AlphaEvolve Explained: Google’s AI Math Discoveries

Quick answer: AlphaEvolve is Google DeepMind’s Gemini-powered coding agent for discovering and improving algorithms. It generates program variants, scores them with an automatic evaluator, and uses the best ones to guide further search. Its best-known math result is a way to multiply 4×4 complex matrices with 48 scalar multiplications instead of the 49 from applying Strassen’s method twice. In 2026 it helped improve the matrix multiplication exponent and nine Ramsey number lower bounds, and became generally available on Google Cloud.

Updated November 22, 2026.

Most AI math headlines are about proofs. AlphaEvolve is different: it is a machine for finding things. Give it a problem where any candidate answer can be scored, such as “fit as many spheres as possible” or “multiply matrices with as few multiplications as possible,” and it evolves programs that produce better answers. Since Google DeepMind announced it on May 14, 2025, it has improved several long-standing records and become a tool working mathematicians use.

How AlphaEvolve Works

  1. Define the task. A person supplies a starting program, the problem it should solve, and an evaluator that scores any output automatically.
  2. Generate alternatives. Gemini models propose changes to the program, from small tweaks to new strategies.
  3. Evaluate the candidates. Each new program is run, checked against the rules and scored.
  4. Use the feedback. The best programs are kept and mutated again, thousands of times, like natural selection for code.

The evaluator is the heart of the system. For a packing problem, it must first check that no shapes overlap before rewarding a denser arrangement. For a speed task, a faster program must still give correct answers. If a computer cannot score an answer, AlphaEvolve cannot improve it.

What Has AlphaEvolve Discovered?

DateResult (old → new)What it establishes
May 20254×4 complex matrix multiplication: 49 → 48 scalar multiplicationsA valid, faster algorithm for this case
May 2025Kissing number in 11 dimensions: lower bound 592 → 593At least 593 spheres fit (later surpassed; see below)
Sept 2025MAX-4-CUT hardness: 0.9883 → 0.987; metric TSP hardness: 117/116 → 111/110Stronger limits on how well these problems can be approximated
Mar–Apr 2026Nine classical Ramsey numbers, e.g. \(R(3,13)\): lower bound 60 → 61Better lower bounds, not exact values
Aug 2026Matrix multiplication exponent: \(\omega < 2.371339\) → \(\omega < 2.371177\)A better upper bound for very large matrices

In its first report, DeepMind tested AlphaEvolve on more than 50 open problems in analysis, geometry and combinatorics: it matched the best known answer in about 75% of cases and improved it in about 20%.

What Do 64, 49 and 48 Mean?

To multiply two matrices, combine rows of the first with columns of the second:

\[ \begin{pmatrix} 1 & 2 \\ 3 & 4 \end{pmatrix} \begin{pmatrix} 5 & 6 \\ 7 & 8 \end{pmatrix} = \begin{pmatrix} 19 & 22 \\ 43 & 50 \end{pmatrix}. \]

The top-left entry is \(1 \times 5 + 2 \times 7 = 19\). Each of the four entries uses two multiplications, so this takes 8 in total (see how to multiply matrices).

Method for a 4×4 productScalar multiplications
Ordinary row-by-column (16 entries × 4)64
Strassen’s 1969 method applied twice (\(7^2\))49
AlphaEvolve (complex numbers, 2025)48

Volker Strassen showed in 1969 that a 2×2 product needs only 7 multiplications instead of 8, and using his trick at two levels gives 49 for 4×4. AlphaEvolve’s 48 uses complex numbers. Later in 2025, Jean-Guillaume Dumas, Clément Pernet and Alexandre Sedoglavic found a 48-multiplication method with rational numbers, building on it. A smaller count does not automatically mean faster code on every computer, because additions and memory use also matter.

The August 2026 result answers a different question. The exponent \(\omega\) describes how the work of multiplying \(n \times n\) matrices grows as \(n\) gets huge. A team including Josh Alman, Virginia Vassilevska Williams and DeepMind researchers combined new mathematics, optimization and AlphaEvolve to lower the best known bound to \(\omega < 2.371177\).

Kissing Numbers and Ramsey Numbers

The kissing number problem asks how many equal spheres can touch a central sphere. In 11 dimensions AlphaEvolve found 593, improving 592. That record has since moved: the standard tables now list a lower bound of 604. The exact answer is unknown.

