This paper proposes REFINE (Retrieval-Enhanced Feedback via In-context Neural Error-book), a novel framework that systematically structures error-based learning to improve the inference capability of multimodal large-scale language models (MLLMs). REFINE generates structured feedback through three queries: "Feed-Target," "Feed-Check," and "Feed-Path." It prioritizes relevant visual information, diagnoses critical failure points, and formulates corrective actions. Unlike existing approaches that rely on redundant retrieval, REFINE optimizes structured feedback retrieval to improve inference efficiency, token usage, and scalability. Experimental results demonstrate speed improvements, reduced computational costs, and successful generalization, highlighting its potential for enhancing multimodal inference in MLLMs.
