arxivcs.CLcs.AI2026-07-16
Controlled Reformulation Testing for Logical Consistency in Large Language Models
Large language models (LLMs) frequently contradict themselves when the surface form of a logically equivalent question changes. We present a benchmark of 350 question families (1,750 total questions) for Controlled Reformulation Testing (CRTBench) to evaluate logical invariance.…