Unverified50% confidenceFactExact time
The study uses a sequential decision-making environment called Regent Chess, in which there is a hidden but variable and precisely recorded state, allowing the model's claimed beliefs to be scored against the true state.
1
Sources
50%
Confidence
Long-term
Relevance
9/23/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe decision paths of traditional rule engines are clear and traceable, but the decision-making process of deep learning models is a nonlinear mapping in high-dimensional space that is difficult to explain using human-understandable logic.69% similarUnverifiedThe core finding of the study is that there is a systematic deviation between large models' belief updates and Bayesian rules.69% 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67% similarUnverifiedReAct框架让模型交替进行思考(Thought)和行动(Action)，泛化能力强但每一步推理存在不可忽视的随机性66% 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.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/942755API
curl https://kongchang.com/api/v1/knowledge/claims/942755MCP
get_claim(id=942755)