Closed labs have been locking reasoning traces away for a while: the API returns an answer plus a short summary, and the chain of thought itself stays hidden. Replaying encrypted blocks across models can steal it back — the Stealing Reasoning Traces paper from this August proved the route works — but it breaks provider terms of service, so no product can build on it long term. On October 21, independent researcher Yan Wang published TraceForge on arXiv, offering a third route: do not steal the teacher's reasoning; write your own, in the style of the student's own model family.
Traces can be synthesized
TraceForge grows two models against each other. In the supervised stage, a generator learns to restore roughly ten thousand real reasoning traces of Qwen3.8-Max from their context and answer. In the adversarial stage, a critic learns to tell real traces from generated ones; samples that fool the critic become the next round's training data, and online co-training converges after five rounds. At inference, the generator writes a "family-native" trace for the teacher's answer, and the critic picks the most convincing of four candidates. Scored by a closed judge (GPT-6.1-SOL), synthetic traces reach 68% perceived authenticity, slightly above the 66% of real traces, and 82% with best-of-four selection. The method section records one honest mistake: the author assumed a smaller model could grade a larger one; within a few hundred steps the critic fitted shortcuts, so the critic must be at least as large as the generator.
A clean five-corpus comparison
Every student is fine-tuned from Qwen3.5-9B-Base with base model, recipe, and budget held fixed. The data pool is 5,000 Claude Code trajectories carrying Fable 5's real reasoning traces and answers (3,500 questions, 256K context window, eight H20 GPUs), and only the source of the trace and the answer varies. The paper reports: replacing Fable 5's own traces with family-native ones lifts HumanEval+ from 64.63 to 68.29 and MBPP+ from 52.38 to 59.52 — gains of 3.7 and 7.1 points; distilling directly from the family's strongest model, Qwen3.8-Max, yields only 46.95 and 24.60 — 21 and 35 points behind; the generic TraceInversion traces collapse to 3.66 and 6.61. Switching to a second teacher, GPT-6.1-SOL, keeps the conclusion: given identical answers, family-native traces reach 59.15 and 60.05 while the third-party synthesizer stays at 3.66 and 6.61.
Why the family accent wins
The paper attributes the gain to smaller catastrophic forgetting: student and traces come from the same family, so the student absorbs them without unlearning its own habits. Two pieces of side evidence: the teacher-trace corpus (S1) is unstable, with HumanEval+ swinging between 66.46, 33.54, and 64.63 — a 33-point drop that repairs itself; and the Qwen-Max and TraceInversion corpora drift into dead loops and repetition as training proceeds, the so-called second-hand-smoke effect, a pattern family-native traces never show. Teacher capability is necessary but not sufficient: the answer can come from the strongest model, but the delivery has to be the student's own.
Limits and take
Honest caveats: single-author independent research, one run per cell, a closed-source judge, and a confidential source for the 5,000 trajectories. Still, five corpora, two teachers, and per-checkpoint cross-checks make this an unusually clean experiment for the direction. For the industry, it moves the fight over reasoning traces from stealing to synthesizing: for the first time, staying legal costs no performance. The paper also offers a variant where zero Claude or OpenAI tokens are ever backpropagated — the family-native trace paired with a Qwen3.8-Max answer — which performs essentially the same as the main setup, removing the legal exposure entirely. Back to the opening question: the best student is not produced by the best teacher alone, but by the best teacher teaching the way the student learns best. Paper: arXiv:2610.29680 (https://arxiv.org/abs/2610.29680).