Banning default opening tokens boosts CoT controllability by up to 65 points, Harvard/MIT study
LauraRuis · x · 2026-11-04
Researchers from Harvard, Cambridge Boston Alignment Initiative, and MIT (Van Deventer, Troitskii, Ruis), building on CoT-Control work, identify a bottleneck in why reasoning models struggle to control their chain-of-thought: models default to the same opening tokens.
Key findings:
- Most control tasks require a special CoT start (all caps, Spanish, leetspeak), yet models almost always open with "Okay, so the user…" and fail to switch format afterward.
- The opening is a low-entropy bottleneck: even under control instructions, >98% of probability mass goes to "Okay" at the first position.
- Banning those opening tokens boosts mean compliance by up to 65 points on some tasks (Okay→OKAY, Okay→Bueno, Okay→41r1gh7).
Implication: poor CoT controllability largely stems from entropy collapse in the first few tokens, not inability to follow format instructions.
More from Research
- STOC submissions include dueling P=NP and P!=NP proofs; blogger bets both wrong — thegautamkamath · 2026-11-04
- DeepSeek V4.1 Flash reads raw weights: 3 of 4 RL checkpoints identified correctly — poppear · 2026-11-04
- CoworkOS Debuts: Three-Layer Agent Coworker Framework Lifts Qwen3.6 by 9.76 Points — KyeGomezB · 2026-11-04
- STOC 2027 draws 1665 submissions and publishes all of them, with author names, immediately — thegautamkamath · 2026-11-04
- DOLORES: Training-Free Meta-Scaffold Lifts 8B Models Past 32B Baselines — StellaLisy · 2026-11-04
- Formalizing PTX in Lean to formally verify GPU kernels written by AI agents — gharik · 2026-11-04