Study separates AI mathematical answers from discovery of the key idea
An October 1 preprint tests whether AI can identify the structural idea that makes a mathematics solution work.
Finding the idea, not just the answer
An October 1 preprint introduces the term mathematical primitive for the compact structural idea that makes a solution possible: for example, a useful invariant, representation or reduction. The authors use it to distinguish four capabilities in AI mathematical reasoning: finding the idea independently, recognizing it in a worked solution, carrying out a solution after receiving it, and producing a correct answer without help.
Across the authors' benchmark, giving models the relevant primitive often improved their ability to execute a solution. The authors therefore argue that independent discovery of the organizing idea, rather than routine execution, is a central weakness in current systems. They also introduce a self-distillation method intended to transfer that guidance during training.
The claims are based on the authors' new benchmark and selected model evaluations, not on independently replicated research mathematics. The benchmark's definition and scoring of a key idea are themselves important assumptions to examine. This is a preprint; MathsAI has not reproduced the experiments.