Meta says Muse Spark helped mathematicians settle five open research questions
Meta AI Research on Friday published six papers from mathematicians working with Muse Spark 1.1 and 1.2 in Thinking Mode on meta.ai, saying five present answers to previously open research questions after a second mathematician group reviewed the work.
Competition math already has an answer key. Open research does not — and Meta is putting Muse Spark into that slower, messier loop with labeled AI drafts and human review, while admitting other teams hit some of the same doors independently.
On Friday, 2 October 2026, Meta AI Research published “Solving Open Research Problems Together.” The page is research.meta.ai. The line under the title says mathematicians and Muse Spark collaborated on six research papers. The page dates the post October 2, 2026, and marks it as an 8 minute read. It does not print an hour. The author line is Meta AI Research. It does not name a person. Muse Spark, on this page, is the model the mathematicians worked with. Those lines are Meta’s.
Why Meta says it tried this. Earlier this year, the page says, its models reached gold-medal-level performance across five high-school Olympiad competitions, in mathematics, physics, and chemistry. Gold-medal-level is Meta’s phrase for that earlier competition work. The words “our models achieved gold-medal-level performance” are a link to a post on X. The blog does not print scores, a medal table, or the names of the five contests. Meta says that result raised a different question: whether a model could help scientists on problems that are still open. A contest problem can be very hard and still have an answer waiting. An open research problem does not. There is no guarantee a line of attack will work. Progress, the page says, means trying an idea, making a new mistake, fixing it, and sometimes starting over. Those lines are Meta’s.
How the mathematicians used the model. Over the past several months, Meta says, it partnered with mathematicians on problems in several areas of mathematics. They used Muse Spark 1.1 and Muse Spark 1.2 in Thinking Mode, through the regular chat on meta.ai. Thinking Mode is the name of the setting on that chat. The page does not describe what the setting changes inside the model. There was no custom research scaffold. A scaffold, here, would be extra software wrapped around the chat to run a special research pipeline. They used the ordinary chat box. The goal, Meta says, was not to mass-produce papers. It was to help researchers develop mathematical insights that other people can understand and build on. Those lines are Meta’s.
The four rules printed on the page. A team of mathematicians guided the research and worked with Muse Spark to explore ideas and develop the arguments. A second group of mathematicians then reviewed the work. Each paper marks which passages were primarily drafted by researchers and which were drafted by AI. Each paper credits the earlier research and the mathematical ideas it builds on. Those four sentences are Meta’s. The page does not say Muse Spark wrote the papers on its own.
What they are sharing. Six papers from that collaboration. Five, Meta says, present answers to previously open research questions. After the work was finished, Meta learned that other teams outside Meta had independently announced solutions to some of the same problems, using different approaches. The papers acknowledge those concurrent works and how they relate. Those lines are Meta’s. The page does not say every one of the six was announced elsewhere first.
The probability paper. The title is “The Strict Threshold for Gaussian Ellipsoid Fitting.” Aykut Arslan with Muse Spark via meta.ai. The reviewers named on the page are Babak Modami, Alexander Roitershtein, Mark Sepanski, and Grigory Sokolov. The question is about fitting random Gaussian points, in high dimensions, to an ellipsoid. A Gaussian point is one drawn from a bell-curve cloud. An ellipsoid is a stretched sphere, the higher-dimension version of an ellipse. The page’s low-dimension picture is dots scattered on a plane, and an ellipse centered at a fixed point that has to pass through all of them. The result is a sharp threshold for how many points can be fitted that way. Below the threshold, such an ellipsoid exists with high probability. Above it, such an ellipsoid almost certainly does not. That is a theoretical benchmark for the limits of exact fitting. What happens exactly at the threshold is still unresolved. The figure on the page puts the switch near n approximately equal to d squared over four: an exact fit is likely below that line, and not above it. The figure does not define those letters in a sentence. In this problem, n is the number of points and d is the number of dimensions. Proof strategies were developed and revised with help from Muse Spark, under Arslan’s guidance, while four other mathematicians on the team checked and refined the arguments. Meta also points to three independent works posted in August 2026. Misiakiewicz and Wen proved the Gaussian threshold. De la Cerda, Potechin, Tulsiani, and Xu established it up to a vanishing multiplicative factor, Meta’s phrase for the same cutoff except for a multiplier that draws closer to 1. Koehler and Sohn obtained a broader universality result that includes the Gaussian threshold as a special case. Universality, here, means the result covers more than the bell-curve points, and the Gaussian cutoff falls out as one case. Meta says these works and its own were developed independently and use different approaches.
