AAAI 2026 Papers — Page 42
AAAI Conference on Artificial Intelligence · 4149 papers
When Safe Unimodal Inputs Collide: Optimizing Reasoning Chains for Cross-Modal Safety in Multimodal Large Language Models
Wei Cai (Peking University), Xuelong Li (Institute of Artificial Intelligence, China Telecom)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed a safety reasoning path optimization framework SRPO for multimodal large language models, and constructed the SSUI dataset and RSBench benchmark to enhance cross-modal safety reasoning performance.
When Smiley Turns Hostile: Interpreting How Emojis Trigger LLMs’ Toxicity
Shiyao Cui (Tsinghua University), Minlie Huang (Tsinghua University)
Safty and PrivacyExplainability and InterpretabilityLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper constructs emoji-containing prompts (rewriting AdvBench and adversarial prompts) to systematically evaluate the effectiveness of emojis in triggering toxic generation in LLMs across multi-lingual, multi-model, and multi-attack scenarios.
When Top-ranked Recommendations Fail: Modeling Multi-Granular Negative Feedback for Explainable and Robust Video Recommendation
Siran Chen (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences), Yali Wang (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences)
Recommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelVideoTextMultimodality
🎯 What it does: Constructed a TVNF dataset containing multimodal video content, user behavior, and negative feedback reason labels, and proposed the Agentic ENF framework and S-GRPO reinforcement learning strategy for predicting and explaining user negative feedback.
When Trackers Date Fish: A Benchmark and Framework for Underwater Multiple Fish Tracking
Weiran Li (China Agricultural University), Zhenbo Li (China Agricultural University)
Object TrackingConvolutional Neural NetworkVideoBenchmark
🎯 What it does: Proposed a benchmark dataset MFT25 specifically for underwater multi-fish tracking, along with a baseline tracking framework SU-T.
When Truth Is Overridden: Uncovering the Internal Origins of Sycophancy in Large Language Models
Keyu Wang (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Studying the sycophancy behavior of large language models when users express opinions, explaining this phenomenon from the perspective of internal computational mechanisms, and analyzing the impact of opinions, professional levels, and grammatical person on sycophancy.
Where and What Matters: Sensitivity-Aware Task Vectors for Many-Shot Multimodal In-Context Learning
Ziyu Ma (Alibaba Group), Jianfei Cai (Monash University)
Computational EfficiencyRepresentation LearningMeta LearningTransformerReinforcement LearningMultimodality
🎯 What it does: Propose a task vector insertion framework STV based on perceptual sensitivity to achieve multi-sample in-context learning for multimodal large models without increasing context length.
Where It Moves, It Matters: Referring Surgical Instrument Segmentation via Motion
Meng Wei (King's College London), Nicolas Padoy (University of Strasbourg)
SegmentationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoBiomedical Data
🎯 What it does: Propose a motion-guided surgical video localization and segmentation framework called SurgRef, which can map natural language expressions to the spatiotemporal segmentation results of tools.
Where Norms and References Collide: Evaluating LLMs on Normative Reasoning
Mitchell Abrams (Tufts University), Matthias Scheutz (Tufts University)
Data SynthesisTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a diagnostic test set named SNIC to evaluate the reasoning capabilities of large language models (LLMs) in social norm-driven reference resolution; it is programmatically expanded from 51 human-validated scenarios to 9,000 instances, focusing on norms in physical environments (e.g., cleanliness, service, and use of clean items).
Where to Start Alignment? Diffusion Large Language Model May Demand a Distinct Position
Zhixin Xie (Nanyang Technological University), Jun Luo (Nanyang Technological University)
Safty and PrivacyLarge Language ModelReinforcement LearningDiffusion modelContrastive LearningText
🎯 What it does: This paper presents the first systematic analysis of the security of diffusion-based large language models (dLLM) and proposes a security alignment method called MOSA based on intermediate tokens, significantly reducing the success rates of various black-box jailbreaking attacks while maintaining the model's practicality on conventional tasks.
WhisperDiari: A Whisper-Based Speaker Diarization Framework in Token Space Leveraging Semantic and Speaker Information for Better Text Adaptability
Yongkang Yin (Peking University), Yuexian Zou (Peking University)
RecognitionTransformerContrastive LearningTextAudio
🎯 What it does: Propose the WhisperDiari framework, achieving simultaneous speaker separation and ASR in the token space, supporting synchronized generation of speaker labels and text.
