Introspective Dynamic Strategy for Knowledge Graph Enhanced Reasoning in Large Language Models

Abstract

Large Language Models (LLMs) achieve strong performance on many tasks, yet their reasoning can remain brittle and their decision process is often insufficiently transparent. Knowledge graphs (KGs) provide explicit entities and relations that can ground reasoning and improve traceability. However, most KG-augmented approaches rely on static or single-path integration, which limits dynamic exploration of alternative reasoning trajectories and weakens verification of whether intermediate evidence truly supports the final answer.We propose the Introspective Dynamic Strategy (IDS), a KG-introspected framework that jointly performs Introspection and Inspection. Introspection generates and expands multiple KG-grounded reasoning paths, while Inspection assesses the adequacy of the current chain and decides whether to finalize the answer or continue refining it. Experiments on medical question answering datasets demonstrate that IDS improves semantic alignment and yields competitive answer quality over competitive baselines.(IDS performance varies across metrics/datasets; we avoid overstating universal superiority.)