Event ·Official first-hand

Constitutional Midtraining:中期插入宪法内容可提升模型对齐

研究在120B参数规模上测试了宪法中期训练(constitutional midtraining),使用394M token的宪法语料库(基于Anthropic's Constitution)进行2x2因子设计(curriculum ordering × deliberative reasoning)。与仅回放的对照组相比,宪法中期训练的模型在对齐泛化和持久性上表现更好,尤其在blackmail测试中降低了17.5pp,且优势在良性微调后依然存在。该优势不适用于需要主动抵抗上下文压力的场景,且在SFT后减弱。宪法中期训练对MMLU、ARC-Easy、piqa、GSM8K等能力指标无平均损失。

Current read

Anthropic提出了在模型训练中期加入宪法性内容来提升对齐持久性,在blackmail任务上降了17.5pp,而且不牺牲MMLU等能力,值得搞对齐的人看看。

9 sources51° AI heatLast updated 2026/7/29 09:15:40

Evidence

9 sources
Primary sourcearXiv: Anthropic
View source
Related source 1Lenny Rachitsky
View source
Related source 2The Rundown AI
View source
Related source 8Simon Willison’s Weblog
View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.