An open-source AI system just completed original mathematics research — without a human in the loop after the initial question.
QED is a multi-agent pipeline that takes a research-level math question and produces a complete proof. It splits the work across three agent types: one decomposes the problem into a proof structure, others generate candidate arguments, and a final layer checks correctness. The team tested it on 18 research projects spanning algebraic geometry, fluid PDEs, probability, and inverse problems. Five resulted in original works; three were assessed by domain experts as comparable in difficulty and scope to papers published in established specialist mathematics venues.
Math has long been the stress test for AI reasoning — it is unambiguous, verifiable, and hard to fake. Where prior systems mostly verified human-written proofs or solved competition problems with known solutions, QED is aimed squarely at open problems, the kind where no answer exists yet. That is a meaningful shift in ambition, and the expert sign-off on three of the five outputs is the hardest part to hand-wave away.
The obvious caveat: 18 problems is a small sample, and the source does not detail how the 18 were selected — cherry-picking is a real risk in any early benchmark. Still, the code is public, so the math community can pressure-test the claims directly.