Large language models are increasingly used to guide Bayesian optimization, drawing on past observations to propose new search directions. However, the paper notes that these proposed directions and any subsequent revisions can stay implicit, making it hard to see how the search is actually evolving.

To address this, the authors propose Feedback-Guided Search State (FGSS), a structured way to represent the search state. While the abstract is truncated, the core idea is clear: by making the search state explicit, FGSS should allow the LLM to manage hypotheses more systematically and respond to feedback in a more principled manner.

The work is described in an arXiv preprint, and the full details of the method and its evaluation are not yet available from the abstract alone. Still, the motivation is a practical one—bringing more structure to a process that otherwise risks being opaque.