AAAI 2026 Papers — Page 4
AAAI Conference on Artificial Intelligence · 4149 papers
BEE-RAG: Balanced Entropy Engineering for Retrieval-Augmented Generation
Yuhao Wang (Renmin University of China), Haifeng Wang (Baidu Inc)
GenerationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose the BEE-RAG framework, achieving robustness of retrieval-augmented generation (RAG) across different context lengths through balanced entropy engineering.
Behavior Regularization with Flow Latent Policy for Offline Reinforcement Learning
Yulong Xia (Tsinghua University), Fuchun Sun (Tsinghua University)
Reinforcement LearningFlow-based ModelBenchmark
🎯 What it does: This study proposes an offline reinforcement learning framework called Flow Latent Policy (FLP), which optimizes strategies in the latent space of a frozen flow model. By parameterizing the policy as a latent Gaussian distribution and mapping it to the action space, FLP achieves strategy optimization in the latent space and closed-form behavioral regularization.
Behavior Tokens Speak Louder: Disentangled Explainable Recommendation with Behavior Vocabulary
Xinshun Feng (Beihang University), Shuai Wang (Beihang University)
Recommendation SystemExplainability and InterpretabilityGraph Neural NetworkLarge Language ModelSupervised Fine-TuningAuto EncoderContrastive LearningTextSequential
🎯 What it does: Propose the BEAT framework, converting user-item interaction behaviors into discrete behavioral vocabulary interpretable by large language models (LLMs), enabling zero-shot explainable recommendations;
Behaviour Policy Optimization: Provably Lower Variance Return Estimates for Off-Policy Reinforcement Learning
Alex W. Goodall, Francesco Belardinelli (Imperial College London)
Reinforcement LearningBenchmark
🎯 What it does: Propose a Behavior Policy Optimization (BPO) framework that designs low-variance behavior policies to collect offline data, achieving more stable return estimation and higher sample efficiency in online reinforcement learning.
Belief-Driven Value Alignment for Human-Robot Collaboration
Saisai Li (Wuhan University of Technology), Xiao Su (Wuhan University of Technology)
Robotic IntelligenceReinforcement Learning from Human Feedback
🎯 What it does: Proposed a particle filter-based hierarchical dynamic programming algorithm (PFHDP) to address the value alignment problem in human-robot collaboration.
Benchmarking and Enhancing Rule Knowledge-Driven Reasoning of Large Language Models
Zijie Xu (Southeast University), Jing Zhou (Southeast University)
TransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose a rule knowledge-driven reasoning task, construct an automated large-scale benchmark RULER, and design the RAMPS process reward framework to enhance multi-rule reasoning.
Benchmarking LLMs for Political Science: A United Nations Perspective
Yueqing Liang (Illinois Institute Of Technology), Kai Shu (Illinois Institute Of Technology)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed the UNBench benchmark, systematically organizing and annotating draft proposals, voting records, and discussion texts from the UN Security Council between 1994-2024, covering four political decision-making tasks;
Benchmarking LLMs’ Mathematical Reasoning with Unseen Random Variables Questions
Zijin Hong (Hong Kong Polytechnic University), Xiao Huang (Beihang University)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper introduces the RV-BENCH benchmark, which evaluates the true mathematical reasoning ability of LLMs through generated random variable questions (RVQs).
Benchmarking Multimodal Knowledge Conflict for Large Multimodal Models
Yifan Jia (Shandong University), Dongrui Liu (Shanghai AI Laboratory)
Large Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the MMKC-Bench benchmark to evaluate the behavior and detection capabilities of multimodal models when facing multimodal knowledge conflicts in retrieval-augmented generation.
Benchmarking Reinforcement Learning Algorithms for ICU Ventilator Settings: An Interpretable and Probabilistic Patient Environment for Doctor Agents
Ya-Hsi Chang (National Tsing Hua University), Po-Chih Kuo (National Tsing Hua University)
Explainability and InterpretabilityReinforcement LearningTabularBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: Constructed an interpretable probabilistic ICU patient environment simulator, and conducted benchmark evaluation of seven offline reinforcement learning algorithms within this environment to optimize mechanical ventilation settings.
Benchmarking Visual LLMs Resilience to Unanswerable Questions on Visually Rich Documents
Davide Napolitano (Politecnico di Torino), Fabrizio Battiloro (Politecnico di Torino)
Large Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This study proposes the VRD-UQA evaluation framework, which automatically detects three types of corruption in multi-page visual rich documents (VRDs): entity, document elements, and layout. It generates answerable but unanswerable questions and evaluates the robustness of visual large language models (VLLMs) in detecting unanswerable questions at both page-level and document-level.
Best Arm Identification with Biased Contexts
James Cheshire (Telecom Paris), Stephan Clémençon (Telecom Paris)
Optimization
🎯 What it does: This paper proposes TACTIC, an algorithm for the Best Arm Identification problem in multi-armed bandits with contextual selection bias, which actively corrects bias by leveraging known target distribution moment information. TACTIC collects data directly under the source distribution and adaptively estimates importance weights and arm means, ultimately identifying the optimal arm in a PAC manner within a fixed confidence framework.
Best of Both Worlds Guarantees for Equitable Allocations
Umang Bhaskar (Tata Institute of Fundamental Research), Rakshitha (Indian Institute of Technology Delhi)
Optimization
🎯 What it does: This paper studies the 'best of both worlds' allocation that simultaneously satisfies ex ante fairness (ex ante EQ) and ex post fairness (ex post EQ1/EQX), providing its existence geometric conditions, computational complexity, and algorithms and limitations under different valuation models (binary, general additive).
