Back

Meta says mathematicians used its Muse Spark model to answer five open research questions

Illustration: five question marks fading into check marks above five geometry proofs
Illustration: FrameSignal (AI-generated)

Meta says mathematicians working with its Muse Spark AI model answered five open research questions, problems with no known answer, unlike the math contests its models have done well on before. Outside teams reached some of the same results on their own, one as early as Aug. 10. Meta has not said how many problems the model was tried on without success.

Meta says it wanted to see whether its AI model could help with open research math, not just contest problems. On Friday, Oct. 2, the company published six math papers written by mathematicians working with its Muse Spark model, and says five present answers to previously open research questions. Two of the papers disprove conjectures, claims other mathematicians had proposed but not proved. Meta says this is a different kind of test from its earlier gold-medal-level results on high-school Olympiad competitions, since contest problems already have a known solution and open problems do not.

Meta says the mathematicians used Muse Spark 1.1 and 1.2 in Thinking Mode through the regular meta.ai chat interface, with no custom research setup. Meta says a team of mathematicians guided the research and a second group reviewed it, though its post names no reviewers for one of the six papers. Each paper marks which passages researchers drafted and which the AI drafted.

The model's role varied. In a paper on groups, mathematical structures used to describe symmetry, Meta says Muse Spark wrote the search program that found a counterexample, a single case that proves a general claim wrong. It was a group with 384 elements that breaks a conjecture by mathematician M. Kida. In a paper linking number theory and a version of string theory, Meta says, it drafted three core technical sections that the researchers then checked and corrected. Other papers cover random data fitting, when a wave must collapse in a model drawn from laser physics, an optimization problem, and a type of algebra inspired by evolutionary biology. The data fitting paper leaves open what happens exactly at the cutoff it finds.

Others had already made public answers to three of the five questions. Meta says it learned after finishing that outside teams had independently announced solutions to some of the same problems. On Aug. 10, Theodor Misiakiewicz and Garrett G. Wen posted a proof of the same data fitting result: the point past which random points almost certainly can't all be fitted to an ellipsoid. Meta acknowledges theirs and two other outside papers from August on that problem. Meta's post also credits an AI agent called Nilradical, which reported a different counterexample to Kida's conjecture on Sept. 16. And Xing-Yu Hu and Ran Wen posted counterexamples to the conjecture in the evolution algebra paper before Meta's papers came out.

It is not yet known how many problems the mathematicians tried with Muse Spark without success, or whether the papers will go through journal peer review. The reviewers are part of Meta's collaboration, and Meta's post does not say whether the papers have been submitted to journals.

More on Meta