From Verdict to Process: Agentic Reinforcement Learning for Multi-Stage Fact Verification

The article introduces ProFact, a novel agentic reinforcement learning framework for optimizing multi-stage fact verification workflows involving claim decomposition, evidence gathering, and verdict prediction. ProFact employs a unified policy that utilizes process-aware rewards to enhance stage-level learning, resulting in improved verification performance and inference efficiency compared to existing methods. This approach is significant for practitioners as it enables more adaptive coordination among verification stages, potentially leading to more accurate automated fact-checking systems.