Meta Shares Six AI-Assisted Math Papers, Saying Five Answer Open Questions

An AI chatbot helped mathematicians search for counterexamples, develop proofs and connect ideas across fields—not just answer problems with known solutions. On October 2, Meta AI Research shared six papers developed with Muse Spark, saying five present answers to previously open research questions. The work paired model-generated ideas with human direction and review.

A chat window, with mathematicians steering

The researchers used Muse Spark 1.1 and 1.2 in Thinking Mode through the regular meta.ai chat interface. There was no custom research scaffold: no purpose-built system surrounding the model to manage the research. Meta says the collaborations ran over several months, with mathematicians guiding the exploration and development of arguments.

That setup did not make the work autonomous. The division of labor varied by paper. In one project, the model wrote a search program; in another, it drafted three core technical sections. Researchers checked, corrected and refined those contributions. Meta describes a second group of mathematicians reviewing the work, rather than treating the model’s output as a finished result.

A search program finds an exception

The group-theory paper gives a concrete example of the mechanism. Groups are mathematical structures used to describe symmetry. A conjecture proposed by M. Kida in 2024 said every finite group with a property called semiabelian also had a property called monomial. One exception would be enough to disprove it.

Muse Spark generated a search program in GAP, a mathematical software system. It found a 384-element group that breaks the conjecture, according to Meta. Milana Golich and her collaborators verified the result and completed the argument, with two other mathematicians reviewing the work. The model supplied the search tool; people established the result.

Proofs with precise boundaries

Another paper tackles whether random points in high-dimensional space can all lie on an ellipsoid. Meta says the team identified a sharp threshold: below it, an exact fit exists with high probability; above it, an exact fit almost certainly does not. The behavior exactly at the threshold remains unresolved.

A wave-collapse paper addresses a model inspired by laser physics. For waves symmetric around a center, with negative energy in two or more dimensions, it proves collapse occurs within finite time. The equation studied is the mass-critical biharmonic nonlinear Schrödinger equation. Meta says this settles a question left open in 2015 for that setting.

Muse Spark helped with calculations, candidate arguments and proof revisions. Mathematician Leonard Dinh selected the problem and key proof ideas; two reviewers helped refine the work. Those details make the claim narrower and more informative than saying a chatbot independently solved wave collapse.

Three more results—and overlapping discoveries

The optimization paper identifies when a particular simplified version of a yes-or-no decision problem preserves the exact answer, and when it leaves a gap. Muse Spark helped recast the problem using probabilities and develop the proof strategy. Aykut Arslan and Kien Trung Le checked the arguments and corrected gaps.

The arithmetic-physics paper connects calculations in number theory and p-adic string theory, showing they describe the same quantity. The evolution-algebra paper supplies a three-dimensional counterexample to a proposed classification test, then offers an alternative rule based on whole subspaces rather than individual elements.

These are not all claims to sole discovery. Meta acknowledges independent work on some of the same problems, including three concurrent ellipsoid-fitting works posted in August and a different group-theory counterexample reported by the AI agent Nilradical on September 16. It says the overlapping results were developed independently using different approaches.