Can machine feedback sustain discovery at test time?

Most demonstrations of LLMs "doing science" stop at hypothesis generation or benchmark gains. AlphaEvolve goes further: an evolutionary coding agent that orchestrates an autonomous pipeline of LLMs making direct changes to code, continuously scored by one or more automated evaluators. The loop produced real, deployed results — a more efficient data-center scheduling algorithm at Google, a functionally-equivalent simplification in hardware-accelerator circuit design, a faster matrix-multiplication algorithm, and an acceleration of the training of the very LLM underpinning AlphaEvolve.

The conceptual keeper: AlphaEvolve is best read as a test-time compute agent where machine feedback sustains compute scaling into the regime of genuine discovery — far beyond repeated sampling. Because the evaluator is automatic and objective, the loop can run long enough to reach novel solutions, and the same problem can be attacked in different ways (search the solution directly, evolve a constructive function, or evolve a search algorithm), each with different inductive biases.

This anchors the autonomous-research cluster on the verification side. Since What limits how much models can improve themselves?, AlphaEvolve works precisely where that gap is wide and cheaply checkable — automated evaluators are the verification advantage made concrete. It complements Can AI research itself without losing human oversight? and extends Can AI systems improve themselves through trial and error? from self-modifying agents to deployed algorithmic artifacts.