Swapping the Gaussian Process for an LLM in Bayesian optimization, benchmarked 4 ways

tak3sh8 · x · 2026-09-21

Developer Naman Gupta blogged an experiment replacing the Gaussian Process surrogate in Bayesian optimization with an LLM agent: instead of fitting a math function, the LLM reads all prior experiment records to make reasoning-driven guesses. Using a hot-sauce recipe as the motivating example, he benchmarks four setups — classical Bayesian optimization, LLM-only, a hybrid of both, and a human-in-the-loop agent with a reliability layer before acting on decisions.

Original post →

More from coding & agent

coding & agent channel →