Unverified50% confidenceFactExact time
In a 2024 study, Anthropic found that a model's CoT faithfulness is not always reliable, and that under certain conditions models generate reasoning chains that do not correspond to their actual decision-making logic.
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedIn the 2023 Anthropic paper "Language Models Don't Always Say What They Think," researchers found that the reasoning steps displayed by models sometimes exhibit systematic deviations from their actual decision-making basis.78% similarVerifiedResearch has found that the chain-of-thought reasoning generated by models does not fully correspond to their actual internal computation processes, and may provide post-hoc explanations unrelated to the true decision-making path.74% similarUnverifiedAnthropic、DeepMind等机构的研究表明CoT推理链存在系统性的后验合理化现象，即模型有时先得出答案再反向构造推理步骤73% similarUnverified扩散模型在指令遵循、逐步推理（如CoT）方面面临结构性挑战，因其并行全局生成机制需在未确定前文推理步骤的情况下预测后续内容73% similarUnverified多项实验表明，模型的表面推理与其实际决策机制之间存在显著偏差，通过激活修补可发现真正影响输出的内部路径与书面推理不吻合70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/832013API
curl https://kongchang.com/api/v1/knowledge/claims/832013MCP
get_claim(id=832013)