Embedding by Elicitation: Dynamic Representations for Bayesian Optimization of System Prompts

In three linesReElicit is a Bayesian optimization framework for tuning system prompts using only aggregate feedback. An LLM dynamically elicits a compact, interpretable feature space, then a Gaussian process selects optimized target vectors refined into deployable prompts. Across 10 tasks with 30-evaluation budget, ReElicit outperforms aggregate-only prompt optimization baselines.

Summary generated by Claude — human-verified