Prompt optimization through textual feedback has shown promising results in improving the performance of large language models (LLMs) on downstream tasks. However, existing approaches often rely on selecting prompt edits from a pool of candidate gradients using random sampling or local heuristics, requiring multiple evaluations to find effective modifications. In this work, we propose a center-aware selection method that identifies high-quality gradient candidates based on their proximity to a robust semantic center representation of the gradient pool. Rather than sampling or scoring candidates iteratively, our method embeds all textual gradients and deterministically selects the top-k closest to the semantic center, which captures the consensus of the candidate pool. Experiments on three diverse datasets demonstrate that our approach not only improves predictive performance but also reduces the number of required model queries. In addition, qualitative analyses reveal that gradients near the center tend to encode more generalizable reasoning patterns. These findings highlight the utility of semantic embedding space as a reliable signal for selecting effective prompt edits in a resource-efficient manner.
Jang et al. (Fri,) studied this question.