GRPO Enhances Multi-Hop Fact Verification via Causal Models

A novel framework establishes a foundation for multi-hop fact verification based on structural causal models (SCMs), conceptualizing verification as a form of constructive causal inference. The researchers discovered an 'inverted U-shaped' relationship between the length of reasoning chains and accuracy, indicating that overly complex chains hinder performance. To tackle this issue, they suggest a rule-based reinforcement learning approach utilizing Group Relative Policy Optimization (GRPO) to maintain a dynamic equilibrium between structural depth and brevity. This research aims to resolve hallucinations and disjointed logical sequences in large language models (LLMs) for effective multi-hop fact verification.