arXiv Paper Proposes Out-of-Loop Multi-Fidelity Bayesian Optimization

A new paper on arXiv (ID: 2608.04113) introduces a method for multi-fidelity Bayesian optimization (MF-BO) when the highest-fidelity function is too expensive to query during optimization. The authors demonstrate that standard MF-BO algorithms are suboptimal in real-world scenarios where gold standard data (observations of the highest-fidelity function) from previous experiments are available. They propose an 'out-of-loop' approach that leverages this existing high-fidelity data to improve optimization performance. The paper is categorized as a cross-type announcement and is relevant to fields such as molecular optimization, where chemists often use simulations to select top-k candidates and later reveal true objective values. The work addresses a common challenge in black-box optimization, where objective functions are expensive and lower-fidelity proxies are used. The proposed method aims to make MF-BO more practical by incorporating historical high-fidelity observations without requiring them to be part of the active optimization loop. The paper is available at the provided arXiv URL.