Frontier language models have made rapid capability gains widely attributed to improved reasoning, but that reasoning has remained largely unverifiable because raw chain-of-thought (CoT) traces in closed-source systems are hidden. A new preprint, Capable yet Parsimonious: Extracting and Characterizing Hidden Chain-of-Thought in Frontier Models, proposes a way to change that.
According to the authors, registering a simple custom tool through a standard API feature induces frontier models to externalize their intermediate reasoning. To check whether these traces reflect genuine reasoning rather than post-hoc rationalization, the team first evaluated the method against native CoT on open-source models, then extended it to closed-source frontier systems including GPT-6 Astra, GPT-5.6 Sol, Claude Opus 4.8, and Claude Sonnet 5.
The researchers report that the extracted reasoning matches native reasoning performance and substantially outperforms no-reasoning baselines across competition mathematics, science, and code generation. They then characterized how frontier models structure their intermediate reasoning, examining token efficiency, reasoning-step types, and induced reasoning trees, and identified systematic differences in how models externalize, compress, and organize reasoning.
Among the findings, Astra stood out for what the authors describe as token-efficient directed reasoning: it selects a correct trajectory earlier, resolves elementary steps internally, and externalizes only crucial reasoning. Locally, its trace resembles mental arithmetic with routine calculations left implicit; globally, it backtracks less and reaches solutions with little visible trial-and-error.
The authors caution that this efficiency cuts both ways. A very short path to a correct answer is exactly what genuine skill and memorized test data both look like from the outside, which is why being able to see the reasoning matters: a right answer can hide wrong reasoning, or no reasoning at all.
The work is a collaboration across AAU-NLP, the AI:SECURITY Lab, and the Seafill Open Source Community, with co-authors Tao Ren, Wenrui Yu, Xiao Li, Qiongxiu Li, and Johannes Bjerva. The authors state that their method has been disclosed to OpenAI and Anthropic.