• 学术搜索
  • 科研智能体
    • Research Labs
    • AI 阅读
    • AI 文库
    • 深度研究
    • 学者亮点
  • 学术资源
    • AI2000
    • 期刊/会议
    • 学者库
    • 学术API
      MCP&Skill
    • 溯源树
    • 数据集
  • 知识沉淀
    • 学术空间
订阅小程序
旧版功能
aminer vip
开通会员低至0.73元/天
一次搞定AI科研
立即登录
  • English
  • 联系方式
    IEEE Transactions on Artificial Intelligence

    IEEE Transactions on Artificial Intelligence

    JournalISSN 2691-4581

    年发文量

    研究主题

    论文(1694)

    排序
    1Explicit-implicit Associative Fusion Network for Multimodal Entity Linking
    Zihao Chen,Chenyang Bu,Xindong Wu

    Multimodal entity linking (MEL) involves using multimodal information to connect a given mention to its corresponding entity in a knowledge base. Existing MEL methods typically start by employing encoders to derive embedded representations of mentions and entities, and then neural networks are used to implement interaction computations between the vectors to obtain semantic similarities between mentions and entities. The variations among methods primarily lie in the design of these neural networks. However, these approaches still face two main limitations: first, they focus mainly on discovering implicit semantic similarities between fine-grained features, ignoring the possibility that explicit symbolic features may still be useful in the linking process. Second, multimodal data may experience missing modalities, introducing noisy interference that impacts the accuracy of linking. To address these challenges, we propose an Explicit-Implicit Associative Fusion Network (EIAFN) to accomplish the linking task by fusing the semantic similarity of mentions and entries with the similarity of symbols. EIAFN first processes the text and image information of mentions and entities using encoders to obtain embeddings at different levels of granularity. Next, multiple computational units are used to compute the similarity between features of different modalities and granularities, respectively, taking into account both fine-grained semantic similarity and coarse-grained symbolic similarity. Finally, the multiple computational results are fused by the matching fusion module, and the weights of the computational units are dynamically updated to minimize the effect of noise caused by missing modalities. Experimental results show that our method outperforms state-of-the-art methods on three publicly available datasets. The code is available at https://github.com/DMiC-Lab-HFUT/EIAFN.

    2026
    引用
    AI阅读
    加入学术空间
    2Prompt-Aware Structured Retrieval for Interpretable Reasoning in Large Language Models
    Zhenzhen Huang, Yitian Zhou,Chaoning Zhang,Jiaquan Zhang, Chi-lok Andy Tai, Pengcheng Zheng,Zeyu Ma,Kuien Liu, Sung-Ho Bae,Caiyan Qin,Jinyu Guo,Yang Yang,

    Recent advances in Retrieval-Augmented Generation (RAG) have enabled large language models (LLMs) to incorporate external knowledge beyond fixed parametric memory, but most existing methods still concatenate prompts, queries, and retrieved passages into a flat input. This unstructured design leads to redundancy, fragmented evidence utilization, and limited interpretability in reasoning-intensive tasks. To address this limitation, we propose Prompt-Aware Synergistic Optimization of Retrieval Content (PASO-RC), a structured RAG framework that rearranges RAG inputs into a unified, step-guided prompt rather than a loose concatenation. Instead of treating retrieved content as independent evidence for reasoning, PASO-RC treats the query, prompt, and retrieval content together as a structured, task-specific context, enabling efficient and transparent use of external knowledge. PASO-RC employs knowledge alignment that dynamically anchors retrieved knowledge units to the corresponding steps of an explicit chain-of-thought (CoT) reasoning process, and then applies prompt fusion to construct a compact, task-specific reasoning template. We further introduce a new evaluation dimension, reasoning fidelity, to evaluate the alignment between intermediate reasoning steps and their supporting evidence beyond task accuracy. Experiments on multiple open-domain QA and verification datasets demonstrate that PASO-RC substantially outperforms state-of-the-art RAG methods in accuracy, coherence, and transparency, providing a new pathway toward interpretable and controllable knowledge-augmented reasoning.

    2026
    引用
    AI阅读
    加入学术空间
    3A Novel Global Context-Driven Feature Pyramid Network for Multi-Domain Semantic Segmentation
    Dayu Tan, Cheng Kong, Kai Zhong,Hai Chen,Dongliang Yang,Weimin Zhong, Min Han

    Accurate multi-domain semantic segmentation in medical images remains challenging due to the complex anatomical structures and variable organ sizes. Although Transformer-based networks have gained popularity for modeling long-range dependencies, these methods often incur substantial computational overhead and exhibit limitations in capturing fine-grained local details. This paper presents LamFormer, a global-context-driven feature pyramid network that targets fine-grained multi-organ delineation. At its core lies a pyramid encoder built on a Linear Attention Mamba (LAM) block, which attains multi-scale contextual modelling at linear cost. A Parallel Hierarchical Feature Aggregation (PHFA) block adaptively recalibrates the features drawn from successive encoder stages, thereby shrinking the semantic gaps between shallow and deep representations while suppressing uninformative responses. A Reduced Transformer (RT) block is further introduced in the decoding path so that global reasoning is preserved after upsampling, which helps recover fine structures. Extensive experiments on seven diverse and challenging public datasets demonstrate that LamFormer achieves superior average overall performance against mainstream state-of-the-art methods across seven datasets, with competitive results on most organ segmentation tasks. In particular, our approach achieves an optimal balance between superior segmentation performance and manageable model complexity. The code has been made available on GitHub: https://github.com/kec1212/LamFormer.

