Keywords: DeepMind · AlphaEvolve · matrix multiplication exponent · combination loss analysis · exact rational certificate · Carnegie Mellon · arXiv 2608.16884
---
How to read this
In 1969, Volker Strassen reduced the arithmetic complexity of multiplying two n×n matrices from n³ down to O(n^log₂7) ≈ O(n^2.807), opening a 57-year chase for the true value of the matrix multiplication exponent ω. The lower bound is 2 (you must at least read every element), and Strassen's 2.807 has never been eliminated—only pushed downward. On August 17, 2026, a DeepMind team—Emilien Dupont, Marvin Eisenberger, Borislav Kozlovskii, Abbas Mehrabian, Francisco J. R. Ruiz, Abigail See, Renfei Zhou (CMU), Josh Alman (Columbia), Virginia Vassilevska Williams (MIT), and Matej Balog—pushed the upper bound from 2.371339 (the SODA 2025 result of Alman et al.) down to ω < 2.371177. Subtracting the two records gives just 0.000162, which sounds like a rounding error, but on the matrix multiplication exponent leaderboard every drop in the sixth decimal place corresponds to a genuinely new mathematical construction. This is not an "AI independently proving things" story, but a collaboration sample of "humans rewriting mathematics, AI rewriting the optimizer's code"—worth unpacking in detail.
1. What the optimization problem looks like
Where does an upper bound on ω come from? Every advance over the past half century has come from the same mother lode—the laser method—whose most recent generation is called combination loss analysis (attributed to Duan et al. 2022, Williams et al. 2024, Alman et al. 2025). This method translates "can ω be compressed to some value" into "can one find a feasible point in a non-convex optimization problem." It sounds like ordinary industrial black-box optimization, but the essential difficulty is: how low a bound you can prove depends entirely on how large you can make the search space; and as the search space grows, the combinatorial variables explode exponentially, so predecessors could only work at the scale of ℓ*=3.
The shape of the optimization problem can be imagined as "a non-convex black box with a few hundred knobs, where each knob setting corresponds to a certificate, and the certificate's value is the ω upper bound derived from that setting." Finding better knob combinations is one of the most vexing shapes in non-convex optimization. The previous record (SODA 2025) of ω < 2.371339 was only sustained by Alman pushing this method's asymmetry to its extreme.
2. Three layers, from 2.371339 to 2.371177
This short paper, submitted to arXiv on 2026-08-17 (2608.16884), splits the work into three parts. Each layer alone is not "sexy," but stacked together they advanced the bound by 0.000162:
**Layer one: algebraic restructuring, making a larger ℓ* solvable.** The authors gave an algebra-level reformulation of the original optimization problem so that a larger parameter setting becomes solvable. The paper calls this parameter ℓ*; the previous record had ℓ* = 3, and this work pushes it to **ℓ* = 4. This is mathematical work—AI had no hand in it.
Layer two: rewriting the solver with JAX + automatic differentiation + Sinkhorn-Knopp normalization. Prior work used a sequential quadratic programming (SQP) solver; DeepMind switched to a gradient-based solver written in JAX, obtaining gradients via automatic differentiation and using Sinkhorn-Knopp for normalization (repeatedly projecting intermediate matrices back onto the set of doubly stochastic matrices). This layer is "numerical methods" work—AI still had no hand in it—but with the tools swapped, the shape of the feasible region changed.
Layer three: handing that solver to AlphaEvolve. AlphaEvolve is DeepMind's own Gemini-driven coding agent, designed specifically to "rewrite the optimizer's code itself." In other words, it keeps rewriting the solver code into new versions, runs them, checks what new upper bound results, and goes back to rewrite the code. The paper states explicitly: "AlphaEvolve proves no propositions about ω; it rewrites the optimizer code that searches for upper bounds on ω."
3. Why this matters
This track has a peculiar iteration rhythm. Each new upper bound means either a new combinatorial structure has appeared, or an existing one has been mined more deeply. In other words, every advance is a genuinely new construction, not a small tweak of parameters. This 0.000162 sounds small, but it counts as "completing another step of the descending staircase of the past 57 years."
More crucially, this time there is an exact rational-arithmetic certificate. DeepMind rewrote the final numerical solution using rational numbers and re-verified it under exact arithmetic, in order to "eliminate the possibility of floating-point error"—any ω upper bound of 2.371177 computed with IEEE floating point could be suspected of being 2.3711769999... after rounding; written in rationals, it can be independently verified with no rounding. This is exactly the posture that news like "AI moved one digit on a 70-year constant" should adopt: the certificate should let third parties re-run it.
