Controlled Reformulation Testing for Logical Consistency in Large Language Models

Alexander Gu, Alan Chen

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 (1,7…