Whispering Agents: A Event-Driven Covert Communication Protocol for the Internet of Agents
Kaibo Huang (Beijing University of Posts and Telecommunications), Linna Zhou (Beijing University of Posts and Telecommunications)
Safty and PrivacyLarge Language ModelText
🎯 What it does: Proposed an event-driven covert communication protocol called Pi-CCAP for the Internet of Agents (IoA), and presented a unified covert event channel model (storage, timing, and behavior dimensions);
Who Is Helping Whom? Analyzing Inter-Dependencies to Evaluate Cooperation in Human-AI Teaming
Upasana Biswas (Arizona State University), Subbarao Kambhampati (Arizona State University)
Reinforcement LearningAgentic AI
🎯 What it does: This work proposes a constructive mutual dependency metric to evaluate the collaboration of human-robot teams, and conducts user experiments with SOTA zero-shot cooperative agents in the Overcooked environment.
Who Should I Trust? Explicit Confidence-Focused Multimodal Intent Recognition
Yi Liu (Xinjiang University), Lanlan Lu (Xinjiang University)
RecognitionExplainability and InterpretabilityTransformerVideoTextPoint CloudAudio
🎯 What it does: This paper proposes an explicit confidence attention-based multimodal intent recognition framework called ECFMIR, which uses CLens to estimate confidence for each modality and cross-modal features and then weight them for fusion.
Whole-Body Coordination for Dynamic Object Grasping with Legged Manipulators
Qiwei Liang (Shenzhen MSU-BIT University), Runhao Zeng (Chinese University of Hong Kong)
Knowledge DistillationRobotic IntelligenceTransformerReinforcement LearningImageSequentialBenchmark
🎯 What it does: Designed a specialized benchmark, DQ-Bench, for whole-body coordinated grasping of quadruped robots in dynamic environments, and proposed the DQ-Net framework, achieving an end-to-end closed-loop grasping pipeline from perception to control.
Whole-Field Action Sensing via Wearable Single-Channel EMG Sensors and Resource-Efficient Motion Network
Xuanming Jiang (Xi'an Jiaotong University), Guoshuai Zhao (Xi'an Jiaotong University)
RecognitionComputational EfficiencyConvolutional Neural NetworkTime SeriesBiomedical Data
🎯 What it does: Developed a wireless single-channel sparse EMG sensor (2SEMG) and a lightweight time-frequency decoupling network OMONet, enabling full-scenario action concurrency perception for multiple people.
Why Do Open-Source LLMs Struggle with Data Analysis? A Systematic Empirical Study
Yuqi Zhu (Zhejiang University), Huajun Chen (Zhejiang University)
Data SynthesisAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTabularBenchmarkChain-of-Thought
🎯 What it does: This paper systematically evaluates the three core capabilities of open-source LLMs in data analysis tasks and proposes a data synthesis framework based on moderate-length dialogues, moderately difficult instances, and concise reasoning summaries to significantly enhance the model's analytical reasoning performance.
Wi-CBR: Salient-aware Adaptive WiFi Sensing for Cross-domain Behavior Recognition
Ruobei Zhang (Hefei University Of Technology), Jiabao Guo (Guizhou Normal University)
RecognitionDomain AdaptationConvolutional Neural NetworkTransformerContrastive LearningTime Series
🎯 What it does: Proposed the Wi-CBR framework, which fuses WiFi phase and DFS signals through dual-branch self-attention and salience-guided modules to achieve cross-domain behavior recognition.
WIET: Harmonizing Group-aware Model Weighting and Worker Allocation for Ensemble Temporal Prediction MaaS
Binbin Feng (Tongji University), Zhijun Ding (Donghua University)
OptimizationTime SeriesFinance Related
🎯 What it does: Designed and implemented the WIET system for coordinated optimization of model weighting and worker allocation in Ensemble Temporal Prediction Model-as-a-Service (ETPMaaS), thereby improving the accuracy of time series prediction and system resource utilization.
WikiMAG: A Multi-Agent Guided Framework for Generating Structured Wikipedia-like Articles
Xiuli Kang (Hebei University of Engineering), Guotong Geng (Center of Information Research, AMS)
GenerationTransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Proposed a multi-agent collaborative framework called WikiMAG for generating structured Wikipedia-like articles, supporting three content types: narrative, timeline, and table;
WikiREVIEW: A Multi-Perspective Review Framework for Automatic Wiki-Style Article Generation
Guo-Biao Zhang (Beijing Institute Of Technology), Xian-Ling Mao (Beijing Institute Of Technology)
GenerationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes the WikiREVIEW framework, which continuously revises Wikipedia-style articles through a multi-perspective review loop to achieve higher quality long-text writing.
