引用brendoelfrendo · 2025-08
“This sometimes leads to the desired outcome but the phenomenon is very brittle and disappears when the AI is pushed outside the bounds of its training.”
拿它说明思维链很脆弱,超出训练范围就失效
思维链只在训练分布内有效,越界即失效
论文用数据分布的视角解释思维链何时有效:CoT 推理反映的是从分布内数据学到的结构化归纳偏置,其效果取决于训练数据与测试问题之间的分布差异。作者构建可控环境 DataAlchemy,从零训练模型,沿任务、长度、格式三个维度做实验,发现 CoT 一旦超出训练分布就变得脆弱,像一个海市蜃楼。
HN 上的人引用它说明思维链带来的收益很脆弱,一旦问题超出训练范围就会消失。
“This sometimes leads to the desired outcome but the phenomenon is very brittle and disappears when the AI is pushed outside the bounds of its training.”
“Examples:”