Meta Publishes 6 Math Papers With Muse Spark, Closing 5 Open Problems

Meta announced on October 2, 2026 that a team of mathematicians produced six research papers working alongside Muse Spark — and that five of them answer previously open questions. The detail that changed the conversation was not the result itself but the route: all of it went through the ordinary Meta.ai chat interface, the same one anyone uses, with no custom research scaffolding.

Quick answer: what did Meta do?

Mathematicians worked for months with Muse Spark 1.1 and 1.2 in Thinking Mode, through the standard Meta.ai chat, across six areas: probability, differential equations, group theory, optimization, arithmetic physics and non-associative algebra. The result was six papers, five of them closing open problems. Humans picked the problems, directed the exploration and verified every argument — the model did not work on its own.

The clearest case: a 384-element counterexample

The example that best explains the method involves a group-theory conjecture proposed by M. Kida in 2024. Muse Spark wrote a search program in GAP (a computational algebra system) that found a 384-element counterexample. The mathematicians verified it — and the conjecture fell.

Notice what happened there: the model did not “prove” anything by intuition. It translated the problem into search code, ran it and returned the object that breaks the claim. That is exactly the kind of task a computer has always been better at than a person; what changed is that translating the problem into the code now comes from the machine too.

What the announcement proves — and what it doesn’t

Why this matters

The practical point is not the mathematics — it is the tool. Meta is saying that the difference between the chat you open on your phone and the system that closes a group-theory problem is the person on the other side, not a secret version of the model. Someone who can frame the problem well, break it into verifiable steps and check every answer gets results from an ordinary chat that, as recently as 2024, would have required a lab.

It is also a useful counterpoint to the week in which OpenAI paused training over rogue agents and the White House stood up an AI task force: while the public debate circles risk, the fastest-moving real use is the boring, checkable kind — generate search code, test a hypothesis, verify the result.

Frequently asked questions

Did Muse Spark solve the problems by itself?

No. Mathematicians chose the problems, led the exploration and verified every argument. The papers describe bounded model contributions, not autonomous discoveries.

Which model version was used?

Muse Spark 1.1 and 1.2 in Thinking Mode, through the standard Meta.ai chat interface — with no custom research system.

Which areas of mathematics were covered?

Probability, differential equations, group theory, optimization, arithmetic physics and non-associative algebra.

Can this be reproduced with a normal account?

The infrastructure is the same: Meta.ai chat in Thinking Mode. What is not common is the human half — picking a problem that matters and knowing how to verify the answer.

At DigitalRadar, we separate what AI actually did from what was announced. Stay on the radar.