Graph-Verified Bootstrapping for Multi-Hop Knowledge Reasoning

Abstract

Multi-hop knowledge reasoning over knowledge graphs remains fundamentally challenging under weak supervision, as lightweight models lack intermediate signals for relation composition, while large language models suffer from hallucinations. We propose a graph-verified bootstrapping framework that exploits the knowledge graph itself as a deterministic verifier to guide reliable self-enhancement. The framework is built upon differentiable relation graph propagation, where a student model traces entity activation signals step by step along relation edges, rendering the entire multi-hop reasoning chain structurally verifiable against the graph. This intrinsic transparency enables a closed-loop bootstrapping process: on low-confidence samples, a large language models teacher is invoked to propose candidate answers, each of which is then strictly examined by graph-connectivity verification, so that only answers that correspond to valid paths in the graph are retained, while unreachable hallucinations are discarded. The verified answers are iteratively injected back into training, allowing the student to progressively expand its reasoning boundary under the continuous supervision of graph-level verifiability. Experiments on a public multi-hop benchmark demonstrate that, relying solely on single-hop annotations, the framework achieves substantial and sustained accuracy gains across complex 2-hop and 3-hop questions over multiple bootstrapping rounds, confirming that graph-verified bootstrapping provides a principled path toward reliable teacher-student collaboration in weakly-labeled reasoning scenarios.