The problem
How it works
AlphaEvolve is an evolutionary coding agent. Gemini Flash proposes many program variants and Gemini Pro adds deeper suggestions; automated evaluators run and score each candidate, and the strongest ones seed the next round. Google DeepMind applied it to its Borg scheduler, Gemini training kernels, a TPU arithmetic circuit and more than 50 open mathematics problems.
Results and limitations
As vendor claims, Google DeepMind reports that a scheduling heuristic recovers about 0.7% of Google's worldwide compute on average, that a 23% kernel speed-up cut Gemini training time by 1%, and that a FlashAttention kernel ran up to 32.5% faster. In mathematics it reports multiplying 4x4 complex matrices with 48 scalar multiplications, matching the best known result on about 75% of the open problems tried and improving it on about 20%.
Before you try it
The method only suits problems whose solutions are code that a program can score automatically; without a trustworthy evaluator there is nothing to optimise. AlphaEvolve was not generally available when announced (Google described an early access programme for selected academics), so the transferable lesson is the pattern: write a strict evaluator first, then let an agent search. Results depend on your setup and were not independently reproduced by Top 200 Repos.
Tools and repositories
No ranked repository is involved. This use case still stands on its own source.
- Tools
- AlphaEvolve (early access) · Gemini Flash and Gemini Pro · An automated evaluator
- Prerequisites
- AlphaEvolve itself was offered through an early access programme; the prompt reuses its evaluator-first pattern with a general coding agent.
Try it yourself: Run an evaluator-first optimisation search
Fill the fields to adapt the prompt. Empty fields stay as visible [placeholders].
I want to improve [Function to improve] in [File containing it], measured by [Metric to optimise]. Before changing anything, write an evaluator script that checks correctness against [Correctness reference] and reports the metric, and show it to me for approval. Then propose up to [Number of candidates] alternative implementations, run the evaluator on each, and keep a table of candidate, correctness result and metric. Reject any candidate that fails a correctness check, however fast it is. Do not change the evaluator or tests after I approve them, and do not edit files outside [File containing it].This applies the AlphaEvolve pattern with any coding agent; the evaluator is the part to review most carefully.
Sources
- AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms (opens in a new tab)
Vendor post · deepmind.google · Published 14 May 2025 · Source checked 11 Nov 2026, 22:05 UTC
What this source supports
- Uses an ensemble of Gemini Flash and Gemini Pro with automated evaluators in an evolutionary loop
- A Borg scheduling heuristic recovers on average 0.7% of Google's worldwide compute
- A kernel speed-up of 23% cut Gemini training time by 1%; a FlashAttention kernel ran up to 32.5% faster
- Matched the state of the art on about 75% of 50+ open maths problems and improved about 20%
- Requires problems with an automated evaluation metric; early access planned for selected academics
- AlphaEvolve: A coding agent for scientific and algorithmic discovery (opens in a new tab)
Paper · arxiv.org · Published 16 Jun 2025 · Source checked 11 Nov 2026, 22:05 UTC
What this source supports
- Describes an evolutionary coding agent that edits code and refines it with evaluator feedback
- Reports multiplying two 4x4 complex-valued matrices with 48 scalar multiplications
- Reports results in data-centre scheduling, hardware circuit simplification and LLM training speed