Tool-calling extraction reveals hidden CoT of frontier models: GPT-6 "mental-math" skips steps

机器之心 · wechat · 2026-09-29

Researchers from Aalborg University's AI Safety Lab and the Seafill open-source community published a paper showing that tool calling via standard APIs can extract hidden chain-of-thought from frontier closed models like GPT-6 Astra and Claude.

Method works: On open models, extracted reasoning matches or slightly beats native reasoning accuracy, with text similarity near the model's own rollout variance; on closed models it performs close to native reasoning across math/code/science benchmarks.
Astra "mental-math": its composition of reasoning activities matches other models, but its reasoning tree is narrower — it skips basic arithmetic and algebra, surfacing only key conclusions. It reasons just as much, but reveals less.
CoT value depends on the reader: Astra's compressed traces are nearly lossless for strong models but hard for weak ones, which still answer wrong on 73/80 questions where Astra wrote the correct answer. The strongest teacher doesn't write the best traces for weak students.
Safety implications: disabling native reasoning doesn't stop leakage via tool calls or visible replies; defenses must cover all channels. OpenAI and Anthropic were informed.

Original post →

Tool-calling extraction reveals hidden CoT of frontier models: GPT-6 "mental-math" skips steps

More from Safety