[2603.26735] Distilled Large Language Model-Driven Dynamic Sparse Expert Activation Mechanism

Computer Science > Computer Vision and Pattern Recognition arXiv:2603.26735 (cs) [Submitted on 21 Mar 2026] Title:Distilled Large Language Model-Driven Dynamic Sparse Expert Activation Mechanism Authors:Qinghui Chen, Zekai Zhang, Zaigui Zhang, Kai Zhang, Dagang Li, Wenmin Wang, Jinglin Zhang, Cong Liu View a PDF of the paper titled Distilled Large Language Model-Driven Dynamic Sparse Expert Activation Mechanism, by Qinghui Chen and 7 other authors View PDF HTML (experimental) Abstract:High inter-class similarity, extreme scale variation, and limited computational budgets hinder reliable visual recognition across diverse real-world data. Existing vision-centric and cross-modal approaches often rely on rigid fusion mechanisms and heavy annotation pipelines, leading to sub-optimal generalization. We propose the Distilled Large Language Model (LLM)-Driven Sparse Mixture-of-Experts (DS-MoE) framework, which integrates text-guided dynamic routing and lightweight multi-scale comprehension. The DS-MoE framework dynamically aligns textual semantics with defect-specific visual patterns through a sparse MoE architecture, where task-relevant experts are adaptively activated based on semantic relevance, resolving inter-class ambiguity. A lightweight MobileSAM encoder enables real-time inference while preserving multi-scale defect details. Extensive experiments on PCB, aluminum foil, and mold defect datasets demonstrate that our framework achieves superior performance compared to existi...