Diagnosing judgment flip risk of large language models in fact verification under conflicting evidence

Abstract

Large language model (LLM)-based fact verifiers can encounter conflicting evidence in retrieval-augmented pipelines, but their robustness to evidence composition remains underexamined. We introduce a flip-rate diagnostic protocol that measures how often initially correct binary judgments become incorrect when the supporting and refuting evidence is perturbed. Using six models and two datasets (FEVER and SciFact), we vary the evidence ratio, quantity, and ordering and evaluate chain-of-thought (CoT) prompting and three lightweight mitigations. On FEVER, one refuting passage alongside three supporting passages flips correct judgments in 3.3%–22.2% of cases; under full conflict, flip rates exceed 60% for every model on FEVER and 90% on SciFact. For most models, placing refuting evidence later also increases flip rates. CoT reduces flips under full conflict but is unreliable under mixed conflict. Conflict-aware stance-grouped prompting is the most consistent mitigation across datasets. These findings indicate that baseline accuracy alone is insufficient for assessing LLM fact-verifier reliability and that evidence composition should be treated as a separate axis of robustness evaluation.