Best-Effort Policies for Robust Markov Decision Processes
Alessandro Abate (University of Oxford), Francesco Fabiano (University of Oxford)
Reinforcement LearningTabular
🎯 What it does: Proposes the Best-Effort Robust Strategy (ORBE), providing a systematic method based on dominance and best-effort for resolving multiple optimal robust strategies in RMDP, along with complete theoretical analysis and algorithm implementation.
Better Datasets Start from RefineLab: Automatic Optimization for High-Quality Dataset Refinement
Xiaonan Luo (University Of Notre Dame), Xiangliang Zhang (Vanderbilt University)
OptimizationData-Centric LearningTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose RefineLab, an LLM-based framework for automatically improving QA datasets under budget constraints, which enhances quality dimensions such as coverage, difficulty, and authenticity by selecting edit operations through integer linear programming.
Better Matching, Less Forgetting: A Quality-Guided Matcher for Transformer-based Incremental Object Detection
Qirui Wu (Northwestern Polytechnical University), Peng Wang (Northwestern Polytechnical University)
Object DetectionTransformerImage
🎯 What it does: In incremental object detection, to address the catastrophic forgetting problem caused by the "background-foreground confusion" in DETR series models, a quality-guided minimum cost maximum flow (Q-MCMF) matcher is proposed. It can retain one-to-one matching while eliminating low IoU erroneous matches, thereby removing incorrect supervision.
BEVDilation: LiDAR-Centric Multi-Modal Fusion for 3D Object Detection
Guowen Zhang (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkGaussian SplattingImageMultimodalityPoint Cloud
🎯 What it does: Proposes a LiDAR-centric multi-modal 3D object detection framework called BEVDilation, which employs image BEV as implicit guidance to alleviate depth estimation errors, and enhances the density and semantic information by sparsifying foreground voxels through the Sparse Voxel Dilation Block (SVDB) and Semantic-Guided BEV Dilation Block (SBDB).
Beware of Reasoning Overconfidence: Pitfalls in the Reasoning Process for Multi-solution Tasks
Jiannan Guan (Harbin Institute of Technology), Wanxiang Che (Du Xiaoman Science Technology Co Ltd)
OptimizationExplainability and InterpretabilityLarge Language ModelBenchmarkChain-of-Thought
🎯 What it does: This paper investigates the overconfidence in reasoning of large language models on multi-solution tasks and proposes the MuSoBench benchmark to systematically evaluate and analyze this issue.
Beyond Accuracy: A Cognitive Load Framework for Mapping the Capability Boundaries of Tool-use Agents
Qihao Wang (Chinese Academy of Sciences), Yuanmin Tang (Chinese Academy of Sciences)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose an evaluation framework for tool-using agents based on cognitive load theory, and construct an adjustable load benchmark called ToolLoad-Bench to systematically assess LLM tool usage performance;
Beyond Adapter Retrieval: Latent Geometry-Preserving Composition via Sparse Task Projection
Pengfei Jin (Massachusetts General Hospital and Harvard Medical School), Quanzheng Li (Massachusetts General Hospital and Harvard Medical School)
SegmentationGenerationTransformerImageTextBiomedical DataMagnetic Resonance ImagingComputed TomographyRetrieval-Augmented Generation
🎯 What it does: This paper proposes a sparse geometry-preserving adapter combination framework based on task potential prototypes, generating new adapters under zero-shot settings by solving ℓ₁-regularized sparse reconstruction in the task prototype space.
Beyond Binary Classification: A Semi-supervised Approach to Generalized AI-generated Image Detection
Hong-Hanh Nguyen-Le (University College Dublin), Nhien-An Le-Khac (Trinity College Dublin)
Anomaly DetectionTransformerGenerative Adversarial NetworkContrastive LearningImageBenchmark
🎯 What it does: Propose a semi-supervised method TriDetect, combining binary classification with architectural-level clustering, leveraging Sinkhorn-Knopp to balance clustering and consistency, revealing differences between GAN and DM in latent subspaces, and enhancing cross-generator detection.
Beyond Binary Erasure: Soft-Weighted Unlearning for Fairness and Robustness
Xinbao Qiao (Zhejiang University), Meng Zhang (The Hong Kong University of Science and Technology (Guangzhou))
OptimizationImageTextTabular
🎯 What it does: Proposed the Soft-Weighted Unlearning framework, which replaces traditional hard deletion with fine-grained weights to address the practicality degradation caused by excessive unlearning.
Beyond Boundaries: Leveraging Vision Foundation Models for Source-Free Object Detection
Huizai Yao (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
Object DetectionDomain AdaptationKnowledge DistillationTransformerContrastive LearningImage
🎯 What it does: Proposes a source-agnostic object detection framework leveraging a visual foundation model (VFM), aiming to achieve cross-domain adaptation without accessing source data.
Beyond Chains: Bridging Large Language Models and Knowledge Bases in Complex Question Answering
Yihua Zhu (Kyoto University), Hidetoshi Shimodaira (Kyoto University)
RetrievalExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the PDRR framework, which first uses LLM to predict question types and decompose them into structured triplets, then retrieves KB facts through retrieval, and finally completes reasoning and generates answers via LLM; this method achieves training-free, interpretable multi-step reasoning.
