模型是否知道它们为何改变主意？知识冲突下思维链的可解释性与忠实性
Do Models Know Why They Changed Their Mind? Interpretability and Faithfulness of Chain-of-Thought Under Knowledge Conflict
AI总结：
通过引入内省忠实性，研究在知识冲突下语言模型的思维链推理是否忠实反映其决策机制，发现CoT高度稳定但置信度携带微弱真实信号。
摘要：
当语言模型看到与其训练知识相矛盾的文档时，它必须做出选择：遵循文档还是相信自己。先前的工作证明这种选择取决于事实的知名程度。我们问：模型的思维链（CoT）推理是否忠实地报告了这一机制？我们引入了内省忠实性，并在200个问题、8个模型和4种提示条件下进行了测试。我们发现CoT推理在相反决策下高度稳定：翻转对保留了96%的相同答案相似度（d=0.34；通过ROUGE-L确认，d=0.45）。然而，自我评定的置信度携带微弱的真实信号：对于实体知名度无信息量的冷门事实，置信度仍能预测决策（p<0.001），并追踪项目级知识（r=0.134）。GPT-4o是唯一具有统计上可靠的推理-决策耦合的模型。Claude Sonnet 4.6显示出最宽的置信度范围（SD=1.39），但汇总相关性接近零，因为置信度-决策关系在不同条件下反转；温度消融实验证实这是模型特有的。内部思考令牌比面向用户的CoT显示出更大的决策敏感性（p=0.033）。CoT分解为决策不变的知识展示（约96%）和一层薄弱的置信度层，后者带有微弱但真实的信号。对于监控：读取置信度，而不是论证。
英文摘要：
When a language model sees a document contradicting its training knowledge, it must choose: follow the document or trust itself. Prior work proved this choice depends on how well-known the fact is. We ask: does the model's chain-of-thought (CoT) reasoning faithfully report this mechanism? We introduce introspective faithfulness and test it across 200 questions, 8 models, and 4 prompt conditions. We find CoT reasoning is highly stable across opposite decisions: flip pairs retain 96% of same-answer similarity (d=0.34; confirmed by ROUGE-L, d=0.45). Yet self-rated confidence carries a faint genuine signal: for obscure facts where entity fame is uninformative, confidence still predicts decisions (p<0.001) and tracks item-level knowledge (r=0.134). GPT-4o is the only model with statistically reliable reasoning-decision coupling. Claude Sonnet 4.6 shows the widest confidence range (SD=1.39) but near-zero pooled correlation because the confidence-decision relationship reverses between conditions; a temperature ablation confirms this is model-specific. Internal thinking tokens show greater decision-sensitivity than user-facing CoT (p=0.033). CoT decomposes into a decision-invariant knowledge display (~96%) and a thin confidence layer with weak but real signal. For monitoring: read confidence, not the argument.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
