Reasoning models can steer their chain of thought far better than earlier tests suggested, paper finds
dair_ai · x · 2026-10-20
A new finding shared by dairai on reasoning models and chain of thought: reasoning models can control their CoT much better than earlier evaluations suggested.
The key difference is where the constraint lives. Earlier evals put the constraint in the user prompt and found very low controllability; researchers placed the same constraint inside the chain of thought as a self-instruction, and models followed it far more often.
A generic reminder to follow the user's constraint also helped a lot. Notably, the self-instruction can be triggered from the user prompt alone — no access to the reasoning channel is needed, which matters for monitoring and steering reasoning model behavior.
More from Models
- Users want open-format export from Claude Docs as charts only save as images — BlackHC · 2026-10-20
- EMNLP study: LLM safety interventions drop up to 4.5x when distress is wrapped in delusion — MaartenSap · 2026-10-20
- Cybersecurity-focused VLM OrcaCyber-Nano-1.0 trends on Hugging Face — orcarouter · 2026-10-20
- 0.8B local model with 2,000 labels beats Jev's frontier API on banking data — kalyan_kpl · 2026-10-20
- New TTS benchmark tests 26 models across 30,564 requests: fastest at 46ms — mattturck · 2026-10-20
- Agents handle unreleased language Jai flawlessly, even compile-time execution — lucasmeijer · 2026-10-20