Grounding Multi-Hop Reasoning in Structural Causal Models via Group Relative Policy Optimization

The article presents a novel framework for Multi-Hop Fact Verification that utilizes Structural Causal Models (SCM) and Group Relative Policy Optimization (GRPO) to enhance reasoning accuracy in Large Language Models (LLMs). By employing directed dependency graphs to model structural relationships between evidence and claims, the framework addresses the challenges of hallucinations and logical fragmentation in existing methods. Empirical results indicate that this approach not only improves performance over strong baselines on datasets like HoVer and EX-FEVER but also provides more traceable reasoning structures, which is critical for practitioners focusing on reliable AI systems.