Agentic Bayesian Optimization Enhances Autoresearch with LLMs.

Bayesian optimization (BO) is a standard for sample-efficient optimization, but its efficiency can be further enhanced by incorporating richer domain priors, which are typically hard to encode. This research proposes "agentic Bayesian optimization," a novel paradigm that integrates large language models (LLMs) to make informal priors from natural language and code directly available to the optimizer. In this setup, an LLM agent serves as the central decision-maker within the BO loop, while a Bayesian backend provides the uncertainty-aware optimization substrate. The agent, instantiated as "Sara" (surrogate-augmented autoresearch agent), configures the problem, queries the backend, selects evaluations, and can dynamically revise the optimization strategy based on new instructions or observed evidence. This includes tightening bounds, switching acquisition functions, or even reframing the problem. Benchmarking on synthetic and real-world tasks shows Sara preserves the reliability of state-of-the-art BO without prior knowledge, outperforms LLM-based baselines, and uses natural-language priors to surpass standard BO. It also demonstrates practical value in dynamic settings by reconfiguring the optimization problem on the fly.