Harvard and MIT Study: Banning 'Okay' Boosts Reasoning Model Control

1 reports archived1 independent sourcesupdated 11/4/2026, 03:43:08

Synthesis & Latest Updates

1 Sources Cross-Validated

A new study reveals that reasoning models' poor chain-of-thought control stems largely from a habitual start with the word "Okay". By banning this token and prefilling one compliant token, average instruction compliance increased by 34-41 percentage points. This suggests that CoT monitoring should distinguish between opening-word bias and general controllability.

LATEST/A new study reveals that reasoning models' poor chain-of-thought control stems largely from a habitual start with the word "Okay". By banning this token and prefilling one compliant token, average instruction compliance increased by 34-41 percentage points. This suggests that CoT monitoring should distinguish between opening-word bias and general controllability.

Hourly heat curve

2 fully observed hours

Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first

Rohan Paul (@rohanpaul_ai)T2Nov 4, 03:43·78 pts
Harvard and MIT Study: Banning 'Okay' Boosts Reasoning Model Control
- Models like Qwen3-32B assign ~99% probability to starting CoT with "Okay"
- Banning "Okay/Alright" and prefilling one example token boosts compliance by 34-41%
- Information-hiding tasks show inconsistent performance, indicating a separate skill from opening-word bias
View Related Reports & Analysis Original Source