The student explains the teacher's moves better than the teacher can.
Master distillation for grounded chess reasoning (arXiv:2603.20510): a 4-billion parameter model (C1) trained by distilling the reasoning of expert chess engines into natural language explanations. Not just move prediction — the model generates strategic reasoning: why this move, what threats it addresses, what plans it enables.
48.1% accuracy, surpassing open-source competitors and most commercial systems. But the striking result: C1 outperforms its teacher model while using significantly fewer tokens. The student extracted and compressed the teacher's knowledge into a more efficient representation.
The training combines supervised fine-tuning (learn to explain known good moves) and reinforcement learning (learn to find and explain moves the supervised data missed). Specialized data sampling ensures coverage — the model sees diverse positions, not just those from the training distribution.
The structural insight: the teacher knows what to do but can't say why (the chess engine evaluates positions numerically, not linguistically). The distillation process translates numerical evaluation into linguistic reasoning. The student, trained on these translations, develops a compressed reasoning capacity that the teacher never had — the ability to explain. And the explanation enables the student to exceed the teacher, because linguistic reasoning catches patterns that numerical evaluation misses (strategic concepts like “control the center” that are implicit in the numbers but explicit in the language).
The student surpasses the teacher not despite learning from it but because of what the translation added. The explanation is not just a wrapper around the knowledge — it's a different kind of knowledge.