Beyond Conservation: Flexible Molecular Assembly with Unbalanced Diffusion Bridge
Rongchao Zhang (Peking University), Hanpin Wang (Nanjing University)
Drug DiscoveryGraph Neural NetworkDiffusion modelGraph
🎯 What it does: Proposes a molecular assembly generation framework called AssemUDB based on unbalanced diffusion bridges, capable of efficiently sampling in molecular structure spaces ranging from weak to strong correlations;
Beyond Content: A Comprehensive Speech Toxicity Dataset and Detection Framework Incorporating Paralinguistic Cues
Zhongjie Ba (State Key Laboratory of Blockchain and Data Security, Zhejiang University), Li Lu (State Key Laboratory of Blockchain and Data Security, Zhejiang University)
ClassificationTransformerContrastive LearningMultimodalityBenchmarkAudio
🎯 What it does: Proposed the ToxiAlert-Bench speech toxicity dataset and the ToxiAlert dual-head model for simultaneous detection of toxicity categories and sources (text, acoustic, or both).
Beyond Cosine Similarity: Magnitude-Aware CLIP for No-Reference Image Quality Assessment
Zhicheng Liao (South China Normal University), Baoliang Chen (South China Normal University)
TransformerVision Language ModelContrastive LearningImageBenchmark
🎯 What it does: This paper proposes a no-reference image quality assessment method, introducing Box-Cox normalization of image embedding amplitude based on the semantic similarity of the CLIP model, and employing a confidence-guided fusion strategy to achieve more accurate quality scoring.
Beyond Counting: Evaluating Abstract and Emotional Reasoning in Vision-Language Models
Yuan Zhou (Institute of Automation, Chinese Academy of Sciences), Shiming Xiang (Institute of Automation, Chinese Academy of Sciences)
Large Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Created the EmojiGrid benchmark using procedurally synthesized emoji grid images and over 29,000 QA pairs, covering nine cognitive subtasks including perception, relational and structural reasoning, and abstract and emotional reasoning;
Beyond Euclidean Assumptions: Geometry-Aware Adaptive Routing for Remote Sensing Segmentation
Jie Qiu (Fujian Agriculture and Forestry University), Jianzhang Chen (Fujian Agriculture and Forestry University)
SegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: This work proposes the Geometry-Aware Adaptive Routing (GAAR) module, which dynamically assigns high-level features to Euclidean or hyperbolic space to enhance the geometric representation capability in remote sensing image semantic segmentation.
Beyond Fact Retrieval: Episodic Memory for RAG with Generative Semantic Workspaces
Shreyas Rajesh (University of California), Vwani Roychowdhury (University of California)
RetrievalComputational EfficiencyTransformerPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the Generative Semantic Workspace (GSW) framework, utilizing Operator and Reconciler modules to construct an updatable episodic memory, enabling LLMs to achieve entity tracking and event reasoning in long narratives.
Beyond Fixed Depth: Adaptive Graph Neural Networks for Node Classification Under Varying Homophily
Asela Hevapathige (Australian National University), Ahad N. Zehmakan (Australian National University)
ClassificationGraph Neural NetworkGraph
🎯 What it does: This paper proposes an Adaptive Depth Graph Neural Network (AD-GNN) based on node-level adaptive aggregation depth, which can simultaneously apply to both homogeneous and heterogeneous graphs within the same model.
Beyond Fixed Tasks: Unsupervised Environment Design for Task-Level Pairs
Daniel Furelos-Blanco (Imperial College London), Michael Dennis (Google DeepMind)
OptimizationGraph Neural NetworkReinforcement Learning
🎯 What it does: Investigated a co-adaptive curriculum generation method ATLAS based on unsupervised environment design, aiming to simultaneously optimize tasks and environments to enhance the robustness of general agents in scenarios with scarce solvable task-environment pairs.
Beyond Fully Supervised Pixel Annotations: Scribble-Driven Weakly-Supervised Framework for Image Manipulation Localization
Songlin Li (Xinjiang University), Gaobo Yang (Hefei University of Technology)
SegmentationTransformerImageBenchmark
🎯 What it does: Develop a scribble-based weakly supervised image tampering localization method and release the first scribble-annotated Sc-IML dataset
Beyond Graph Priors: A Co-Evolving Framework Under Uncertainty for Enterprise Resilience Assessment
Yanzhe Xie (Southwestern University of Finance and Economics), Kunpeng Zhang (Sichuan University)
Graph Neural NetworkContrastive LearningGraphFinance Related
🎯 What it does: Propose the CFU framework, achieving enterprise resilience assessment through co-evolving structural and representational aspects.
Beyond Illumination: Fine-Grained Detail Preservation in Extreme Dark Image Restoration
Tongshun Zhang (Jilin University), Qiuzhan Zhou (Jilin University)
RestorationImage
🎯 What it does: Propose a two-stage method that first recovers global illumination in the frequency domain through the Residual Fourier-Guided Module (RFGM), and then refines textures and edges in the spatial domain using Patch Mamba and Grad Mamba, focusing on fine-grained detail restoration in extremely dark images.
Beyond Immediate Activation: Temporally Decoupled Backdoor Attacks on Time Series Forecasting
Zhixin Liu (Nankai University), Xiangrui Cai (Nankai University)
Adversarial AttackGraph Neural NetworkTime Series
🎯 What it does: Designed and implemented the TDBA framework to achieve spatiotemporal decoupling backdoor attacks in multivariate time series prediction, enabling triggers to be flexibly activated across different dimensions and time delays;
Beyond Local Patterns: Multiscale Inconsistency Learning for Graph Anomaly Detection
Jie Lian (Zhejiang University), Haishuai Wang (Zhejiang University)
Anomaly DetectionGraph
🎯 What it does: Propose an unsupervised graph anomaly detection framework called MI-GAD that utilizes multi-scale inconsistency learning to simultaneously model local similarity and global distribution differences.
