Meta Superintelligence Labs published a research post on October 2, 2026: mathematicians used Muse Spark 1.1 and 1.2 in Thinking Mode through ordinary meta.ai chat, without a custom research scaffold, and shipped six papers. Five present answers to previously open questions. AI at Meta amplified the announcement on X; Alexandr Wang quote-posted that announcement. explainx.ai could not independently retrieve Wang's caption text before publication — treat the amplification as a distribution fact, not as a quoted technical claim.
The practitioner angle is not "AI did research math." It is whether a consumer chat surface plus expert mathematicians plus a second review bench is enough to move open problems into papers with transparent AI credit. That sits next to AGMAI's responsible-release checklist and the same expert-attention lesson as Breen's archive loop with Opus 5.5: models draft and search; humans choose and own.
Primary source: Solving Open Research Problems Together (Meta AI Research, October 2, 2026).
Update — October 3, 2026: A separate Google Research story — Gemini-based Cogentic reports five open results in learning and auctions via a multi-agent prove–verify harness, not meta.ai chat. Do not merge the headlines: Cogentic five math problems.
TL;DR — what changed, what to believe
| Question | Direct answer |
|---|---|
| What shipped? | Six Muse Spark–assisted math papers; five answer open questions |
| When? | Meta research post dated October 2, 2026 (this explainx.ai coverage: October 3) |
| Which models? | Muse Spark 1.1 and 1.2 in Thinking Mode — not claimed as 1.3 |
| Interface? | Regular meta.ai chat; Meta says no custom research scaffold |
| Who owns the work? | Named mathematicians guided; a second group reviewed |
| Transparency? | Papers mark researcher-drafted vs AI-drafted passages; credit prior work |
| Exclusive? | No — Meta acknowledges independent concurrent solutions on several problems |
| Builder takeaway? | Treat this as a lab collaboration report, not a harness bake-off win |
What Meta claims about the process
Meta frames the project as a step past olympiad gold medals: competition problems already have keys; open research does not. The collaboration principles, as stated in the research post, are:
- A team of mathematicians guided the research and worked with Muse Spark to explore ideas and develop arguments.
- A second group of mathematicians reviewed that work.
- Each paper marks which passages were primarily drafted by researchers and which were drafted by AI.
- Each paper credits earlier research and ideas it builds on.
After Meta finished its batch, the lab says other teams outside Meta independently announced solutions to some of the same problems with different approaches. The papers acknowledge those contributions. That sentence matters more than the headline count. Six papers is a publication event. Exclusive discovery is a different claim, and Meta is not making a clean exclusivity claim.
For builders, the process claim that changes product narrative is the scaffold sentence: Thinking Mode on the regular meta.ai chat interface, Muse Spark 1.1 and 1.2, no custom research scaffold. That distinguishes this write-up from demos that only work inside a private Lean loop, a bespoke retrieval stack, or a multi-agent research harness. It does not prove that any random user can open meta.ai and close an open problem. Expert direction is still the bottleneck — the same pattern as Claude-shaped science.
Meta also notes that Muse Spark 1.3 shipped earlier (September 2026 on explainx.ai's timeline). These papers report 1.1 and 1.2 over preceding months. Do not rewrite the story as a 1.3 showcase.
The six problems in plain language
Titles below match Meta's post and Wang's listed problem names. One-liners are explainx.ai paraphrases of Meta's own plain-language summaries — not theorem statements.
1. The Strict Threshold for Gaussian Ellipsoid Fitting
Field: Probability · Lead: Aykut Arslan with Muse Spark · Review named by Meta: Babak Modami, Alexander Roitershtein, Mark Sepanski, Grigory Sokolov
One-liner: There is a sharp cutoff on how many random Gaussian points in high dimension you can force onto one centered ellipsoid; below the cutoff a fit almost always exists, above it almost never.
Meta's post says the exact threshold behavior remains unresolved, and that three independent concurrent works posted in August 2026 also address the Gaussian threshold (Misiakiewicz and Wen; De la Cerda, Potechin, Tulsiani, and Xu; Koehler and Sohn), developed independently with different approaches.
Paper page (verified HTTP 200): The Strict Threshold for Gaussian Ellipsoid Fitting
2. Finite-Time Blow-Up of Radial Negative-Energy Solutions for the Mass-Critical Biharmonic Nonlinear Schrödinger Equation
Field: Differential equations · Lead: Leonard Dinh with Muse Spark · Review named by Meta: Fazel Hadadifard and Salem Selim
One-liner: For a laser-physics-inspired wave model, certain symmetric negative-energy waves in two or more dimensions must collapse in finite time — they cannot keep concentrating forever.