    2026
    引用
    AI阅读
    加入学术空间
    4FuzzyCS: Adaptive Client Selection Via Fuzzy Rough Approximations in Federated Learning
    Qianlong Dang,Baosheng Li, Binquan Guo,Qiqi Liu,Shoufei Han,Guoqing Li

    The client selection of federated learning presents a pivotal dilemma necessitating the harmonization of performance and efficiency. Prevailing studies frame this dilemma within the context of a multi-armed bandit optimization problem. With an expanding client base, the complexity of the task increases exponentially, demanding advanced algorithms to efficiently navigate the decision space. Fuzzy sets offer a clear depiction of the uncertainties associated with client selection, eliminating the need for intricate numerical computations. To this end, this paper proposes a fuzzy client selection (FuzzyCS) method. Specifically, we construct a fuzzy multi-criteria information matrix based on several selection criteria and use it to introduce a criteria-oriented fuzzy selection model. By leveraging the fuzzy rough approximations, we can estimate the fuzzy relationship between the criteria-oriented fuzzy concepts of each client and its likelihood of being selected. FuzzyCS can actively balance performance and efficiency by constructing an adaptive fuzzy approximation set, which incorporates the upper and lower approximate fuzzy rough sets. Experimental results demonstrate that FuzzyCS achieves the top Friedman rank in both performance and efficiency. The code is available at https://github.com/dqlme/FuzzyCS.

    2026
    引用
    AI阅读
    加入学术空间
    5Off-Policy Receding-Horizon Reinforcement Learning for Formation-Containment Control of Multi-Robot Systems
    Changxin Zhang,Ronghua Zhang,Xin Xu,Xinglong Zhang,Qingwen Ma, Xinxin Yao,Quan Xiong

    Multi-robot formation-containment control, which requires multiple leader agents to maintain a prescribed formation while keeping follower agents inside the convex hull they form, is a representative cooperative decision-making problem for artificial intelligence. Solving its underlying nonlinear optimal control problem in an online and data-efficient manner remains an open challenge for learning-based controllers. As a powerful data-driven framework for nonlinear optimal control, current reinforcement learning (RL) approaches still suffer from insufficient online adaptability and low sample efficiency, while probing actions used for excitation can directly contaminate on-policy learning targets. To address these limitations, this paper develops an off-policy receding-horizon reinforcement learning (Off-RHRL) controller. The method rolls out a behavior policy on a known local prediction model, reuses the resulting transitions, and algebraically corrects each transition for a target policy. We establish the exact ideal finite-horizon recursion, prove a per-sample behavior-action cancellation identity, and give a conditional local ISS/UUB estimate for the forced coupled learned closed loop under assumed nominal decrease, bounded actor error, and robust-invariance conditions. The implementation is evaluated in multi-seed simulation studies against on-policy RHRL, TD3, and matched nonlinear MPC, with additional sensitivity and model-mismatch tests; a four-robot experiment is presented as a feasibility demonstration.

    2026
    引用
    AI阅读
    加入学术空间
    立即登录,查看全部 1694 篇论文

    高被引作者

    作者引用发文
    Huan Liu5322
    Houbing Herbert Song5242
    Jian Wang4901
    Yongxin Liu4901
    Shuteng Niu4901
    Feng Xia4574
    Ke Sun4501
    Shirui Pan4501
    Abdul Aziz4501
    Liangtian Wan4501

    高产作者

    作者引用发文
    C. L. Philip Chen14117
    Yu Kang7715
    Prayag Tiwari589
    Shaocheng Tong229
    Zhenghua Chen1318
    Yang Cao738
    Ali Etemad487
    Mengjie Zhang (Meng)1067
    Nishchal K. Verma307
    Yi Mei337

    相关期刊

    International Conference on Control, Automation, Robotics and Vision
    International Conference on Cognitive Systems and Signal Processing
    International Conference on Robotics and Artificial Intelligence
    Robotics, Automation and Mechatronics
    Robot Intelligence Technology and Applications
    Conference proceedings. IEEE International Conference on Systems, Man, and Cybernetics
    Journal of Robotics, Networking and Artificial Life
    Annual Meeting of the IEEE Industry Applications Society
    International Conference Robotics and Computer Vision
    Cognitive Computation and Systems