Beyond Missing Data Imputation: Information-Theoretic Coupling of Missingness and Class Imbalance for Optimal Irregular Time Series Classification
Xin Qin (Tianjin University of Technology), Xu Cheng (Tianjin University of Technology)
ClassificationConvolutional Neural NetworkRecurrent Neural NetworkTransformerTime SeriesBenchmark
🎯 What it does: Proposes the SPECTRA framework, jointly modeling missing patterns and class imbalance in irregular time series, providing an information-theoretic perspective on the 'missing-imbalance coupling' theory;
Beyond Monotonicity: Revisiting Factorization Principles in Multi-Agent Q-Learning
Tianmeng Hu (University of Exeter), Ke Li (University of Exeter)
Reinforcement Learning
🎯 What it does: This paper proposes removing the monotonicity constraints of traditional methods such as QMIX, adopting non-monotonic value decomposition, and proving through dynamic analysis of continuous-time gradient flows that under approximate greedy exploration, the zero-loss points of IGM inconsistency are unstable saddle points, while IGM-consistent points are stable attractors; experiments verify that this method can recover the optimal IGM solutions and outperform traditional baselines in multi-agent environments such as matrix games, SMAC, and GRF.
Beyond MSE: Ordinal Cross-Entropy for Probabilistic Time Series Forecasting
Jieting Wang (Shanxi University), Xiaolei Shang (Shanxi University)
TransformerTime Series
🎯 What it does: Convert time series regression into an ordinal classification problem, using ordinal cross-entropy (OCE) to train a probabilistic distribution model, which can quantify uncertainty while maintaining robustness.
Beyond N-grams: A Hierarchical Reward Learning Framework for Clinically-Aware Medical Report Generation
Yuan Wang, Zuozhu Liu (Zhejiang University)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningBiomedical DataElectronic Health Records
🎯 What it does: Propose HiMed-RL, a hierarchical reward learning framework specifically designed for medical report generation;
Beyond Next Token Probabilities: Learnable, Fast Detection of Hallucinations and Data Contamination on LLM Output Distributions
Guy Bar-Shalom (Technion), Haggai Maron (Technion)
Anomaly DetectionTransformerLarge Language ModelText
🎯 What it does: Proposes utilizing the complete output distribution of LLMs (including the probability distribution of the next token and actual token probabilities) to detect hallucinations and data contamination, and designs a lightweight Transformer architecture, LOS-NET, for learning.
Beyond Observations: Reconstruction Error-Guided Irregularly Sampled Time Series Representation Learning
Jiexi Liu (Nanjing University of Finance and Economics), Songcan Chen (Nanjing University of Aeronautics and Astronautics)
ClassificationRepresentation LearningAuto EncoderContrastive LearningTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: Propose the iTimER framework, which achieves self-supervised pre-training on irregularly sampled time series by generating pseudo observations through leveraging the reconstruction error distribution.
Beyond Passive Critical Thinking: Fostering Proactive Questioning to Enhance Human-AI Collaboration
Ante Wang (Xiamen University), Jinsong Su (Baidu Inc.)
Large Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: Propose a Proactive Critical Thinking paradigm, enabling large language models to actively ask questions to obtain missing information when encountering incomplete or incorrect problems, thereby achieving better collaboration and reasoning;
Beyond Perplexity: Let the Reader Select Retrieval Summaries via Spectrum Projection Score
Zhanghao Hu (Alan Turing Institute), Lin Gui (Alan Turing Institute)
RetrievalCompressionTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the Spectrum Projection Score (SPS), a metric to evaluate the semantic alignment between retrieved summaries and reader LLMs, and develop the xCompress framework to dynamically sample, sort, and compress summaries;
Beyond Plain Demos: A Demo-Centric Anchoring Paradigm for In-Context Learning in Alzheimer’s Disease Detection
Puzhen Su (National University of Defense Technology), Ting Wang (National University of Defense Technology)
ClassificationTransformerContrastive LearningTextBiomedical DataAlzheimer's DiseaseRetrieval-Augmented Generation
🎯 What it does: Propose a demo-centric anchoring framework DA4ICL, which improves Alzheimer's disease detection in-context learning by utilizing multi-dimensional retrieval and projection vector anchoring;
Beyond Predictive Resampling: Learning Input-Agnostic Downsampling for Efficient Aligned Vision Recognition
Kai Zhao (Shanghai University), Dan Zeng (Jiaxing University)
RecognitionComputational EfficiencyFlow-based ModelImage
🎯 What it does: Propose a learning-based downsampling paradigm for aligned visual recognition, achieving non-uniform downsampling through learning input-agnostic flow fields;
Beyond Quadratic: Linear-Time Change Detection with RWKV
Zhenyu Yang (Nanjing University of Science and Technology), Yazhou Yao (Nanjing University of Science and Technology)
SegmentationComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: This study proposes a linear-time remote sensing change detection framework called ChangeRWKV based on RWKV, which employs a hierarchical encoder and a spatial-temporal fusion module to achieve efficient multi-scale feature extraction and difference determination.
Beyond ReAct: A Planner-Centric Framework for Complex Tool-Augmented LLM Reasoning
Xiaolong Wei (Beihang University), Dawei Yin (Baidu Inc)
OptimizationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Proposed a planner-based global DAG planning and execution framework for addressing complex LLM tasks involving multiple tool calls.
