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

Capable yet Parsimonious. Extracting and Characterizing Hidden Chain-of-Thought in Frontier Models. By Xiaoyu Luo and five others, from Aalborg University and Seafill. Published on the twenty-second of September. No code repository is linked to it.

The capability gains of frontier language models are widely attributed to better reasoning, but that cannot be verified for closed-source systems, because their raw chain-of-thought traces are hidden. The authors show that registering a simple custom tool through a standard API feature induces frontier models to externalize their intermediate reasoning, and that what comes out tracks native reasoning closely enough to study.

Extracted traces might be post-hoc rationalization rather than genuine reasoning. So the method is first evaluated against native chain of thought on open-source models, then extended to closed-source frontier models including GPT-6 Astra. The extracted reasoning matches native reasoning performance, and it substantially outperforms baselines without reasoning, across competition mathematics, science, and code generation. That comparison is the safeguard. If the extracted traces were only decoration written after the fact, they would have no reason to perform like the real thing, and no reason to beat answering without reasoning.

With the traces in hand, the authors characterize how frontier models structure their intermediate reasoning, along three lines: token efficiency, the types of reasoning steps, and the reasoning trees the traces induce. They identify systematic differences in how models externalize, compress and organize reasoning. Astra shows token-efficient, directed reasoning. It selects a correct trajectory earlier, resolves elementary steps internally, and externalizes only the crucial reasoning. Put plainly, it writes less and reaches the right path sooner, keeping the simple steps to itself and showing only the ones that matter.

The claims are behavioral. They describe what models write down when a tool invites them to, checked against native traces where those can be seen, and they cover competition mathematics, science and code generation. They say how reasoning looks from the outside when it is written down. They do not open the model's internals, and the characterization of Astra describes its externalized reasoning, not whatever it resolves internally.

The title names the pattern found in Astra: capable yet parsimonious. What the work offers is a behavioral lens on frontier reasoning that reaches past benchmark scores, opened through a feature the standard interface already provides.