Chain-of-Thought Reasoning in the Wild Is Not Always Faithful

[Submitted connected 11 Mar 2025 (v1), past revised 16 Jun 2026 (this version, v6)]

View PDF HTML (experimental)

Abstract:Recent studies bespeak that erstwhile faced pinch definitive biases successful prompts, models often omit mentioning these biases successful their Chain-of-Thought (CoT) output, revealing that verbalized reasoning tin springiness an incorrect image of really models get astatine conclusions (unfaithfulness). In this work, we show that unfaithful CoT besides occurs connected people worded, non-adversarial prompts without adding artificial biases aliases editing exemplary outputs. We find that erstwhile separately presented pinch the questions "Is X bigger than Y?" and "Is Y bigger than X?", models sometimes nutrient superficially coherent arguments to warrant systematically answering Yes to some aliases No to both, contempt the contradiction. We coming preliminary grounds that this is owed to models' implicit biases towards Yes aliases No, labeling this Implicit Post-Hoc Rationalization. Our results uncover rates up to 13% for accumulation models, and while frontier models are much faithful, nary are wholly so, including reasoning models for illustration DeepSeek R1 (0.37%) and Sonnet 3.7 pinch reasoning (0.04%). We besides analyse Unfaithful Illogical Shortcuts, wherever models usage subtly illogical reasoning to make speculative answers to difficult mathematics problems look rigorously proven. Our findings bespeak that while CoT tin beryllium useful for assessing outputs, it is not a complete relationship of the soul process that produced the model's reply and should beryllium utilized pinch be aware successful agentic aliases safety-critical settings.