DASH: Decoupled Adaptive Surrogate-Acquisition Harness for Automated Bayesian Optimization

A recent paper on arXiv (2608.00641v1) presents DASH, which stands for Decoupled Adaptive Surrogate-Acquisition Harness, designed for enhancing Automated Bayesian Optimization (AutoBO) with large-language models (LLMs). Bayesian optimization (BO) utilizes both a surrogate model and an acquisition function, but the best configurations differ depending on the task and stage of optimization. Current AutoBO approaches either modify a single component, resulting in mismatches, or select surrogate-acquisition pairs based on a unified criterion, neglecting their unique functions: surrogate selection is influenced by predictive reliability, while acquisition adaptation must consider the context of the campaign. DASH resolves this by choosing surrogates based on predictive reliability, uncertainty calibration, and ranking consistency. Additionally, its two-stage acquisition controller adjusts quotas among acquisition functions periodically. The paper is accessible on arXiv and is classified as a preprint.