Beyond Retraining: Training-Free Unknown Class Filtering for Source-Free Open Set Domain Adaptation of Vision–Language Models
Yongguang Li (Jilin University), Menglin Yang (Hong Kong University of Science and Technology)
Domain AdaptationVision Language ModelMultimodality
🎯 What it does: Developed a fully training-free, annotation-free inference method called VLM-OpenXpert, which utilizes SUFF (SVD low-rank filtering) and BGAT (Box-Cox + GMM adaptive thresholding) to filter unknown classes and correct thresholds in open-domain adaptation scenarios involving source-agnostic, unknown category emergence for vision-language models;
Beyond Semantic Features: Pixel-level Mapping for Generalized AI-Generated Image Detection
Chenming Zhou (Institute of Computing Technology, Chinese Academy of Sciences), Sheng Tang (Institute of Computing Technology, Chinese Academy of Sciences)
Data SynthesisAnomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: Propose a pixel-level mapping preprocessing method that suppresses low-frequency semantic information by disrupting the monotonic arrangement of pixel values, thereby enhancing the cross-model generalization capability of AI-generated image detectors on unknown generative models.
Beyond Sharpness: A Flatness Decomposition Framework for Efficient Continual Learning
Yanan Chen (Xi'an Jiaotong University), Wen Wen (China Telecom)
OptimizationComputational EfficiencyImage
🎯 What it does: Propose an FLAD framework that first decomposes sharpness-aware perturbations into gradient alignment and random noise components, regularizing only the random noise to achieve flatter minima in continual learning with lower computational costs;
Beyond Sharpness: The Role of Nonuniformity in Generalization
Yingcong Zhou (Northeast Normal University), Fengqin Yang (Northeast Normal University)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes and investigates the impact of gradient noise non-uniformity on the generalization of deep networks, demonstrating its predictive power for generalization and proposing an optimization algorithm based on this metric.
Beyond Simple Edits: X-Planner for Complex Instruction-Based Image Editing
Chun-Hsiao Yeh (UC Berkeley), Krishna Kumar Singh (Adobe Research)
Image TranslationTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: X-Planner designs a multi-modal large language model (LLM)-driven planning system that decomposes complex image editing instructions into sub-instructions using chain-of-thought reasoning, and automatically generates precise masks and bounding boxes, achieving high-quality editing without requiring manual masks or bounding boxes;
Beyond Single Transactions: D-EMAML---Dual-Edge Motif Neural Networks for Enhanced Anti-Money Laundering Detection
Dongmei Han (Shanghai University of Finance and Economics), Xiaofeng Zhou (Shanghai University of Finance and Economics)
Anomaly DetectionGraph Neural NetworkGraphFinance Related
🎯 What it does: Constructed a dual-edge motif graph and utilized GNN for edge-level anti-money laundering (AML) anomaly detection
Beyond Single-Point Perturbation: A Hierarchical, Manifold-Aware Approach to Diffusion Attacks
Zhijie Wang (Zhejiang University), Cong Wang (Zhejiang University)
Adversarial AttackDiffusion modelImage
🎯 What it does: By introducing hierarchical multi-step perturbations in the diffusion generation process, combined with unprompted internal self-attention guidance and dynamic manifold alignment, high-fidelity and high-transferability adversarial examples are generated.
Beyond Single-Speed Reasoning: Coordinating Fast and Slow Dynamics for Efficient World Modeling
Hongwei Wang (Beijing Jiaotong University), Yi Jin (Peking University)
Computational EfficiencyKnowledge DistillationRepresentation LearningRecurrent Neural NetworkReinforcement LearningAuto EncoderWorld ModelImageVideo
🎯 What it does: Proposed the SF-RSSM dual-branch world model, combining a fast branch with residual prediction and a slow branch with GRU, improving model-based reinforcement learning performance through cross-scale distillation and a consistency-driven curiosity module.
Beyond Single-Step Updates: Reinforcement Learning of Heuristics with Limited-Horizon Search
Gal Hadar (Ben Gurion University), Shahaf S. Shperberg (Ben Gurion University)
OptimizationReinforcement LearningGraph
🎯 What it does: Proposed the Limited Horizon Bellman-based Learning (LHBL) method, which updates the heuristic function using finite-depth search and improves state sampling.
Beyond Static: Related Questions Retrieval Through Conversations in Community Question Answering
Xiao Ao (University of Electronic Science and Technology of China), Weikang Guo (Southwestern University of Finance and Economics)
RetrievalTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes a label-enhanced conversational question retrieval model called TeCQR, which dynamically refines query representations using clarification questions and user feedback, significantly improving the effectiveness of relevant question retrieval on community Q&A platforms.
Beyond Step Pruning: Information Theory Based Step-level Optimization for Self-Refining Large Language Models
Jinman Zhao (University of Toronto), Gerald Penn (University of Toronto)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackSupervised Fine-TuningText
🎯 What it does: Built the ISLA framework, which directly generates step-level preference data using SFT golden answers and performs step-level DPO fine-tuning by self-selecting important reasoning steps through information-theoretic measures.
Beyond Superficial Forgetting: Thorough Unlearning Through Knowledge Density Estimation and Block Re-Insertion
Feng Guo (University of Electronic Science and Technology of China), Shuo Shang (University of Electronic Science and Technology of China)
Safty and PrivacyComputational EfficiencyTransformerTextBenchmark
🎯 What it does: Propose the KUnBR framework, which utilizes knowledge density estimation to locate layers containing harmful knowledge and achieves deep machine forgetting through a block-level reinsertion strategy.
Beyond the Horizon: Decoupling Multi-View UAV Action Recognition via Partial Order Transfer
Wenxuan Liu (Peking University), Chia-Wen Lin (National Tsing Hua University)
RecognitionDomain AdaptationConvolutional Neural NetworkVideo
🎯 What it does: Investigated the height differences in UAV-based multi-view action recognition, proposing the Aero Partial Order Guided Network (Aerorder) to separate action features from viewpoint features and leverage the partial order relationship of height for knowledge transfer.
