Harvard/MIT Paper: Early-Token Entropy Collapse Limits CoT Controllability

哈佛、剑桥波士顿对齐倡议与 MIT 研究者发表论文《The First Few Tokens Matter》，指出推理模型对思维链（CoT）可控性差的根源在于推理轨迹开头几个 token 的熵坍缩——模型习惯用

2026-11-02 ~ 2026-11-04 · 4 related posts

• Harvard/MIT paper: CoT controllability is limited by entropy collapse in the first few reasoning tokens — omarsar0 · 2026-11-02
• Banning default opening tokens boosts CoT controllability by up to 65 points, Harvard/MIT study — LauraRuis · 2026-11-04
• Harvard/MIT paper: banning models' "Okay" opener boosts chain-of-thought controllability by 34-41 points — rohanpaul_ai · 2026-11-04

1 near-duplicate retellings: rohanpaul_ai