STaSBO: LLM-Guided Strategy Planning and Configuration Selection for Bayesian Optimization
Summary
- Researchers introduced STaSBO, an LLM-driven meta-controller for strategic planning and configuration selection in Bayesian optimization.
- STaSBO uses a structured memory system to connect prior plans, configuration selections, and candidate behaviors to optimization states.
- The framework combines strategy planning with surrogate fit evidence and prospective BO-action diagnostics for joint configuration selection.
- Across 12 synthetic mathematical and 10 scientific application tasks, STaSBO achieved the best average rank compared to 11 baselines.
Why it matters
Integrating large language models into the strategic management of Bayesian optimization can improve sample efficiency and configuration tuning across complex scientific and engineering tasks.