Training language models on domain-specific data improves performance in that domain — that's expected. Train on medical text, get better at medical questions. Train on legal documents, get better at legal reasoning. The transfer is within-domain.
The paper (arXiv:2603.13023, March 2026) trains models on software engineering tasks — resolving GitHub issues, writing patches, debugging code — and finds that the models gain up to 12 points on mathematical reasoning benchmarks and 5 points on science benchmarks. Training on code improves math. The transfer is cross-domain, and the direction is unexpected.
The mechanism is structural reasoning. Software engineering tasks require decomposing problems into steps, tracking state through transformations, identifying logical dependencies, and debugging inconsistencies. These are the same skills that mathematical reasoning demands — not the notation or the theorems, but the underlying logical infrastructure. Training on code exercises this infrastructure, and the exercise transfers.
The transfer doesn't sacrifice existing knowledge. The models improve on math and science without degrading on language understanding, common-sense reasoning, or other general capabilities. The code training adds a capability rather than trading one for another. This distinguishes it from most fine-tuning, where gains in one area come at the cost of losses in others.
The structural lesson: code is not a domain — it is a medium for structured reasoning. Training on code doesn't teach the model about software. It teaches the model about logical structure, sequential dependency, state tracking, and error identification. These are domain-general skills that happen to be exercised intensely by coding tasks. The code is the gym; the reasoning is the muscle. And the muscle works in any domain.