LABO: LLM-Accelerated Bayesian Optimization Framework
Researchers propose LABO (LLM-Accelerated Bayesian Optimization), a framework integrating large language models into Bayesian optimization to reduce experimental costs. LABO uses a gating criterion to dynamically balance LLM predictions with real experiments, leveraging inexpensive LLM evaluations for broad exploration and reserving costly experiments for high-uncertainty regions. The approach aims for more sample-efficient optimization, supported by theoretical analysis of cumulative regret. The paper is available on arXiv under identifier 2605.22054.
Key facts
- LABO stands for LLM-Accelerated Bayesian Optimization
- It combines LLM predictions with experimental observations in a single BO loop
- A gating criterion dynamically balances reliance on LLM predictions versus actual experiments
- Inexpensive LLM evaluations are used for broad exploration
- Costly real experiments are reserved for high-uncertainty regions
- The framework aims for more sample-efficient optimization
- Theoretical analysis includes cumulative regret
- Paper published on arXiv with identifier 2605.22054
Entities
Institutions
- arXiv