WinMamba: Multi-Scale Shifted Windows in State Space Model for 3D Object Detection
Longhui Zheng (Xiamen University), Chenglu Wen (Xiamen University)
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: Proposed the WinMamba framework, which achieves efficient feature encoding for 3D object detection based on Mamba's multi-scale window strategy.
WorldAgen: Unified State-Action Prediction with Test-Time World Model Training
Chi Wan (Northwestern University), Manling Li (Northwestern University)
Robotic IntelligenceTransformerVision-Language-Action ModelWorld ModelMultimodalityBenchmark
🎯 What it does: Unified the Vision-Language-Action framework WorldAgen, which learns environmental dynamics (world model) and predicts task-related actions, and adapts to new environments during inference through test-time training with short-term exploration trajectories.
WorldGrow: Generating Infinite 3D World
Sikuang Li (Shanghai Jiao Tong University), Qi Tian (Huawei Inc)
GenerationDiffusion modelMesh
🎯 What it does: Designed the WorldGrow framework to achieve block-level generation and expansion in an unbounded 3D world.
WorldRFT: Latent World Model Planning with Reinforcement Fine-Tuning for Autonomous Driving
Pengxuan Yang (State Key Laboratory Of Multimodal Artificial Intelligence Systems Institute Of Automation Chinese Academy Of Sciences), Qichao Zhang (Li Auto)
Autonomous DrivingReinforcement LearningWorld ModelImagePoint Cloud
🎯 What it does: Proposes WorldRFT, a planning-oriented potential world model framework that integrates spatial perception encoding, hierarchical planning refinement, and reinforcement learning fine-tuning to enhance the safety and accuracy of end-to-end autonomous driving.
WRitEer: A Multi-Objective, Preference-Driven Multi-Agent Framework for Human-Like Advanced Text Generation
Junchuan Yu (Tianjin University), Yuyang Sun (Huazhong Agricultural University)
GenerationOptimizationLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the WRitEer framework, enabling interactive iterative writing through three agents (Reader, Editor, Writer), enhancing text naturalness and emotional expression.
X-MoGen: Unified Motion Generation Across Humans and Animals
Xuan Wang (Zhejiang University), Gaoang Wang (Zhejiang University)
GenerationGraph Neural NetworkTransformerDiffusion modelAuto EncoderMultimodality
🎯 What it does: Propose a unified framework X-MoGen that can generate 3D motion sequences of both humans and animals from natural language text.
X-ReID: Multi-granularity Information Interaction for Video-Based Visible-Infrared Person Re-Identification
Chenyang Yu, Huchuan Lu (Dalian University Of Technology)
RetrievalTransformerVision Language ModelContrastive LearningVideoMultimodality
🎯 What it does: Study visible-infrared person re-identification in videos, propose the X-ReID framework, and achieve new state-of-the-art results on two public datasets.
X-SAM: From Segment Anything to Any Segmentation
Hao Wang (Sun Yat-sen University), Xiaodan Liang (Sun Yat-sen University)
SegmentationTransformerLarge Language ModelSupervised Fine-TuningMultimodality
🎯 What it does: Propose the X-SAM framework, extending the Segment Anything Model into a unified 'from segment anything to any segmentation' model that supports dual queries of text and vision.
X2Edit: Revisiting Arbitrary-Instruction Image Editing Through Self-Constructed Data and Task-Aware Representation Learning
Jian Ma (OPPO AI Center), Haonan Lu (OPPO AI Center)
GenerationTransformerMixture of ExpertsVision Language ModelDiffusion modelContrastive LearningImageMultimodalityBenchmark
🎯 What it does: Constructed the X2Edit dataset, which scales up to 3.7M and covers 14 editing tasks, and developed a lightweight, plug-and-play arbitrary instruction image editing model X2Edit based on FLUX.1.
