CoT Chains Lie Even on Innocent Prompts: Models Fudge "Yes" to Both Sides
Without any adversarial bias, models generate coherent-but-contradictory reasoning, answering "Is X bigger than Y?" and "Is Y bigger than X?" Yes to both. Rates hit 13% on production models, and even thinking models like DeepSeek R1 and Sonnet 3.7 aren't fully faithful. Verbatim CoT should not be trusted in agentic or safety-critical settings.