What's this about?
Researchers register a custom tool that encourages models to externalise intermediate steps, then compare those traces with native chain-of-thought on open models before applying the method to closed systems.
What exactly is hidden chain-of-thought?
Some reasoning models use private intermediate computation that users cannot inspect. An elicited trace is behaviour produced at the interface, not a direct recording of the model's internal process.
Why does this matter?
author-reported preprint; 11 HF upvotes.
- Extracted traces outperform no-reasoning baselines on maths, science and code tests.
- GPT-6 Astra reportedly externalises fewer elementary steps and commits earlier to a correct path.
- Matching performance does not eliminate the possibility of post-hoc rationalisation.