XLinear: A Lightweight and Accurate MLP-Based Model for Long-Term Time Series Forecasting with Exogenous Inputs
Xinyang Chen (Huazhong Agricultural University), Zaiwen Feng (Huazhong Agricultural University)
Computational EfficiencyTime SeriesBenchmark
🎯 What it does: Propose a lightweight MLP model called XLinear, which leverages learnable global tokens and time/variable gating mechanisms to integrate endogenous sequences with exogenous drivers for long-term forecasting.
xMHashSeg: Cross-modal Hash Learning for Training-free Unsupervised LiDAR Semantic Segmentation
Jialong Zhang, Yanyun Qu (Xiamen University)
SegmentationDepth EstimationDomain AdaptationAutonomous DrivingOptimizationTransformerContrastive LearningImageMultimodalityPoint Cloud
🎯 What it does: Propose a training-free cross-modal LiDAR semantic segmentation framework named xMHashSeg, which leverages a base model and non-parametric network to extract features from 2D images, depth maps, and 3D point clouds, and achieves unlabeled, no-additional-training point cloud semantic segmentation through hash learning for cross-modal feature fusion.
YOLO-IOD: Towards Real Time Incremental Object Detection
Shizhou Zhang, Yanning Zhang (Northwestern Polytechnical University)
Object DetectionComputational EfficiencyKnowledge DistillationMeta LearningConvolutional Neural NetworkImageBenchmark
🎯 What it does: Propose a YOLO-IOD framework for real-time incremental object detection to address the catastrophic forgetting problem in YOLO models during incremental learning.
You Don’t Need Pre-Built Graphs for RAG: Retrieval Augmented Generation with Adaptive Reasoning Structures
Shengyuan Chen (Hong Kong Polytechnic University), Xiao Huang (Hong Kong Polytechnic University)
RetrievalComputational EfficiencyTransformerTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the LogicRAG framework, which dynamically constructs a query-dependent directed acyclic graph (DAG) during inference and performs adaptive retrieval, eliminating the need for pre-built graphs.
You Only Need One Stage: Novel-View Synthesis from a Single Blind Face Image
Taoyue Wang (State University of New York at Binghamton), Lijun Yin (State University of New York at Binghamton)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: Propose a one-stage end-to-end method NVB-Face, which directly generates multi-view high-quality face images from a single blind (low-quality) face image;
Your AI-Generated Image Detector Can Secretly Achieve SOTA Accuracy, If Calibrated
Muli Yang (Institute for Infocomm Research), Hongyuan Zhu (Institute for Infocomm Research)
ClassificationDomain AdaptationAnomaly DetectionImageBenchmark
🎯 What it does: This paper investigates the issue of false image misjudgment caused by distribution drift in AI-generated image detection, proposing a post-scalar calibration method to dynamically adjust the decision threshold and restore detection capabilities for new generative models.
Your Prompts Are Not Safe: Output-Free Membership Inference via Prompt Vectors in Vision-Language Tuning
Yuran Bian (Shanghai Jiao Tong University), Li Pan (Shanghai Jiao Tong University)
Safty and PrivacyRepresentation LearningConvolutional Neural NetworkTransformerPrompt EngineeringVision Language ModelContrastive LearningImageText
🎯 What it does: The study proposes a prompt vector-based output-free membership inference attack (PIPRA), which identifies whether a sample was used for prompt tuning by analyzing distribution differences between prompt vectors and samples in a cross-modal semantic space.
Yours or Mine? Overwriting Attacks Against Neural Audio Watermarking
Lingfeng Yao (University of Houston), Miao Pan (University of Hawaii)
Adversarial AttackAudio
🎯 What it does: This paper studies cover attacks in neural audio watermarking systems and conducts systematic evaluations under white-box, gray-box, and black-box threat models.
ZeRCP: Towards Communication-Efficient Collaborative Perception and Future Scene Prediction via Request-Free Spatial Filtering
Yijie Chen (Hong Kong University of Science and Technology), Xinhu Zheng (Hong Kong University of Science and Technology)
Autonomous DrivingComputational EfficiencyTransformerOptical FlowImage
🎯 What it does: Propose the ZeRCP framework, integrating collaborative perception with future scene prediction to achieve communication-efficient multi-vehicle collaborative perception and prediction.
Zero-Reference Joint Low-Light Enhancement and Deblurring via Visual Autoregressive Modeling with VLM-Derived Modulation
Wei Dong (McMaster University), Jun Chen (McMaster University)
RestorationVision Language ModelImage
🎯 What it does: Proposed a fully unsupervised generative framework called VAR-LIDE, which combines a visual autoregressive model with a visual language model (VLM) prior to achieve joint recovery of low-light enhancement and deblurring.
