Meta publishes six math papers made with Muse Spark and human mathematicians

Meta published six research papers on October 2nd describing mathematicians' work with its Muse Spark 1.1 and 1.2 models in Thinking Mode through the regular Meta AI chat interface, with Meta saying five of the papers answer previously open questions. The papers credit named researchers including Aykut Arslan, Leonard Dinh, Joseph Phillip Brennan and Milana Golich, and report results such as a group with 384 elements that disproves a group-theory conjecture and a sharp threshold near n = d squared divided by four for fitting random Gaussian points to an ellipsoid. Meta acknowledged that other teams independently announced solutions to some of the same problems, including work on the Gaussian ellipsoid threshold posted in August 2026 and a separate group-theory counterexample reported by the AI agent Nilradical on September 16th.

Meta publishes six math papers made with Muse Spark and human mathematicians Researchers used Muse Spark 1.1 and 1.2 through Meta AI's regular chat interface; Meta says five papers answer previously open questions. By Ryan Merket https://runtimewire.com/author/ryan-merket · Published Primary source: X https://x.com/AIatMeta/status/2106099776035152231 Why it matters The papers offer concrete examples of Muse Spark assisting with proofs, code, and mathematical arguments, while showing how much the results depended on researcher direction and review. That context helps readers assess the model's role in the research. Meta https://about.fb.com/news/2026/04/introducing-muse-spark-meta-superintelligence-labs/ published six research papers on October 2nd describing mathematicians' work with its Muse Spark models, including proofs addressing five questions the company says had been open. According to Meta's account https://research.meta.ai/blog/solving-open-research-problems-together , the collaboration used Muse Spark 1.1 https://runtimewire.com/models/native-meta/muse-spark-1.1-7789041b785c52db and 1.2 in Thinking Mode through the regular Meta AI chat interface https://meta.ai/ . The papers identify the researchers and their responsibilities. Aykut Arslan worked with Muse Spark on two papers, one in probability and another in optimization. Leonard Dinh led work on a differential-equations problem; Joseph Phillip Brennan and Milana Golich worked on a group-theory conjecture; and other researchers tackled arithmetic physics and non-associative algebra. Meta says mathematicians chose and guided the research, while separate mathematicians reviewed the work. Each paper marks passages drafted primarily by researchers or AI. The papers describe an assistant contributing to research under expert direction, including by generating search code, proposing proof approaches, and drafting sections. They do not describe a system choosing its own research agenda or independently validating its results. In one example, Muse Spark generated a search program in GAP that found a counterexample to a group-theory conjecture; researchers checked the example and completed the proof. In another, it generated a counterexample and proposed alternative characterizations for a conjecture about evolution algebras, which a researcher then refined. The results range from concrete counterexamples to technical proofs. In probability, Arslan's paper identifies a sharp threshold, near n = d squared divided by four, for whether random Gaussian points can be fit exactly to an ellipsoid. The paper explicitly leaves behavior at the threshold unresolved. In a wave-equation problem, Dinh's paper proves finite-time blow-up for a specified class of radial solutions, addressing a question left open in 2015. In group theory, Brennan and Golich exhibit a group with 384 elements that disproves a conjecture asserting that a certain class of groups must have another property. The three other papers address when a particular optimization simplification exactly captures its original problem, extend a connection between number theory and string theory calculations, and disprove a conjecture about evolution algebras while offering an alternative characterization. Meta says the optimization paper answers a question posed in 2026. For the arithmetic-physics paper, the company says Muse Spark helped researchers extend a connection first developed for the Tate curve. It also says the model drafted three technical sections that researchers checked and revised. Meta's account also qualifies the novelty of the batch. Other teams independently announced solutions to some of the same problems, using different approaches. The company specifically acknowledges independent work on the Gaussian ellipsoid threshold posted in August 2026, a separate counterexample to the group-theory conjecture reported by the AI agent Nilradical on September 16th, and other independent work on the evolution-algebra conjecture. Those acknowledgments make the papers a less tidy demonstration of six exclusive AI discoveries. They also clarify the contribution claims: the papers describe how the model assisted researchers while crediting work that arrived independently. Meta introduced Muse Spark in April 2026, released version 1.1 in July, and described version 1.2 in August. On September 2nd, it announced Muse Spark 1.3 https://runtimewire.com/models/meta/muse-spark-1.3 . The October papers report use of earlier versions, 1.1 and 1.2, over the preceding months; they do not establish that version 1.3 produced these results. Meta says researchers used the ordinary chat interface, distinguishing this work from demonstrations built around custom tools or a bespoke research workflow. Meta presents the project as a test of whether a general-purpose assistant can help with problems lacking an answer key. For the company, that expands the case for Muse Spark beyond consumer assistance and coding into scientific work. The evidence is a set of six papers with explicit human-review and AI-use disclosures, not a measure of how often the model can produce valid research or how much time it saved researchers. Meta has not attached a cost, researcher count, or productivity measure to the collaboration in its announcement.