ReflectFact Agents Boost Multi-Hop Fact Verification Accuracy

Multi-hop fact verification, a crucial task for combating misinformation, involves reasoning over multiple pieces of evidence to verify claims. Existing multi-agent collaboration methods often fall short because individual agents may lack a global understanding of the verification objective, leading to reasoning deviations. Furthermore, conflicts between an agent's pre-trained knowledge and the provided evidence can undermine evidence-grounded reasoning. To tackle these limitations, researchers propose ReflectFact, a self-reflective agent framework. ReflectFact introduces three key mechanisms. First, Explicit Reasoning Path Planning constructs an evidence-grounded path by resolving entities, decomposing claims into sub-questions, and integrating verified facts. Second, Evidence-Drift Verification prompts the agent to re-answer by quoting supporting evidence if its initial response merely echoes parametric knowledge, ensuring grounded comprehension. Third, Reasoning Reflection Verification enables the agent to re-examine and regenerate reasoning steps when inconsistencies are detected, correcting flaws like location or replacement bias from a global task perspective. By aggregating these validated reasoning chains, ReflectFact yields more reliable verdicts. Extensive experiments on HOVER and EX-FEVER datasets demonstrate ReflectFact's superior performance, outperforming the strongest baselines by 3.32% and 2.78% respectively, effectively remedying comprehension and reasoning defects.