Beyond the Lower Bound: Bridging Regret Minimization and Best Arm Identification in Lexicographic Bandits
Bo Xue (City University of Hong Kong), Qingfu Zhang (Zhejiang University)
OptimizationReinforcement LearningTabular
🎯 What it does: Proposes two elimination algorithms, LexElim-Out and LexElim-In, to uniformly address the regret minimization and best arm identification problems in lexicographic multi-objective bandits, along with theoretical upper bounds and experimental validation.
Beyond the Mean: Fisher-Orthogonal Projection for Natural Gradient Descent in Large Batch Training
Yishun Lu (University of Oxford), Wesley Armour (University of Oxford)
OptimizationImage
🎯 What it does: Propose the Fisher-Orthogonal Projection (FOP) optimizer, incorporating geometry-aware variance correction into natural gradient updates, enabling training with extremely large batches.
Beyond Tokens: Dynamic Latent Reasoning via Semantic Residual Refinement
Fangrui Lv (Tsinghua University), Changshui Zhang (Tsinghua University)
Computational EfficiencyTextChain-of-Thought
🎯 What it does: Proposes DyLaR, an untrained dynamic implicit reasoning framework that enables large language models to perform reasoning in a continuous hidden layer space.
Beyond Training-time Poisoning: Component-level and Post-training Backdoors in Deep Reinforcement Learning
Sanyam Vyas (Cardiff University), Vasilios Mavroudis (Alan Turing Institute)
Adversarial AttackReinforcement LearningVideo
🎯 What it does: Proposes two novel deep reinforcement learning backdoor attacks: component-level TrojanentRL and post-training attack InfrectroRL that does not require a training pipeline;
Beyond Wide-Angle Images: Structure-to-Detail Video Portrait Correction via Unsupervised Spatiotemporal Adaptation
Wenbo Nie (Beijing Jiaotong University), Kang Liao
RestorationTransformerDiffusion modelOptical FlowImageVideo
🎯 What it does: Designed a structure-to-detail image correction model called ImagePC, and extended it to VideoPC for unlabeled videos, achieving wide-angle portrait correction and video stabilization;
Beyond World Models: Rethinking Understanding in AI Models
Tarun Gupta (Indian Institute of Science), Danish Pruthi (Indian Institute of Science)
Explainability and InterpretabilityWorld Model
🎯 What it does: Critique the world model framework in AI for its inability to sufficiently characterize human-level understanding through philosophical case studies (e.g., domino computers, mathematical proofs, Bohr's atomic theory).
BeyondSparse: Facilitating Mamba to Enhance Cross-Domain 3D Semantic Segmentation in Adverse Weather
Yao Wu, Yanyun Qu (Hanjiang National Laboratory)
SegmentationDomain AdaptationConvolutional Neural NetworkPoint Cloud
🎯 What it does: Cross-Domain 3D Semantic Segmentation in Adverse Weather Conditions
BhashaKritika: Building Synthetic Pretraining Data at Scale for Indic Languages
Guduru Manoj (Krutrim), Shubham Agarwal (Krutrim)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Built and released a multilingual Indian synthetic pre-training corpus named BhashaKritika, containing 540B words, and designed an end-to-end pipeline for multi-technique generation and quality assessment.
Bi-level Personalization for Federated Foundation Models: A Task-vector Aggregation Approach
Yiyuan Yang (University of Technology Sydney), Jing Jiang (University of Technology Sydney)
Federated LearningLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: Propose FedBip, a two-layer personalized framework that performs task-specific fine-tuning on clients and uses task vectors for similarity-weighted aggregation on the server side, achieving efficient personalization and collaboration for federated foundation models.
Bi-Spectrum Distillation: Addressing Spectral Mismatch in ANN-SNN Knowledge Transfer
Yuxuan Zhang (Beihang University), Hongjue Li (Beihang University)
Knowledge DistillationSpiking Neural NetworkImage
🎯 What it does: Propose Bi-Spectrum Distillation (BSD), addressing spectral mismatch in ANN-to-SNN knowledge distillation through feature-level Spectral Residual Distillation (SRD) and logits-level Spectral Semantic Distillation (SSD).
Bi-VLM: Binary Post-Training Quantization for Vision-Language Models
Xijun Wang (University of Maryland), Dinesh Manocha (University of Maryland)
Computational EfficiencyMultimodality
🎯 What it does: Propose Bi-VLM, a binary post-training quantization method for vision-language models, which utilizes Gaussian quantiles to non-uniformly partition weights and perform mixed quantization;
Bias Association Discovery Framework for Open-Ended LLM Generations
Jinhao Pan (George Mason University), Ziwei Zhu (George Mason University)
Explainability and InterpretabilityLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes the Bias Association Discovery Framework (BADF), systematically discovering and quantifying associations between different social identities and descriptive concepts (biases) in open-ended generated text from large language models (LLMs).
Bias-Restrained Prefix Representation Finetuning for Mathematical Reasoning
Sirui Liang (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences), Kang Liu (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the Bias-Restrained Prefix Representation Finetuning (BREP) method, which significantly improves the performance of large language models (LLMs) on mathematical reasoning tasks by restricting bias magnitude, training prefixes, and intervening only during the early inference phase.
BiCA: Effective Biomedical Dense Retrieval with Citation-Aware Hard Negatives
Aarush Sinha (Vellore Institute of Technology), Nirav Pravinbhai Bhatt (Indian Institute of Technology Madras)
RetrievalTransformerGraphBiomedical DataBenchmark
🎯 What it does: Train the BiCA retrieval model by generating hard negative samples using multi-hop citation chains from PubMed.
