To address the knowledge conflict problem arising in augmented search generation (RAG) systems, this paper proposes Micro-Act, a framework with a hierarchical action space that automatically detects contextual complexity and decomposes each knowledge source into a fine-grained comparison sequence. Micro-Act enables inference beyond superficial context through fine-grained comparison steps, demonstrating improved QA accuracy compared to existing state-of-the-art models on various benchmark datasets. In particular, it excels in temporal and semantic conflict types and maintains robust performance even for non-conflict questions, enhancing its applicability in real-world RAG applications.
