LLINBO: Trustworthy LLM In-the-Loop Bayesian Optimization

Bayesian optimization is a sequential decision-making tool for optimizing expensive black-box functions. Large language models have recently shown strong performance in low-data settings, making them promising tools for proposing informative query points using contextual knowledge.

However, relying solely on language models as optimization agents creates risks because they do not provide explicit surrogate modeling or calibrated uncertainty. This work proposes LLINBO, a hybrid framework that combines language models with statistical surrogate experts such as Gaussian processes. The framework uses language models for contextual early exploration while relying on principled statistical models for efficient exploitation, and concludes with a proof-of-concept application in 3D printing.

Recommended citation: C.-Y. Chang, M. Azvar, C. Okwudire and R. A. Kontar. "LLINBO: Trustworthy LLM In-the-Loop Bayesian Optimization." arXiv preprint, 2026.
Download Paper