NLP pipelines often aggregate claims extracted from retrieved passages, documents, or generated reasoning steps. Existing factuality methods typically verify claims individually or in pairs, overlooking the coherence of the entire claim set. Consequently, a set may consist of locally plausible and pairwise compatible claims yet be globally inconsistent. We formalize this as global consistency repair: given an inconsistent natural‑language claim set, retain the largest coherent subset while pinpointing the compact sources of inconsistency. To address it, we introduce the QXR framework, which treats a large language model (LLM) as a noisy set‑level consistency judge. QXR first diagnoses minimal conflicting subsets and then repairs the set by deleting a small hitting set of claims. Experiments on benchmarks derived from VitaminC, FEVER, and a conflicting‑evidence RAG dataset, using various LLM judges, show that QXR preserves substantially more valid claims than a naïve all‑at‑once judgment while maintaining high precision. These findings suggest that factuality in claim‑aggregation systems should be tackled as a set‑level repair problem rather than merely as local verification.
Review