Meta Says Muse Spark Helped Mathematicians Settle Five Open Problems — Using Nothing but the Regular Chat Window

Meta AI Research on Thursday published six mathematics papers developed with its Muse Spark models, and the company says five of them answer research questions that had been open until now. The problems span probability, differential equations, group theory, optimization, arithmetic physics and non-associative algebra — and the striking part is not just the results, but the setup: the mathematicians involved used Muse Spark 1.1 and 1.2 in Thinking Mode through the regular meta.ai chat interface, with no custom research system wrapped around the model.

"Open research is different. There is no answer key, no guarantee that an approach will work," Meta wrote in its blog post, drawing a line under the olympiad-style gold medals its models had already earned on problems with known solutions. "Making progress means trying and retrying ideas, making new mistakes and resolving them, and sometimes going back to the beginning."

The clearest example of the division of labor comes from the group theory paper. A 2024 conjecture by M. Kida proposed that every finite semiabelian group must also be monomial. Muse Spark generated a search program for GAP, the standard computational algebra system, and the program surfaced a 384-element counterexample. Human mathematicians then verified the result and completed the argument — the model supplied the search tool, people established the theorem. Meta also notes that an independent AI agent called Nilradical reported a different counterexample to the same conjecture on September 16.

The other results are similarly concrete. In probability, the team identified a sharp threshold for fitting random Gaussian points in high dimensions onto an ellipsoid: below it, an exact fit exists with high probability; above it, almost certainly not — though behavior exactly at the threshold remains unresolved, and three independent groups posted related work in August 2026. In differential equations, a paper proves that certain symmetric negative-energy waves in a laser-physics-inspired model must collapse in finite time, settling a question open since 2015. A third paper improves an approximation bound in polynomial optimization from 5.2 to 3.52 times a baseline measure, and an arithmetic-physics result connects a number-theory idea with a string-theory calculation, with Muse Spark drafting three core technical sections that researchers then checked and refined.

Meta has built unusually explicit attribution into the papers. Each one marks which passages were primarily drafted by researchers and which by the model, credits the earlier work it builds on, and was reviewed by a second group of mathematicians rather than treated as the model's output. One team guided the research and developed the arguments; another checked them. It is a documented collaboration, not a claim of autonomous machine mathematics — and Meta is not claiming exclusivity, acknowledging that several of the same problems were independently solved by other groups using different approaches.

The announcement lands in a week crowded with AI-does-science claims, and the comparisons are instructive. Google Research separately described Cogentic, a multi-agent harness running on Gemini that produced new proofs on five open problems in online learning, auction theory and mechanism design — a system-level approach with orchestrators and adversarial verifiers, in contrast to Meta's bare chat window. OpenAI's earlier claim that its model solved the Navier-Stokes Millennium Problem, by contrast, was challenged by two researchers who said the company may have built on their work.

For the field, the signal is that frontier chat models are starting to contribute to genuine research without bespoke scaffolding — but the contributions remain bounded. Meta's papers describe model-generated search programs, calculations and draft sections, all steered and verified by humans, with company-reported claims that have not been through third-party audit. The bottleneck is still expert direction. What changed this week is that the tool doing the grunt work now sits in a chat window anyone can open.