Autotune mixes CMA-ES and LLM proposals to warm-start hyperparameter search

RobertTLange · x · 2026-07-26

The follow-up explains how Autotune works in more detail: the agent reads the script and proposes tunable parameter ranges, refinement rounds narrow the search space, a “Centaur” sampler mixes CMA-ES with LLM proposals, and trial transfer warm-starts from prior best runs.

The post also links the code and a blog write-up, making it a more technical companion to the original launch.

Related event: Autotune Combines LLMs and CMA-ES for Hyperparameter Optimization(2 posts)→

Original post →

More from coding & agent

coding & agent channel →