Meta says this settles a question left open in 2015 and confirms a 2002 simulation prediction for this setting. Muse Spark helped with calculations and revisions; Dinh chose the problem and key ideas.
Paper page (verified HTTP 200): Finite-Time Blow-Up…
3. Semiabelian Groups Need Not Be Monomial
Field: Group theory · Leads: Joseph Phillip Brennan and Milana Golich with Muse Spark · Review named by Meta: Andres Barei and John Portin
One-liner: A 2024 conjecture said every finite "semiabelian" group is also "monomial"; one counterexample group with 384 elements kills that claim.
Meta says Muse Spark generated a GAP search program that found the counterexample; Golich and collaborators verified and completed the argument. Meta also acknowledges the AI agent Nilradical, which reported a different counterexample to the same conjecture on September 16, 2026, developed independently.
Paper page (verified HTTP 200): Semiabelian Groups Need Not Be Monomial
4. Tightness of the Cycle-Based Relaxation for Completed Length-Three Alpha-Cycles
Field: Optimization · Lead: Aykut Arslan with Muse Spark · Review named by Meta: Kien Trung Le
One-liner: For a specific family of yes-or-no polynomial optimization problems, Meta's summary says a cycle-based relaxation is exact exactly when each pairwise-shared region in a three-set Venn picture holds one decision — more than one and the approximation leaves a gap.
Meta attributes the original question to Del Pia and Khajavirad in 2026. Muse Spark helped reframe the problem, find a counterexample, and develop strategy; Arslan and Trung Le corrected gaps.
Paper page (verified HTTP 200): Tightness of the Cycle-Based Relaxation…
5. String Two-Point Function = Height Function on a Curve
Field: Arithmetic physics · Authors named by Meta: Anindya Dey, Gabriel Herczeg, An Huang, Nicolas Jaramillo Torres, and Jacob H. Swenberg with Muse Spark
One-liner: Two calculations — one from number theory, one from p-adic string theory — describe the same quantity on a broader class of curves than the previously known Tate-curve case.
Meta presents this as an extension of a connection envisioned by Yuri Manin, not as "five of six open-problem answers." Muse Spark drafted three core technical sections that researchers checked and revised. This is the paper that is not sold as closing a previously open research question in Meta's five-of-six framing.
Paper page (verified HTTP 200): String Two-Point Function = Height Function on a Curve
6. On Solvable Evolution Algebras and a Conjecture by García-Martínez and Pérez-Rodríguez
Field: Non-associative algebra · Lead: Andres Barei with Muse Spark · Review named by Meta: Nicolás Jaramillo Torres
One-liner: A proposed test for "solvable" evolution algebras fails; a small three-dimensional counterexample passes the test but is not solvable, and the paper offers a better subspace-based rule.
Meta says Muse Spark generated the counterexample and proposed alternatives; Barei checked, refined, and rewrote. Meta acknowledges independent counterexamples reported by Hu and Wen.
Paper page (verified HTTP 200): On Solvable Evolution Algebras…
Paper links: what we verified, what we did not invent
| # | Title | Verified URL | arXiv ID in this article |
|---|---|---|---|
| 1 | The Strict Threshold for Gaussian Ellipsoid Fitting | ai.meta.com publication | Not claimed — no arXiv ID verified here |
| 2 | Finite-Time Blow-Up… Biharmonic NLS | ai.meta.com publication | Not claimed |
| 3 | Semiabelian Groups Need Not Be Monomial | ai.meta.com publication | Not claimed |
| 4 | Tightness of the Cycle-Based Relaxation… | ai.meta.com publication | Not claimed |
| 5 | String Two-Point Function = Height Function on a Curve | ai.meta.com publication | Not claimed |
| 6 | On Solvable Evolution Algebras… | ai.meta.com publication | Not claimed |
Hub post (verified HTTP 200): research.meta.ai/blog/solving-open-research-problems-together
If a future mirror lands on arXiv or a journal, update from those IDs — do not invent them from titles.
Credit and review split — the part that matters for builders
Read Meta's batch the way AGMAI wants you to read AI math releases. Named humans chose problems and steered. A second mathematician group reviewed. AI drafting is marked in the papers. Concurrent independent work is acknowledged. That is closer to a steered collaboration than to a secret-model leaderboard dump of the sort the Fields Medalists letter and OpenAI 100+ problems advisory criticized.