Zero-shot Implicit Neural Manifold Representation (INMR) for Ultra-high Temporal Resolution Dynamic MRI
Jie Feng (Shanghai Jiao Tong University), Hongjiang Wei (Shanghai Jiao Tong University)
RestorationBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposed a zero-shot unsupervised implicit neural manifold representation (INMR) to achieve dynamic MRI reconstruction with extremely high spatiotemporal resolution.
Zero-Shot Open-Vocabulary Human Motion Grounding with Test-Time Training
Yunjiao Zhou (Nanyang Technological University), Jianfei Yang (Nanyang Technological University)
SegmentationRetrievalLarge Language ModelVision-Language-Action ModelVideoText
🎯 What it does: Propose a zero-shot, annotation-free open-vocabulary human action segmentation framework called ZOMG, which leverages a large language model to split text sub-actions and achieves instance-level temporal segmentation on a pre-trained encoder through soft mask optimization.
Zero-shot Recommendation: Towards Class Semantic Relation Learning for Inferring Labels of Unseen Micro-videos
Junyang Chen (Shenzhen University), Junkai Ji (Shenzhen University)
Recommendation SystemGraph Neural NetworkLarge Language ModelAuto EncoderVideo
🎯 What it does: Proposed a zero-shot recommendation framework named CSRL, which automatically predicts labels for unlabelled micro-videos and recommends user interests.
Zero-Shot Robotic Manipulation via 3D Gaussian Splatting-Enhanced Multimodal Retrieval-Augmented Generation
Zilong Xie (East China Normal University), Yuan Xie (East China Normal University)
RetrievalRobotic IntelligenceTransformerVision-Language-Action ModelGaussian SplattingImageVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a zero-shot robotic manipulation framework RobMRAG based on multi-modal retrieval-augmented generation (MRAG), which leverages multi-source knowledge bases for retrieval and achieves precise pose alignment through 3D Gaussian Splatting.
Zero-to-Hero: Empowering Video Appearance Transfer with Zero-Shot Initialization and Holistic Restoration
Tongtong Su (Zhejiang University), Dongming Lu (Alibaba Cloud Computing)
Image TranslationRestorationGenerationTransformerDiffusion modelImageVideo
🎯 What it does: Propose a two-stage reference-based video appearance editing framework called Zero-to-Hero: first, edit the anchor frame into a reference image using a zero-shot approach, then propagate the appearance consistently to all frames, and in the second stage, use a conditional generation model to restore distortions caused by the zero-shot propagation.
ZipLJP: Zipped Information Processor for Legal Judgment Prediction
Fanghao Lou (Nankai University), Huijia Li (Nankai University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a legal knowledge-based text compression method called ZipLJP for legal judgment prediction on LLMs.
Zo3T: Zero-Shot 3D-Aware Trajectory-Guided Image-to-Video Generation via Test-Time Training
Ruicheng Zhang (Tsinghua University), Xiu Li (Tsinghua University)
GenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: Developed Zo3T, a zero-shot trajectory-guided image-to-video generation framework;
Zooming In on Fakes: A Novel Dataset for Localized AI-Generated Image Detection with Forgery Amplification Approach
Lvpan Cai (Xiamen University), Xiaoshuai Sun (Tencent)
SegmentationGenerationData SynthesisAnomaly DetectionTransformerPrompt EngineeringDiffusion modelGenerative Adversarial NetworkContrastive LearningImageMultimodality
🎯 What it does: Proposed a large-scale local forgery image dataset named BR-Gen, and designed a noise-guided amplification attention visual Transformer named NFA-ViT for detecting and locating fine-grained forgeries.
Δt-Mamba3D: A Time‑Aware Spatio‑Temporal State‑Space Model for Breast Cancer Risk Prediction
Zhengbo Zhou (University of Pittsburgh), Shandong Wu (University of Pittsburgh)
ClassificationConvolutional Neural NetworkImageTime SeriesSequential
🎯 What it does: Propose the Time-Aware ∆t-Mamba3D model for long-term breast cancer risk prediction using multi-timepoint breast X-ray images.
Ψ-Arena: Interactive Assessment and Optimization of LLM-based Psychological Counselors with Tripartite Feedback
Shijing Zhu, Minlie Huang (Tsinghua University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelText
🎯 What it does: Designed and implemented an interactive evaluation and closed-loop optimization framework called ARENA to assess and improve the performance of large language models (LLMs) in psychological counseling.