Capable yet Parsimonious: Extracting and Characterizing Hidden Chain-of-Thought in Frontier Models
A custom API tool made closed models expose intermediate reasoning traces that behaved like real chain-of-thought in the authors’ tests.
The paper first checks the method on open-source models, comparing extracted traces with native chain-of-thought. It then applies the same setup to frontier closed models, including GPT-6 Astra, across math, science, and code tasks. The extracted reasoning matched native reasoning performance and beat no-reasoning baselines. Astra is described as using fewer tokens, choosing a correct path earlier, and leaving simpler steps internal while externalizing key moves. HF Daily Papers' note
The paper first checks the method on open-source models, comparing extracted traces with native chain-of-thought. It then applies the same setup to frontier closed models, including GPT-6 Astra, across math, science, and code tasks. The extracted reasoning matched native reasoning performance and beat no-reasoning baselines. Astra is described as using fewer tokens, choosing a correct path earlier, and leaving simpler steps internal while externalizing key moves. HF Daily Papers' note
score 6