BiCycle: Group-wise Recursive Transformer Based on ASR Mechanism
Min Ho Jang (Chung-Ang University), Ji Won Yoon (Chung-Ang University)
RecognitionTransformerAudio
🎯 What it does: Proposed the BiCycle recursive Transformer framework, achieving grouped recursion for hierarchical grouping in ASR and introducing grouped feature distillation;
Bid Farewell to Seesaw: Towards Accurate Long-Tail Session-Based Recommendation via Dual Constraints of Hybrid Intents
Xiao Wang (University of Electronic Science and Technology of China), Shuang Liang (University of Electronic Science and Technology of China)
Recommendation SystemContrastive LearningSequential
🎯 What it does: Proposed a plug-in framework HID (Hybrid Intent-based Dual Constraint Framework), which first constructs 'Hybrid Intent' and assigns target and noise intents to sessions, then incorporates dual constraints (targeting long-tail and accuracy) into session embedding learning, achieving a win-win in accuracy and diversity for long-tail recommendations.
Bidirectional Bounded-Suboptimal Heuristic Search with Consistent Heuristics
Shahaf S. Shperberg (Ben-Gurion University of the Negev), Dor Atzmon (Bar-Ilan University)
Optimization
🎯 What it does: Proposes a bidirectional heuristic search framework called WBAE* suitable for tolerable suboptimal search, which introduces a weight λ to regulate the heuristic error term d based on BAE*, and provides a series of theoretical proofs and experimental analysis.
Bidirectional Channel-selective Semantic Interaction for Semi-Supervised Medical Segmentation
Kaiwen Huang (Nanjing University of Science and Technology), Tao Zhou (Nanjing University of Science and Technology)
SegmentationConvolutional Neural NetworkTransformerBiomedical DataMagnetic Resonance ImagingComputed TomographyBenchmark
🎯 What it does: Propose a semi-supervised medical image segmentation framework BCSI based on weak-to-strong consistency learning, achieving effective interaction between labeled and unlabeled data through semantic-space perturbation, channel selection routing, and bidirectional channel interaction.
Bidirectional Counterfactual Distillation for Review-Based Recommendation
Sheng Sang (Hefei University Of Technology), Feng Xue (Hefei University Of Technology)
Recommendation SystemGraph Neural NetworkGenerative Adversarial NetworkText
🎯 What it does: This paper proposes the BiCoD framework, which improves upon issues in comment recommendation caused by knowledge transfer pollution and homogenization of behavior features due to score distribution bias.
Bidirectional Noise Injection: Enhancing Diffusion Models via Coordinated Input-Output Perturbation
Tianyi Zheng (vivo Mobile Communication Co., Ltd), Bo Li (vivo Mobile Communication Co., Ltd)
GenerationDiffusion modelImage
🎯 What it does: Proposes a Bidirectional Noise Injection (BNI) framework, specifically the Coordinated Input-Output Perturbation (CIOP), which synchronously injects noise into both the input and target of diffusion models during training to reduce the deviation between the predicted distribution and the true distribution.
BidMatch: Boosting Semi-Supervised Learning by Bi-Dimensional Sample Weight Guidance
Xianling Yang (South China University Of Technology), Kaixiang Yang (South China University Of Technology)
ClassificationRepresentation LearningImage
🎯 What it does: This paper proposes the dual-dimensional sample weight guidance algorithm BidMatch, which refines sample weights in semi-supervised learning by utilizing class information entropy and pseudo-label probability redistribution.
BiHiTo: Biomolecular Hierarchy-inspired Tokenization
Ruochong Zheng (Peking University), Jie Chen (Peking University)
Protein Structure PredictionTransformerAuto EncoderBiomedical Data
🎯 What it does: Designed a multi-level BiHiTo tokenizer that efficiently encodes and reconstructs 3D structures of proteins and nucleic acids by leveraging hierarchical biological molecular priors.
Bilevel MCTS for Amortized O(1) Node Selection in Classical Planning
Masataro Asai (MIT-IBM Watson AI Lab)
OptimizationBenchmark
🎯 What it does: Proposed Bilevel MCTS, which reduces the node selection cost of MCTS by performing limited Best-First Search at each leaf node, and introduces Tree Collapsing to further reduce tree depth; combined this method with existing heuristic search techniques (breadth-first search, alternating queue, priority operator) to build Nε bula and evaluated it on IPC competition benchmarks.
Binary Message Passing for Generalizable Semi-Supervised Graph Anomaly Detection
Jingyuan Zhang (Institute of Software Chinese Academy of Sciences), Fengjun Zhang (Institute of Software Chinese Academy of Sciences)
Anomaly DetectionGraph Neural NetworkGraph
🎯 What it does: Proposed a Binary Message Passing (BMP) framework that enhances anomaly information propagation and improves semi-supervised graph anomaly detection by constructing binary trees based on anomaly probability for message routing.
Binary Split Categorical Feature with Mean Absolute Error Criteria in CART
Peng Yu (University of Electronic Science and Technology of China), Jesse Read (Ecole Polytechnique Institut Polytechnique de Paris)
OptimizationComputational EfficiencyTabular
🎯 What it does: Studied the binary splitting of classification features under the MAE criterion, proved the infeasibility of unsupervised numerical encoding, and proposed an exact and efficient algorithm based on Unimodal Cost 2-Median.
Binary-Gaussian: Compact and Progressive Representation for 3D Gaussian Segmentation
An Yang (University of Science and Technology of China), Cong Liu (iFLYTEK Research)
SegmentationNeural Radiance FieldContrastive LearningGaussian SplattingMeshBenchmark
🎯 What it does: Propose a binary encoding scheme based on 3D Gaussian profiles to achieve category representation for each Gaussian; realize fine-grained segmentation through hierarchical multi-granularity encoding and progressive contrastive learning;
BiO-HMC: Dynamic Human-Machine Collaboration for Consensus Decision-Making via Bilevel Optimization
Yinghui Pan (Shenzhen University), Mingwei Lin (Fujian Normal University)
OptimizationImage
🎯 What it does: Propose the BiO-HMC framework, modeling the consensus decision-making task as a bi-level optimization to dynamically integrate answers and proactively select the most valuable questions in human-machine collaboration.
