Ordinary distillation shows the student questions and answers. Reasoning distillation shows it the working, and for multi-step problems that difference is large.
Why answers alone are not enough
Given a hard multi-step problem and only the final answer, a student has to infer the entire procedure from input and output. On arithmetic, logic, or code, that is an enormous inferential leap, and small models generally fail to make it. They learn to guess plausibly rather than to compute.
Given the intermediate steps, the task changes completely. The student is no longer inferring a hidden procedure. It is being shown one, repeatedly, and asked to reproduce it.
How it is done
- Prompt a strong teacher to solve problems while showing its reasoning
- Keep the traces that reach correct answers, discard those that do not
- Fine-tune the student on the full traces, reasoning and answer together
- The student learns to produce reasoning before answering, and answers better as a result
Step two is the one that matters most, and it is the reason this works at all. Filtering by correctness turns a noisy teacher into a clean training set. This is sometimes called rejection sampling, and it is doing a great deal of the work.
The characteristic failure
The known risk is that a student learns the appearance of reasoning without the substance. It produces confident, well-structured, correctly formatted chains of thought that do not actually constrain the answer, and it produces them just as fluently when the answer is wrong.
This is worse than a model that simply guesses, because the visible reasoning creates unearned confidence in the reader. Evaluating reasoning-distilled models requires checking whether the steps are load-bearing, not whether they look right.
Why it mattered in 2026
Reasoning capability is the most valuable thing to transfer and the most expensive to develop independently, which makes it the natural target of any serious extraction effort. The disclosures during 2026 said so explicitly: the alleged campaigns were described as targeting multi-step reasoning, software engineering and agentic tool use, not general conversation.
That targeting is itself informative. It indicates an effort aimed at specific capability transfer rather than broad imitation, which is a meaningfully different activity from scraping a model for style.
Common questions
What is reasoning distillation?
Reasoning distillation trains a student on a teacher's intermediate reasoning steps rather than only its final answers. For multi-step tasks this transfers considerably more capability than answer-only training, because the student learns a procedure rather than a mapping.
Can a small model learn to reason from a larger one?
To a significant degree, yes. Small models trained on high-quality reasoning traces substantially outperform equivalent models trained on answers alone. The limit is that the student can imitate the form of reasoning more easily than the reliability of it, so verification of the traces matters more than volume.