The differential-equations paper. The title is “Finite-Time Blow-Up of Radial Negative-Energy Solutions for the Mass-Critical Biharmonic Nonlinear Schrödinger Equation.” Leonard Dinh with Muse Spark via meta.ai. Review by Fazel Hadadifard and Salem Selim. The question is about wave collapse in a model inspired by laser physics. Picture a tug-of-war. One effect squeezes a wave inward. Another spreads it out. Can the wave keep concentrating forever, or must it eventually collapse? For waves that are symmetric around a center and have negative energy, in two or more dimensions, the paper proves that collapse happens within a finite time. Symmetric around a center is what radial means in the title. Blow-up means the wave becomes infinitely tall in that finite time. The figure on the page shows the profile getting narrower and taller as time approaches that moment. Meta says this settles a question left open in 2015 and confirms a prediction from computer simulations in 2002, for this setting. The page does not name those earlier papers. Muse Spark helped work through calculations, test possible arguments, and revise the proof. Dinh chose the problem and the key proof ideas. A separate pair of mathematicians reviewed the work and helped refine it. The long name picks which equation they studied. A nonlinear Schrödinger equation is a standard way to write how a wave evolves. Biharmonic and mass-critical are the names of this version. The page explains the question with the tug-of-war, not with a formula.
The group-theory paper. The title is “Semiabelian Groups Need Not Be Monomial.” Joseph Phillip Brennan and Milana Golich with Muse Spark via meta.ai. Review by Andres Barei and John Portin. A group, the page says, is a mathematical structure used to describe symmetry. The paper disproves a conjecture M. Kida proposed in 2024: that every finite group with a property called semiabelian must also have a property called monomial. The page does not define those two properties past the claim that one was supposed to force the other. One exception is enough. Meta’s image for that is a black swan: finding one disproves the claim that all swans are white. The exception is a group with 384 elements. Order 384 means the group has 384 members. The figure on the page labels the example SmallGroup(384, 20127), the kind of index the software GAP uses for one group of that size, and draws cycles in blue, red, green, and navy. The two properties do not always travel together. That, Meta says, helps mathematicians see how these groups are classified. Muse Spark generated the search program in GAP, which the page calls a mathematical software system, that found the counterexample. Golich and her collaborators verified the result and completed the argument. Two other mathematicians reviewed the work. Meta also acknowledges an AI agent called Nilradical, which reported a different counterexample to the same conjecture on September 16, 2026. Meta says its own result was developed independently.
The optimization paper. The title is “Tightness of the Cycle-Based Relaxation for Completed Length-Three Alpha-Cycles.” Aykut Arslan with Muse Spark via meta.ai. Review by Kien Trung Le. The question was first posed by Del Pia and Khajavirad in 2026: when a relaxation of a binary polynomial optimization problem captures the original exactly. A binary choice is yes or no. A relaxation is a simpler stand-in for a harder problem. Tight means the stand-in matches the original, with nothing left over. The page’s picture is three overlapping circles, like a Venn diagram, each holding a set of yes-or-no decisions. For the family they study, the approximation is exact when each region shared by two circles, but not the third, contains exactly one decision. If any of those regions contains more than one, the approximation leaves a gap. The figure shows the blue approximation matching the orange exact region on the left, and including extra possibilities on the right. That is a rule for when this particular simplification loses nothing, and when it needs to be stronger. Muse Spark helped reframe the problem using probabilities, identify a counterexample, and develop the proof strategy. Arslan and Trung Le checked the arguments, corrected gaps, and refined the final proof.
