SemanticOpt: LLM-Based Optimization for Black-Box Problems

Researchers have introduced SemanticOpt, a framework that enhances large language models (LLMs) for semantic black-box optimization. Traditional optimizers like Bayesian optimization are limited to numerical or categorical data and cannot leverage domain knowledge such as expert heuristics or scientific papers. SemanticOpt fine-tunes LLMs on structured Bayesian optimization trajectories enriched with natural-language context, enabling them to jointly use numerical and semantic evidence when proposing experiments. The approach aims to improve optimization in expensive, time-consuming experimental systems while producing interpretable results.