On October 2, 2026, Meta’s research team published six papers co-developed with mathematicians using Muse Spark — and the setup matters more than the headline number. Five of the papers answer previously open research questions. The other resolves a 2024 conjecture the other way, by finding a counterexample.
The constraint that makes this interesting: no custom research scaffold. The mathematicians used Muse Spark 1.1 and 1.2 in Thinking Mode through the regular meta.ai chat interface, the same product surface any user gets.
Open problems are a different test than Olympiads
Earlier in 2026, Meta’s models hit gold-medal-level performance across five high-school Olympiad competitions. That result motivated this project. But as the Meta team points out in their announcement, competition problems — however brutal — have a known solution. Open research has no answer key and no guarantee any approach works. Progress means trying ideas, making new mistakes, and sometimes starting over.
That distinction matters if you’re building expert-facing AI tools. A benchmark score on solved problems tells you less about usefulness in the unsolved regime than most product decks imply.
The workflow beats the model
Look at how the papers were actually produced, because the process repeats across all six:
- A lead mathematician chose the problem and guided the exploration with Muse Spark.
- The model helped develop proof strategies, generate search programs, work through calculations, or draft technical sections — varying by paper.
- A separate group of mathematicians reviewed and refined everything.
- Each paper marks which passages were primarily drafted by researchers versus AI, and credits earlier work it builds on.
Two examples show the range. In group theory, Muse Spark generated a search program in GAP that found a 384-element counterexample disproving Kida’s 2024 conjecture that semiabelian groups must be monomial — one black swan settles it. In optimization, the model reframed a problem in terms of probabilities, which surfaced a counterexample and a proof strategy that human reviewers then corrected and tightened.
The division of labor is consistent: the model proposes and generates, the human decides and verifies. That’s not a limitation of the tooling; it looks like the working design.
Independent discovery as a validation signal
A detail worth noticing: after finishing, Meta learned that outside teams had independently announced solutions to some of the same problems using different approaches. On the Gaussian ellipsoid fitting threshold, three concurrent works from August 2026 are acknowledged. An AI agent called Nilradical reported a different counterexample to the same semiabelian conjecture on September 16, 2026.
Convergent independent findings don’t prove correctness, but they’re a useful signal in a space where hallucinated proofs are a real failure mode. The papers acknowledge these overlaps explicitly rather than treating them as competition.
What this means if you’re building expert tools
Meta’s stated goal wasn’t mass-producing papers but helping researchers develop insights others can build on. For product builders, the practical lessons are modest and concrete. First, a general-purpose chat interface plus a strong model was sufficient — the leverage came from pairing genuine domain experts with verification discipline, not from bespoke agent infrastructure. Second, provenance features (drafted-by labels, credit to prior work) shipped as part of the workflow, not bolted on after.
That second point echoes something I keep coming back to: the teams making AI useful in high-stakes domains are the ones designing the verification loop first. It’s the same instinct behind putting safety gates before a training run rather than before deployment, as in Meta’s earlier framework update — decide how you’ll check the output before you generate it.
One caveat: these results are in mathematics, where verification is unusually tractable — a proof either checks or it doesn’t. Domains with fuzzier ground truth won’t transfer this workflow as cleanly. If you’re building for experts in those fields, the open question is what plays the role of the reviewing mathematician.