Beyond Prompting: How Algorithmic Evolution Doubled our Training Speed

Rex Lin and Valeria Verzi at Klarna, with Anant Nawalgaria at Google, on using AlphaEvolve to optimise the training pipeline behind their models rather than prompting an assistant to do it. The setup is a search loop: engineers fence off which code may change, name the metric, and fix the constraints that must hold, then the system writes candidate programs, runs them, scores them, and breeds from the survivors. Nearly 6,000 candidates over three weeks, with no human reviewing individual outputs.

Throughput went from 49 samples per second to roughly 72 in the mixed-precision phase, and to about 97 once run under deterministic constraints, hence the doubling in the title. Determinism was not optional here: if you can’t reproduce the exact result, you can’t audit it, and you can’t deploy it in regulated financial services, so the constraint was built into the fitness function rather than checked afterwards.

Two details worth keeping. Model quality decides whether this is viable at all: an earlier model produced syntactically broken programs about 67% of the time, while a newer one landed in the 86 to 97% success range, which is the difference between a search that mostly wastes compute and one that mostly explores. And progress is lumpy, with the largest experiment sitting through 631 consecutive evaluations that found nothing before improving again, which is a hard thing to fund if you are watching a dashboard.

The framing the authors want is that this is a category apart from prompt-driven coding, useful where the search space is too large to reason through and success is cheaply measurable. Note that all figures here are Klarna’s own, about Klarna’s internal pipeline.