Up to this point, DeepMind did what mathematicians do—it's just that the boundary of "AI contributing decisive ideas as a collaborator" has, for the first time, simultaneously landed on "machine-verifiable certificates." This differs from the Sendov conjecture (07-19) and the Crouzeix conjecture (07-20) episodes: in those two cases it was "humans wrote the proofs, AI helped find ideas"; this time it is "mathematicians define the problem, AI rewrites the solver, and the certificate is machine-verifiable." Three modes are now appearing side by side in the late summer of 2026.
4. Three honest limitations
First, the result has not yet been independently reproduced. The ω < 2.371177 output by AlphaEvolve is DeepMind's own computation. We must wait for someone on arXiv to repeat it, or for the authors to release the verification repository. The paper mentions that verification code and the found feasible solution are in a "code repository in preparation"—not yet published. Until that step is released, independent cross-verification has to wait.
Second, this upper bound will almost certainly not make real matrix multiplication faster. ω is an infimum, and the "crossover point" with existing algorithms—the matrix order beyond which these methods beat the schoolbook algorithm—is far larger than any order real computation would use. These are galactic algorithms; however pretty the asymptotic bound, BLAS libraries will not update. In other words, there is no effect on the speed of training a matrix multiplication protocol or a diffusion model.
Third, AlphaEvolve's credit is bounded. What it rewrote was not the proof path for "ω" but "the optimizer's code." The problem layer, the parameter space, and the feasible combinatorial methods were all worked out by humans; the machine searched an already-defined search space. On this point, headlines like "AI advances a 57-year-old math problem" are somewhat misleading; the accurate description is "DeepMind + mathematicians collaborating to scale the search to a new magnitude." But this is still a crucial advance, because enlarging the search scale is a necessary condition for every step of progress in the past 57 years.
5. What it means
This track has one notable feature: the intervals between records keep shrinking. From Pan's 2.781 in 1978, to Coppersmith-Winograd's 2.3755 in 1990 (the section later nicknamed "the CW convex-optimization bound"), to Stothers' 2.387 in 2010, to Alman-Williams' 2.372869 in 2020, to Alman's 2.371339 in 2024, to AlphaEvolve's 2.371177 in August 2026. Every advance is hard, but after AI's involvement, search scale has become a new tool: previously human optimization algorithms were bottlenecked by their own compute, and now AI can run thousands of attempts in parallel, iteratively modifying code. This mode—"AI doesn't prove mathematics, but accelerates the tools of mathematics"—will recur on more records.
The arXiv abstract page lists no institutions, and the team consists mainly of nine DeepMind researchers plus one person each from Carnegie Mellon, Columbia, and MIT. First author Emilien Dupont is a DeepMind research scientist long working on neural network interpretation and AlphaEvolve-style automated code discovery; corresponding author Matej Balog is also at DeepMind; Josh Alman and Virginia Vassilevska Williams are the very people who pushed the previous record of ω < 2.371339 to near this result, and they are co-authors here. That is why this seemingly "sudden small advance" is in fact the continuation of the same math-circle + AI-circle collaboration chain.**
---
References
- arXiv 2608.16884: Improving the matrix multiplication exponent with modern optimization and AlphaEvolve (DeepMind / CMU / Columbia / MIT, submitted 2026-08-17 17:59 UTC): https://arxiv.org/abs/2608.16884
- AI Brief 18 August 2026: a sharper matrix multiplication exponent, with AlphaEvolve in the loop (independent technical commentary, 2026-08-18): https://muhammad-ahmed.com/blog/ai-brief-matrix-multiplication-exponent
- DEV.to: AlphaEvolve helped tighten the matrix multiplication exponent, and the proof was checked in exact arithmetic (independent technical analysis, 2026-08): https://dev.to/breachprotocol/alphaevolve-helped-tighten-the-matrix-multiplication-exponent-and-the-proof-was-checked-in-exact-2fhi
- Alman, Duan, Williams, Xu et al.: More asymmetry yields faster matrix multiplication (SODA 2025, source of the previous record bound 2.371339)
- Duan, Wu, Zhou et al.: Fast Matrix Multiplication via Group Leaders (origin of combination loss analysis, 2022)
- Williams, Xu, Xu, Zhou: New Bounds for Matrix Multiplication with applications to GRH and beyond (2024)