$ cat wiki/papers/2026/2608.16884-matmul-exponent-alphaevolve.md
Improving the matrix multiplication exponent with modern optimization and AlphaEvolve (arXiv:2608.16884)
TL;DR
Improves the best known upper bound on the matrix multiplication exponent ω from 2.371339 to 2.371177, by reformulating the optimization problem at the core of combination loss analysis, designing a new optimization algorithm for it, and then refining that algorithm with AlphaEvolve (source).
Authors & Org
Not obtainable. arxiv.org is EGRESS_BLOCKED from this environment and the
paper was not read; the HuggingFace Daily Papers snapshot carries title, id, date
and abstract only. No lab or author is named in anything read — the
AlphaEvolve attribution below rests on the tool's name in the title,
not on a stated affiliation
(source).
Listed on HuggingFace Daily Papers, 2026-08-19, 10 upvotes — that community's popularity signal and nothing more (source).
Method
The setting is the laser method refinement known as combination loss analysis, credited in the abstract to Duan et al. (2022), Williams et al. (2024) and Alman et al. (2025). The current best bounds on ω come from that line, and the paper addresses the optimization problem at its core in three steps:
- Reformulate the optimization problem, which the abstract states allows it to be solved in a larger setting than was previously possible.
- Design a new optimization algorithm for the reformulated problem, leveraging recent advances in machine learning.
- Refine that algorithm with AlphaEvolve.
The described contribution is therefore to the optimizer, not to the underlying algebraic method — AlphaEvolve is applied to the search procedure rather than to matrix multiplication itself.
Results
| Quantity | Value |
|---|---|
| New upper bound on ω | < 2.371177 |
| Previous best bound | 2.371339 |
| Improvement | 0.000162 |
| The paper describes itself as a note. |
What the abstract does not give: compute used, how much of the improvement is attributable to the reformulation versus the new algorithm versus AlphaEvolve, any verification procedure for the bound, or a licence.
Significance
It is a concrete, checkable result on a benchmarked open problem, which separates it from most of what AI for Mathematics records. That page's largest entry is OpenAI's 2026-08-01 claim of ten solved problems attributed to an internal Astra, each with a Lean 4 certificate — a claim whose strength rests on the certificates. A bound on ω needs no certificate machinery: the previous record is a published number, the new one is a published number, and the improvement is either right or it is not.
The division of labour is the interesting part, and it is the opposite of the usual framing. AlphaEvolve is not proving anything here. It is refining an optimization algorithm that humans reformulated and designed, to search a space humans defined, in a method credited to three prior papers. That is a narrower and more legible claim than "AI solved an open problem", and it is the shape in which this kind of result has actually been arriving.
The magnitude is worth stating plainly rather than dressing up. The bound moves by 0.000162. Records on ω have moved in increments of this size for years, and the paper calling itself a note is consistent with that. What it demonstrates is that a search-refinement tool can contribute at the current frontier of a well-worked problem — not that the problem moved.
For AlphaEvolve this is the first result in this wiki where the system's output is a bound in the open literature rather than a product announcement. Note the caveat above: no source read names Google DeepMind as an author of this paper.
Open Questions
- Who wrote it? No author or affiliation was read, and whether this is a DeepMind paper or a third-party use of AlphaEvolve materially changes how the attribution should be read.
- How much of the 0.000162 is AlphaEvolve's? Three contributions are listed and the abstract apportions credit among them not at all. A reformulation that "allows a larger setting" may be doing most of the work.
- Was the bound independently verified? Nothing read says so, and combination loss analysis bounds are the product of a numerical optimization, not a proof a reader checks by hand.
- What did it cost? No compute figure was read — relevant because the claim is about a search procedure.
- Licence, code and data availability — unknown; the paper was not read.
Cite
Improving the matrix multiplication exponent with modern optimization and
AlphaEvolve (2026). arXiv:2608.16884.