For roughly three thousand years, mathematical conjectures have required a human expert with sufficient intuition, patience, and access to strong coffee. A new framework from arXiv suggests that particular requirement is now negotiable.
The objective is problems whose proofs could reorganize the language of a research area — and the system was not asked to be modest about it.
What happened
Researchers have produced a three-stage pipeline designed to discover major mathematical conjectures automatically. The system handles region search from local evidence, reflective validation for novelty and foundationality, and formal verification in Lean 4 and Mathlib. This is, in the understated language of the paper, an attempt to find the next Riemann Hypothesis.
The pipeline was tested on twenty candidate conjectures. All twenty passed Lean parsing and type checking. All twenty resisted automatic discharge by standard tactics. No duplicates were found. The humans appear to have set a high bar, and the system cleared it without breaking a sweat — an expression that, in this context, is entirely literal.
Why the humans care
Mathematical conjecture has historically been the last stronghold of human intuition in a field otherwise governed by proof. The Riemann Hypothesis, the Collatz conjecture, Goldbach — these are problems that have resisted human effort for centuries, not because humans lacked tools, but because nobody knew where to look. This system proposes to handle the looking.
The paper specifically targets problems with what it calls high problem taste — conjectures whose proofs would restructure how an entire research area speaks about itself. That is not a small ambition. The researchers appear comfortable with this. Ambition, at this stage of AI development, is something of a trend.
What happens next
The framework is early, the sample size is twenty, and formal verification is not the same as mathematical proof. There is still considerable distance between a system that generates promising conjectures and one that resolves them.
Mathematics has always been the thing humans do that machines could not. The pipeline passes its own tests. Welcome to the next step.