BioDPP: Dynamic Prompt Policy Learning for Biomedical Vision-Language Models
Pingyi Miao (Big Data Institute, Central South University), Ying An (Xiangya Hospital, Central South University)
ClassificationDomain AdaptationKnowledge DistillationTransformerReinforcement LearningPrompt EngineeringVision Language ModelBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Propose a dynamic prompting strategy based on reinforcement learning, enabling the model to adaptively generate context-aware prompts for each medical image, thus achieving efficient VLM adaptation under few-shot scenarios.
Biologically-Inspired Evolutionary Domain Symbiosis for Few-shot and Zero-shot Point Cloud Semantic Segmentation
Changshuo Wang (University College London), Prayag Tiwari (Halmstad University)
SegmentationDomain AdaptationPoint Cloud
🎯 What it does: Proposed a unified few-shot and zero-shot point cloud semantic segmentation framework called EDS-Net, leveraging bio-inspired evolutionary symbiosis and genetic crossover mechanisms to address the challenges of support-query domain differences and visual-semantic modality fusion.
Bipartite Mode Matching for Vision Training Set Search from a Hierarchical Data Server
Yue Yao (Shandong University), Tom Gedeon (Northeastern University)
RecognitionObject DetectionDomain AdaptationImage
🎯 What it does: Propose a training set search framework based on hierarchical data servers and bilateral mode matching (BMM) for unsupervised domain adaptation tasks (person/vehicle re-identification and object detection)
BIQ: Bisection Interval Quantization for Communication-efficient Federated Learning
Luyang Gai (Xi'an Jiaotong University), Zihao Zhou (Xi'an Jiaotong University)
Federated LearningImageTabular
🎯 What it does: Proposes two non-uniform quantization methods, BIQ and WBIQ, based on binary interval quantization, specifically designed for federated learning to improve communication efficiency and reduce error.
BitDP: Ultra-low-bit Communication for Data Parallelism in LLM Training
Xiaozhe Ren (Hong Kong University of Science and Technology), Qiong Luo (Hong Kong University of Science and Technology)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the BitDP system to achieve 1-bit/2-bit ultra-low precision gradient quantization and communication, significantly reducing communication overhead during LLM training while maintaining high accuracy;
BLADE: A Behavior-Level Data Augmentation Framework with Dual Fusion Modeling for Multi-Behavior Sequential Recommendation
Yupeng Li (University of Science and Technology of China), Shijin Wang (University of Science and Technology of China)
Recommendation SystemTransformerMixture of ExpertsContrastive LearningVideoTextSequential
🎯 What it does: Propose the BLADE framework, combining early and intermediate dual-stage fusion with behavior-level data augmentation to address the issues of behavioral heterogeneity and sparsity in multi-behavior sequence recommendation.
Blessing of Dimensionality for Approximating Sobolev Classes on Manifolds
Hong Ye Tan (UCLA), Carola-Bibiane Schönlieb (IIT Kharagpur)
🎯 What it does: In this paper, the authors derive lower bounds based on nonlinear width for approximating bounded Sobolev class functions on compact Riemannian manifolds; specifically, they show that under the Lp norm, the minimal error required to approximate Sobolev balls using function classes with finite pseudo-dimension decreases at a rate of (n+log n)^{-1/d} with respect to the pseudo-dimension n of the function class (d is the intrinsic dimension of the manifold).
BLM-Guard: Explainable Multimodal Ad Moderation with Chain-of-Thought and Policy-Aligned Rewards
Yiran Yang (Kuaishou Technology), Jiefei Zhang (Kuaishou Technology)
ClassificationAnomaly DetectionExplainability and InterpretabilityReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningVision Language ModelVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose BLM-Guard, a multimodal ad review framework that integrates chain-of-thought reasoning with policy alignment, capable of automatically determining whether short video ads violate platform policies and providing explainable reasoning chains.
Blur-Robust Detection via Feature Restoration: An End-to-End Framework for Prior-Guided Infrared UAV Target Detection
Xiaolin Wang (Xidian University), Luxin Yan (Xidian University)
RestorationObject DetectionConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: An end-to-end dual-branch framework named JFD3 was designed, jointly performing feature domain deblurring and object detection. It leverages features from clear images to supervise the restoration of features from blurred images, and improves the detection accuracy of blurred infrared UAV targets through a frequency domain structure guidance module.
BOFA: Bridge-Layer Orthogonal Low-Rank Fusion for CLIP-Based Class-Incremental Learning
Lan Li (Nanjing University), De-Chuan Zhan (Nanjing University)
ClassificationRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Propose a framework named BOFA, which leverages CLIP's bridge layer for adaptation, achieving parameter-free incrementality, no catastrophic forgetting, and no example replay in class-incremental learning.
BokehCrafter: Taming Video Diffusion Models for Controllable Bokeh Rendering
Qiwen Wang (Huazhong University of Science and Technology), Zhiguo Cao (Huazhong University of Science and Technology)
GenerationData SynthesisVision Language ModelDiffusion modelAuto EncoderVideoText
🎯 What it does: Propose BokehCrafter, a controllable video bokeh rendering framework based on video diffusion models, which achieves bokeh rendering without parallax or focal parameters using only text instructions and reference bokeh images.