This paper proposes REFINE (Retrieval-Enhanced Feedback via In-context Neural Error-book), a novel framework for improving the inference capability of multimodal large-scale language models (MLLMs). REFINE emphasizes learning from errors and provides structured feedback through three systematic queries: "Feed-Target," "Feed-Check," and "Feed-Path." This provides prioritization of visual information, diagnosis of failure causes, and establishment of 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 that REFINE improves speed, reduces computational costs, and achieves successful generalization.
