yan (Shanghai Jiao Tong University), Guancheng Li (Tencent)
CodeAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper proposes a fine-grained, attacker-perspective backdoor risk assessment framework with high stealthiness, combining large language models for semantic analysis to automatically evaluate backdoor risks in open-source software supply chains, and experimentally validates it on 66 high-priority packages in the Debian ecosystem.
An LLM-based Simulation Framework for Embodied Conversational Agents in Psychological Counseling
Lixiu Wu (Tsinghua University), Jiangtao Gong (Tsinghua University)
CodeGenerationData SynthesisTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Developed an 'ECAs' framework based on a large language model to simulate embodied dialogue agents in psychological counseling and generate high-fidelity, context-rich dialogue data.
Analyzing and Mitigating Object Hallucination: A Training Bias Perspective
Yifan Li (Renmin University of China), Jirong Wen
CodeExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper systematically evaluates the object hallucination phenomenon caused by training data bias in large vision-language models by constructing the POPEv2 benchmark.
Anatomical Region-Guided Contrastive Decoding: A Plug-and-Play Strategy for Mitigating Hallucinations in Medical VLMs
Xiao Liang (Xidian University), Yuanyuan Shi (Chinese PLA General Hospital)
CodeTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningBiomedical Data
🎯 What it does: Proposed an anatomy-region guided three-layer contrastive decoding (ARCD) framework to suppress hallucinations in medical vision-language models;
Anchor Watermark: Robust Attribution for Diffusion-based Text-to-Audio Model
Xianjin Rong (Hefei University of Technology), Donghui Hu (Hefei University of Technology)
CodeGenerationData SynthesisDiffusion modelAudio
🎯 What it does: In text-to-audio diffusion models, watermarks are embedded into the initial latent vector and extracted in real-time through optimization during reverse diffusion of the generated audio, achieving audio attribution without additional training.
🎯 What it does: Designed an end-to-end graph-level clustering framework named ANGC, combining GNN encoding with learnable Nyström anchors to compute kernel similarity.
Anchor-Guided Discriminative Subspace Alignment and Clustering for Cross-Scene Hyperspectral Imagery
Yongshan Zhang (China University of Geosciences), Zhihua Cai (Wuhan University)
CodeDomain AdaptationImage
🎯 What it does: Proposed a cross-scenario hyperspectral image clustering framework (ADSAC), which obtains source scenario clustering labels through anchor-enhanced graph learning (APGL), then eliminates distribution shift by discriminative cross-scenario subspace alignment (DCSA), and infers target scenario labels using KNN in the aligned subspace.
🎯 What it does: Proposes COOL-SD, an acceleration method for autoregressive image generation based on theoretical analysis, improving the relaxation acceptance strategy and resampling distribution of Speculative Decoding;
Yuxin Jiang (Huazhong University of Science and Technology), Yunkang Cao (Hunan University)
CodeGenerationAnomaly DetectionPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: The study proposes a zero-shot anomaly generation framework that utilizes cross-modal prompts (visual + text) to drive Stable Diffusion for generating anomalous images in painting;
🎯 What it does: Propose a lightweight zero-shot anomaly generation method, AnoStyler, which generates high-quality and diverse anomaly images on a single normal image by utilizing text-guided local style transfer;
Answering the Unanswerable Is to Err Knowingly: Analyzing and Mitigating Abstention Failures in Large Reasoning Models
Yi Liu (Nanjing University), Wei Hu (Nanjing University)
CodeAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper analyzes the failure of large reasoning models to deny unsolvable problems and proposes a two-phase cognitive monitoring and intervention method during reasoning.
Anti-adversarial Learning: Desensitizing Prompts for Large Language Model
Xuan Li (Shanghai Jiao Tong University), Beijun Shen (Shanghai Jiao Tong University)
CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose PromptObfus, a method for desensitizing LLM prompts through 'oppositional adversarial learning,' replacing sensitive words to make them unrecognizable to humans while preserving the original task output.
CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningMultimodalityBenchmark
🎯 What it does: Propose the Any2RSI framework, enabling the flexible combination of multiple spatial controls for remote sensing text-to-image generation;