The arithmetic-physics paper. The title is “String Two-Point Function = Height Function on a Curve.” Anindya Dey, Gabriel Herczeg, An Huang, Nicolas Jaramillo Torres, and Jacob H. Swenberg with Muse Spark via meta.ai. The page does not name a separate review pair on this paper’s byline. Meta’s count is that five of the six papers present answers to previously open research questions. This write-up does not use that wording. It says Muse Spark helped the researchers connect two fields, number theory and p-adic string theory, along a direction Yuri Manin envisioned in the 1980s. They started from a connection already known for the Tate curve. The model helped the team see how that connection could extend to a much broader class of curves. The paper shows that two calculations, written in different mathematical languages, describe the same quantity. In simpler cases, the calculation comes down to counting how many initial digits the coordinates of two points share in a base-p number system. Base-p means the digits are written in a number system whose base is a prime. Sharing the opening digits is the page’s picture of closeness in that system. The figure is a graph of a genus-two Mumford curve: two central loops, with trees branching outward. Beyond finding the connection, Muse Spark generated candidate proofs and drafted three core technical sections. The researchers then checked, corrected, and refined those drafts.
The algebra paper. The title is “On Solvable Evolution Algebras and a Conjecture by García-Martínez and Pérez-Rodríguez.” Andres Barei with Muse Spark via meta.ai. Review by Nicolás Jaramillo Torres. An evolution algebra, the page says, is a mathematical structure inspired by evolutionary biology. The section heading calls the area non-associative algebra. Non-associative means the grouping in a product can change the answer, so (a times b) times c need not equal a times (b times c). García-Martínez and Pérez-Rodríguez suggested a test for a class called solvable algebras. The paper disproves that proposal. The team found a small three-dimensional example that passes their test and still does not belong to the class. Three-dimensional means the example is built from three independent pieces. It is a small case. The paper goes further than the counterexample. It gives an alternative rule that looks at whole subspaces, not single elements. A subspace is a slice of the algebra, not one member of it. Meta says that is a more accurate way to understand these structures. Working from the researchers’ prompts, Muse Spark generated the counterexample and proposed alternative characterizations and proofs. Barei checked, refined, and rewrote the material. Meta also acknowledges independent work by Hu and Wen, who reported counterexamples to the same conjecture.
What Meta says this is for. The closing section says the work continues Meta’s investment in scientific research. Progress, it says, will require close collaboration between experts and AI, with results that are carefully verified, transparent about how they were produced, and clearly communicated to the wider research community. The acknowledgments thank the researchers who brought their questions, everyone who checked proofs and helped make the papers clearer, and the broader mathematics community whose earlier work these results build on. Those lines are Meta’s.
The picture is Meta’s card for the post. Blue paths start apart and bend into one loop. The card reads Meta AI Research and Muse Spark, then Open Research, with a line that six papers sit beside five open problems. The lower left reads October 2, 2026, the same day as the post. The lower right reads Thinking Mode and human review. The frame does not print a clock. It is the release graphic. It is not a photograph of a proof on a board.
In plain terms, Meta AI Research said on Friday that mathematicians working with Muse Spark 1.1 and 1.2, in Thinking Mode, on the ordinary meta.ai chat, are sharing six papers. Five of them present answers to questions that had been open. A second group of mathematicians reviewed the work. Each paper labels which passages the researchers drafted and which the model drafted, and each one credits earlier research. The six topics are fitting a cloud of random points to an ellipsoid, proving that a certain wave must collapse in finite time, a 384-element exception in the theory of symmetry, when a simpler optimization problem matches the real one, a link between a string-theory calculation and a number-theory one, and a small counterexample in an algebra inspired by evolution. Meta says other teams independently announced solutions to some of the same problems, with different approaches, and the papers say so. The page does not say the model did the mathematics alone.