An AI capable of doubt can optimize scientific discovery
EPFL researchers developed GOLLuM, a framework that trains large language models to act as uncertainty‑calibrated optimizers for experimental design. By coupling an LLM with a Gaussian process doubt detector, GOLLuM learns from past experiments to reorganize the search space, achieving faster convergence and higher success rates across 23 benchmark tasks in chemistry and materials science. The method reduces the number of required experiments by over 40% compared to traditional Bayesian optimization.
EPFL researchers developed GOLLuM, a framework that trains large language models to act as uncertainty‑calibrated optimizers for experimental design. By coupling an LLM with a Gaussian process doubt detector, GOLLuM learns from past experiments to reorganize the search space, achieving faster convergence and higher success rates across 23 benchmark tasks in chemistry and materials science. The method reduces the number of required experiments by over 40% compared to traditional Bayesian optimization.
This demonstrates Swiss AI’s capacity to integrate language models with probabilistic reasoning for efficient scientific discovery, aligning with SNSF’s NCCR Catalysis objectives.
Original source record. Open the original record to verify the underlying announcement.