JOURNAL / 2026.10.03

Muse Spark contributes to six mathematics papers: what it added and what needs checking

Meta publishes work done by mathematicians with Muse Spark. Two cases show concrete contributions to search and proof, while review, priority and the cost of obtaining them remain separate questions.

On October 2, Meta published six mathematics papers developed with Muse Spark, five of which present answers to previously open questions, according to its announcement. Researchers used versions 1.1 and 1.2 in Thinking Mode through the regular meta.ai chat interface, without a purpose-built research system. The news is not a newly released model: it is published evidence of what researchers managed to do with models already available.

The results span probability, differential equations, group theory, optimization, arithmetic physics and non-associative algebra. They do not all tell the same story about AI's contribution. Two deserve a closer look because they let us follow the work beyond the headline.

A search that ends in a checkable object

In Semiabelian Groups Need Not Be Monomial, Joseph Phillip Brennan and Milana Golich present a counterexample to a conjecture by Kida about two properties of finite groups. A group is a mathematical structure that describes, among other things, symmetries. The conjecture asserted that a particular way of constructing it guaranteed another property of its representations. Finding one group that satisfies the first and fails the second is enough to disprove it.

The paper identifies that object in GAP's SmallGroups library: a group with 384 elements. Muse Spark helped generate the code that extended the search; the authors ran the program and documented the check through a decomposition of the group. The manuscript includes algorithms, code and output, and marks human and AI-assisted material. This leaves a concrete contribution another researcher can inspect: the model helped turn a question into an executable search.

An earlier result changes how this news should be told. The paper acknowledges that Nilradical announced another counterexample on September 16. Its repository pinned to a specific revision describes a group with 2,592 elements, a formal Lean proof and verification checks with a stated scope. I have not run those checks. An earlier solution means Meta's announcement cannot be presented as the first public disproof; the different counterexample and computational route retain their own interest.

An unresolved alternative in wave evolution

Leonard Dinh's work on the biharmonic nonlinear Schrödinger equation offers another kind of contribution. It claims that, in the mass-critical regime studied, radial solutions with negative energy and initial data in H² develop a singularity in finite time, in dimensions two and above. Radial symmetry and those conditions are part of the result; this is not a statement about every wave.

The difference from the previous state of knowledge is precise. Boulenger and Lenzmann's work, first submitted in 2015, allowed two possibilities in the case without lower-order dispersion: a finite-time singularity or unbounded growth as evolution continues indefinitely. Dinh claims to rule out the second through a new exponential localization argument. The change closes a mathematical alternative; it does not observe a physical explosion or solve Navier–Stokes.

The manuscript credits the model with exploring ideas, developing candidate arguments and drafting; researchers directed, corrected and took ownership of the final text. It names two mathematical reviewers. This is a proof submitted for examination, not independent validation performed by this journal or a guarantee of journal acceptance.

My reading is that these cases make useful research collaboration more credible: writing a search and helping develop a proof are substantive tasks, even when they depend on someone who knows how to choose the question and spot errors. The ordinary interface lowers an access barrier; it does not remove the expertise needed to evaluate the answer.

What is missing to turn these cases into a capability measure is the denominator: how many problems were attempted, how many routes failed, how much time review consumed, and what the same team would have achieved without the model. Six selected manuscripts do not answer those questions. Nor do they let us attribute the result to the current version of Muse Spark or compare it with other systems.

For a laboratory adopting this method, I would prioritize tasks that produce artifacts others can challenge and review: code, an explicit counterexample or a lemma with clear hypotheses. Provenance markers help attribute the work; correctness requires examining its arguments. The value of this publication lies in offering objects and proofs others can work on, with visible human credits and prior results. That is a more useful foundation for the next investigation than a general proclamation about machines doing mathematics.

← Back to journal