News · 2026-08-18

AlphaEvolve helped tighten the matrix multiplication exponent, and the proof was checked in exact arithmetic

The upper bound on the matrix multiplication exponent has been tightened to 2.371177, down from the previous best of 2.371339. The improvement, published as a short note on arXiv, came from reformulating the optimization problem at the heart of the current best method, designing a new machine-learning-derived algorithm to solve it, and then refining that algorithm with DeepMind's AlphaEvolve. The final numerical solution was rounded to rationals and re-evaluated in exact rational arithmetic, eliminating any possibility of floating-point error in the certificate.

Key facts

• New certified upper bound on omega: 2.371177, improving the previous best of 2.371339.
• Published August 2026 as arXiv:2608.16884.
• The method builds on combination loss analysis, a refinement of the laser method credited in the paper to Duan et al. 2022, Williams et al. 2024, and Alman et al. 2025.
• The AI component is AlphaEvolve, DeepMind's Gemini-powered coding agent for designing algorithms.

The improvement is in the sixth decimal place. That sounds like a rounding error and is not, and understanding why requires knowing what omega is.

Multiplying two large square matrices the obvious way, row by column, costs roughly n cubed operations for n-by-n matrices. In 1969 Volker Strassen showed that you could do it with fewer, using a clever recursive decomposition, and started a chase that has run for fifty-seven years. Omega is the finish line: the smallest exponent any matrix multiplication algorithm could possibly achieve. It is known to be at least 2, because you have to at least look at every entry. Nobody knows the true value. What the field has instead is a slowly descending ceiling, and every step down that ceiling has required a genuinely new mathematical construction.

That is why a sixth-decimal-place move is a result rather than a rounding. Nobody nudges omega. You either build a new certified construction or you do not.

The paper's own summary of its route is precise. "First, we reformulate the optimization problem allowing us to solve it in a larger setting than was previously possible. Second, we leverage recent advances in machine learning to design a new optimization algorithm for this problem. Finally, we refine the resulting optimization algorithm with AlphaEvolve. Our combined approach yields an upper bound of omega less than 2.371177, improving the previous best bound of 2.371339."

The analogy that fits is a very long chain of inequalities, each one a lever. The laser method and its combination-loss refinement give you a machine with hundreds of knobs, where any setting of the knobs yields some valid bound and the best settings yield the best bound. Finding those settings is a brutal high-dimensional optimization problem, not a proof problem. What the authors did was rebuild the machine so more knobs could be turned at once, then hand the knob-turning to better optimizers, with AlphaEvolve improving the optimizer itself rather than the mathematics.

Why it matters is not speed. It is what "AI made a discovery" ought to mean. There has been a run of claims this year about models advancing mathematics, and the field's response has been justifiably sceptical; we have written about an AI tightening a seventy-year-old constant where the paper itself called the judgment weak and about a Claude bound on the Riemann zeta critical line that we assessed as not real. This result is different in exactly one respect that matters: the answer is machine-checkable. The authors did not ask anyone to trust a model's reasoning. They produced a rational-arithmetic certificate that any competent reader can verify independently, which is the standard we have argued for on every previous claim of this kind.

Two honest caveats. The first is practical irrelevance. These algorithms are galactic, meaning the crossover point where they beat straightforward methods sits at matrix sizes larger than any real computation. Nobody's training run gets faster because of this. The second is scope of credit. AlphaEvolve refined an optimization program the humans had already reformulated and scaled; it did not invent the approach, and the paper does not claim it did. This is a collaboration with a clearly delineated division of labour, and it should be reported as one.

One clarification, since it is easy to conflate: this is not DeepMind's earlier AlphaEvolve announcement about a faster 4x4 complex matrix multiplication scheme. That was a concrete small-case improvement. This is the asymptotic exponent, which is the stronger claim.

For the underlying question of why matrix multiplication dominates AI compute in the first place, our lesson on why LLM inference is memory bound covers where the real bottleneck sits in practice, which is somewhere quite different from where omega lives.

Primary source, verified: read the paper → (arXiv 2608.16884)

Ground Truth. (2026, August 18). AlphaEvolve helped tighten the matrix multiplication exponent, and the proof was checked in exact arithmetic. Ground Truth. https://groundtruth.day/news/alphaevolve-tightened-the-matrix-multiplication-exponent.html

@misc{groundtruth:alphaevolve-tightened-the-matrix-multiplication-exponent,
  title  = {AlphaEvolve helped tighten the matrix multiplication exponent, and the proof was checked in exact arithmetic},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/alphaevolve-tightened-the-matrix-multiplication-exponent.html}
}

Topics: mathematics · alphaevolve · deepmind · algorithms · ai-for-science · theory · proof