Ramsey numbers measure how large a network must be before a pattern is forced. In 2026, Ansh Nagda, Prabhakar Raghavan and Abhradeep Thakurta used AlphaEvolve to build larger pattern-free networks, raising lower bounds for nine classical Ramsey numbers.

Can AlphaEvolve Prove Theorems?

AlphaEvolve searches for algorithms and constructions. A rigorously checked construction can itself prove something: that at least 593 spheres fit, or that a conjecture is false. But a high score alone does not establish a general theorem, and showing that nothing better exists needs a proof.

In “Mathematical exploration and discovery at scale” (November 2025), Bogdan Georgiev, Javier Gómez-Serrano, Terence Tao and Adam Zsolt Wagner tested AlphaEvolve on 67 problems. It matched the best known solutions in most cases and improved several. For one finite-field Kakeya problem, AlphaEvolve combined with Deep Think and AlphaProof helped go from a construction to a proof formalized in Lean. The authors also warn that a flawed evaluator can be “gamed” by answers that only look good.

Can You Use AlphaEvolve?

Yes. On July 9, 2026, Google made AlphaEvolve generally available on Google Cloud, through an API where you supply the evaluator. DeepMind also publishes the mathematical results, with code to verify them, in a public GitHub repository.

Beyond Math

Google reported that a scheduling rule found by AlphaEvolve recovers on average 0.7% of its worldwide computing capacity, and that a 23% faster matrix kernel cut Gemini’s training time by 1%. In May 2026 DeepMind described further uses in chip design, power grids, genomics and quantum circuits.

How AlphaEvolve Compares

SystemWhat it searches for
FunSearch (2023)Single functions that score well
AlphaTensor (2022)Matrix multiplication algorithms only
AlphaProof (2024)Formal proofs in Lean
AlphaEvolve (2025)Whole programs for algorithms and constructions

For the wider picture, see AI math breakthroughs of 2026.

Common Misconceptions

“AlphaEvolve is just a chatbot.” No. It combines language models with automatic scoring and an evolutionary search over programs.

“593 is the kissing number in 11 dimensions.” No. It was a lower bound, and the bound has since been raised.

“Saving one multiplication is trivial.” Not for an operation computers repeat trillions of times, and not for a case where 49 had stood since 1969 for complex matrices.

Watch: AlphaEvolve Explained in Video

If you learn better by watching, these videos are worth your time.

AlphaEvolve: Using LLMs to solve Scientific and Engineering Challenges (AI Coffee Break with Letitia). A clear 2025 explanation of how the evolutionary loop works; see this page for later results.
DeepMind’s AlphaEvolve AI: History In The Making! (Two Minute Papers). A quick, enthusiastic tour of the original results.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. What does AlphaEvolve evolve during its search?
Found a bug? Let us know!
2. What must the evaluator do?
Found a bug? Let us know!
3. What does the number 48 count?
Found a bug? Let us know!
4. What did the 593-sphere arrangement establish?
Found a bug? Let us know!
5. Does a high evaluator score prove a general theorem?
Found a bug? Let us know!

Frequently Asked Questions

What is AlphaEvolve?

AlphaEvolve is Google DeepMind’s Gemini-powered coding agent, announced May 14, 2025. It proposes changes to programs, scores the results with an automatic evaluator, and keeps the best ones in an evolutionary loop to discover algorithms and mathematical constructions.

What did AlphaEvolve discover?

A way to multiply 4×4 complex matrices with 48 scalar multiplications instead of 49, a 593-sphere kissing arrangement in 11 dimensions, stronger hardness bounds for MAX-4-CUT and metric TSP, better lower bounds for nine Ramsey numbers, and in August 2026 a better matrix multiplication exponent bound.

Can I use AlphaEvolve?

Yes. It became generally available on Google Cloud on July 9, 2026, and DeepMind publishes its mathematical results with verification code on GitHub.

Does 48 multiplications mean it is always faster?

No. It reduces the multiplication count for that case, but real speed also depends on additions, memory movement and hardware.

Is 593 still the best 11-dimensional kissing number bound?

No. 593 was AlphaEvolve’s 2025 result; standard tables now list a lower bound of 604. The exact value is unknown.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "AlphaEvolve Explained: Google’s AI Math Discoveries - Effortless Math"?

No one replied yet.

Leave a Reply