Generating training data for one LLM using another is one of the most general training methods and remains common practice. Some leading teams have blocked this route for their own models, returning only an answer or a short summary without the detailed reasoning trace, which is ironic because those models were themselves built by distilling human knowledge. Prior work recovers hidden reasoning by replaying encrypted blocks across models; it works but violates providers' terms of service, making long‑term productization infeasible. We propose an alternative: given the context and the teacher's answer, we generate a reasoning trace in the style of the student model family. Training proceeds in two stages—supervised then adversarial. A generator learns to reconstruct the family’s own reasoning traces from context and answer, while a critic learns to assess how realistic a trace looks. When applied to a teacher answer, the generator writes a trace matching the family’s style. An independent judge rates authenticity on a 0‑1 scale: our traces average 0.68, real traces 0.66, and the critic‑selected best of four candidates scores 0.82. These scores measure authenticity, not correctness. We evaluate on five benchmarks with all other factors fixed. On general tasks the five training corpora perform similarly; on coding tasks the generated reasoning outperforms Fable 5’s native thinking content. With the same 5,000 questions and teacher answers, HumanEval+ improves by 3.7 points and MBPP+ by 7.1 points, and the MBPP+ lead persists at every checkpoint. We suspect that reasoning written in the student family’s style is easier for the model to learn from.
Review