EPFL's GOLLuM teaches LLMs to doubt, cutting lab trials by 40%

MacrinePhD · x · 2026-09-03

Researchers Bojana Ranković and Philippe Schwaller at EPFL built GOLLuM, a framework pairing LLMs with Gaussian processes — a probabilistic "doubt detector" from Bayesian optimization — to turn uncertainty into a training signal for finding optimal experimental recipes.

Classic Bayesian optimization transfers poorly across scientific domains, restarting from scratch each time;
LLMs encode broad scientific knowledge but hallucinate with unwarranted confidence, risking costly detours;
GOLLuM attaches uncertainty estimates to LLM suggestions so the system knows when it might be wrong.

On chemistry benchmarks, GOLLuM matches traditional methods while requiring about 40% fewer lab trials.

Original post →

EPFL's GOLLuM teaches LLMs to doubt, cutting lab trials by 40%

More from Apps