Translation in progress Some entries still include text in the original language.
Redwood Research· Arun Jose·Indexed by this site · Originally publishedSelectedAttention50
Redwood Research：CoT 可控性评测存在严重提示词欠激发
CoT controllability evals seem very under-elicited
AI summary
Redwood Research 的 Arun Jose 发现，CoTControl 评测对提示词高度敏感，用 Claude Opus 4.6 迭代提示模板后，开源模型的合规率提升约 2-3 倍，例如 GPT-OSS-120B 从 5.5% 升至 15%。
Why it was selected
作者用提示词迭代把开源模型的 CoT 可控性提升 2-3 倍，说明现有评测可能低估了模型隐藏思维链的能力。
Read originalblog.redwoodresearch.org