Meta puts Muse Spark in the maths race with six papers

Meta is staking a claim in the contest over which lab's models can do real research, not just pass exams. On 2 October 2026 the company released six mathematics papers produced by mathematicians working with Muse Spark 1.1 and 1.2 in Thinking Mode.

The detail that matters commercially is the tooling. According to Meta, the work ran through the regular meta.ai chat, with no custom research scaffold. The same product consumers use, in other words, was the research instrument. Five of the six papers answer previously open research questions. All were peer reviewed, and each paper marks which passages the AI drafted and which the researchers wrote.

The papers span a wide range. One sets a sharp threshold for fitting random Gaussian points to an ellipsoid in high dimensions. Another shows finite-time blow-up, or wave collapse, for radial negative-energy solutions of the mass-critical biharmonic nonlinear Schrödinger equation. A third, "Semiabelian Groups Need Not Be Monomial", offers a 384-element counterexample that disproves a 2024 conjecture by Kida.

The remaining three cover optimisation, arithmetic physics and algebra. The optimisation paper addresses the tightness of the cycle-based relaxation for completed length-three alpha-cycles. The arithmetic physics paper shows a string two-point function equals a height function on a curve, linking number theory and p-adic string theory. The last gives a three-dimensional counterexample to a conjecture by García-Martínez and Pérez-Rodríguez on solvable evolution algebras.

Meta says the model developed proof strategies, worked through calculations, wrote search programs and drafted technical sections. Humans set the direction and verified the results. Meta's research blog describes the process this way: "Making progress means trying and retrying ideas, making new mistakes and resolving them, and sometimes going back to the beginning."

The release follows earlier results. Muse Spark had already reached gold-medal level on five high-school olympiad competitions. Muse also recently hit #1 on the App Store.

The timing is deliberate. OpenAI and Google DeepMind have both published AI-assisted maths results this year, so Meta is matching rivals on a benchmark that investors and enterprise buyers read as a proxy for reasoning ability. Pairing a consumer chart-topper with peer-reviewed research gives Meta a story its competitors' lab-only demonstrations lack.

The strongest part of the pitch is the absence of special infrastructure. If the results came from the public chat product, Meta can argue its consumer model and its research model are the same asset. The per-passage disclosure of AI-drafted text also gives sceptical reviewers something to check. Watch whether Meta turns this into a product for researchers, and whether rivals answer with their own accounts of how much of the work the model did.