Capable yet Parsimonious: Extracting and Characterizing Hidden Chain-of-Thought in Frontier Models

Researchers show that registering a simple custom tool through a standard API induces frontier models, including GPT-6 Astra, to externalize hidden chain-of-thought; extracted traces match native CoT on open models, and Astra externalizes only key reasoning…

A paper listed on Hugging Face daily papers (2026-09-21) and arXiv cs.CR (2026-09-22) reports that registering a simple custom tool via a standard API feature can induce closed frontier models to externalize intermediate reasoning that is otherwise hidden. On open-source models, the extracted traces match native chain-of-thought performance and substantially beat no-reasoning baselines across competition mathematics, science, and code generation. When the method is extended to closed models, including GPT-6 Astra, both sources describe Astra as token-efficient: it selects a correct trajectory earlier, resolves elementary steps internally, and externalizes fewer tokens, surfacing only the crucial reasoning steps. The two reports are consistent, with no disagreements on facts or figures.

• Registering a simple custom tool through a standard API feature induces frontier models to externalize hidden chain-of-thought.
• On open-source models, extracted traces match native chain-of-thought performance.
• Extracted traces substantially beat no-reasoning baselines on competition mathematics, science, and code generation.
• In closed models, GPT-6 Astra selects a correct trajectory earlier, resolves elementary steps internally, and externalizes fewer tokens than other models tested.
• Astra is characterized as token-efficient, externalizing only crucial/key reasoning steps.
• Sources: Hugging Face daily papers entry dated 2026-09-21 and arXiv cs.CR listing dated 2026-09-22; both describe the same paper with consistent findings.