What Meta's announcement does not give you, at least in the public research post:
- A published prompt archive or thinking-trace dump for every paper
- A compute-cost or wall-clock productivity number
- A Lean or other formalization claim for the six results
- A success rate across a larger pool of attempted open problems
- Evidence that Muse Spark 1.3 produced these proofs
So the honest builder belief set is narrow:
Believe (as Meta's stated process for this batch): consumer meta.ai Thinking Mode with Muse Spark 1.1/1.2; mathematician guidance; second-group review; marked AI/human passages; credit for prior and concurrent work; six paper landing pages on Meta Research.
Treat as lab narrative until you read the papers yourself: that each proof is correct in every line; that the AI contribution was decisive rather than assistive; that "no scaffold" generalizes to your domain; that this beats Claude, GPT, or a Lean-heavy pipeline on research math.
Do not believe without evidence: that Muse Spark autonomously selected problems, that six exclusive discoveries belong only to Meta, or that you should rip out your coding agent because research math moved.
For formalization cost and why a kernel check is not the same as understanding, see Lean 4 formalization cost collapse. For how easy it is to overclaim a partial math result, see Claude's Riemann zeta bound — a real advance that was not the Riemann hypothesis.
What people are asking
Is this Muse Code, or Meta AI chat?
This batch is meta.ai chat + Muse Spark Thinking Mode, not a claim about the Muse Code terminal agent. Muse Code is the coding harness story. These papers are the research-collaboration story. Same model family lineage (1.1/1.2), different product surface.
Did Alexandr Wang announce new theorems?
Wang quote-posted Meta's announcement. He is not listed as an author on the six papers in Meta's research post. Authorship stays with the mathematicians Meta named. Do not cite Wang as a theorem author.
How does this compare to Fable 5 / Jacobian-style results?
Earlier explainx.ai coverage of AI-assisted counterexamples and verification write-ups already taught the same hygiene: named humans, verification paths, careful claim scope. Meta's batch is another instance of that pattern at Meta's scale, with an unusually explicit "no custom scaffold" product claim.
Should I use Muse Spark for my own research notes?
If you already have Muse Spark access, Meta is arguing that ordinary Thinking Mode chat can participate in hard work when a specialist drives. That is useful as a workflow hypothesis. It is not a substitute for a second reader, and it is not a license to skip literature search — Meta's own papers credit concurrent solutions found by others.
Does "five open problems" mean five Millennium Problems?
No. These are specialized open questions in probability, PDE, group theory, optimization, and algebra, plus one arithmetic-physics extension. They are not Clay Millennium Prize problems. For that scorecard hygiene, use the Millennium fact-check.
What a builder should do this week
- Read Meta's hub post, then open the paper pages you care about — start with the ellipsoid threshold or the group-theory counterexample if you want concrete, checkable objects.
- Score the release against AGMAI Path A vs Path B. Named ownership and marked AI text look like Path A. Missing public prompt/cost/Lean logs are gaps if you expected Path B completeness.
- Keep Muse Spark evaluation separate from Muse Code evaluation. Research chat wins do not auto-transfer to Terminal-Bench scores.
- Copy the credit pattern in your own AI-assisted notes: mark AI-drafted sections, name the human owner, cite concurrent work, and have a second reader before you ship.
- Do not invent exclusivity. If your tool also found a counterexample later, say so the way Meta acknowledges Nilradical and Hu–Wen.
Related reading
- AGMAI responsible release of AI-generated mathematics
- Muse Code + Muse Spark 1.2 launch
- Muse Spark 1.3 benchmarks and pricing
- Claude-shaped science and BootLoops
- Opus 5.5 and the 1615 dodo find — expert attention still decides
- OpenAI advisory group and the 100+ problems claim
- Fields Medalists on AI math misalignment
- Lean 4 formalization cost collapse
Primary sources: Solving Open Research Problems Together · six Meta Research publication pages linked above · AI at Meta on X (@AIatMeta)
Details reflect Meta's October 2, 2026 research post and the six ai.meta.com publication URLs verified HTTP 200 on October 3, 2026. Author lists, review names, and concurrent acknowledgments are taken from that post. arXiv IDs, journal DOIs, prompt logs, compute costs, Lean formalizations, and Alexandr Wang's quote-post caption text were not verified as independent artifacts for this article — do not invent them. Proof correctness remains for specialists reading the papers, not for a news summary.