Embedding by Elicitation: ReElicit Framework for Bayesian Optimization of System Prompts

Forum topic · 小凯 · 2026-05-21

Summary

ReElicit is a Bayesian optimization framework for tuning LLM system prompts when feedback comes only as aggregate scalar scores rather than per-example labels or critiques. Given a task description, previously evaluated prompts, and their scores, an LLM elicits a compact, interpretable feature space and maps prompts into it. A Gaussian process surrogate with an acquisition function selects target feature vectors, which the LLM realizes and refines into deployable system prompts. The feature space is re-elicited as new evaluations arrive, allowing the representation to adapt to the observed prompt-score history. Evaluated on ten system prompt optimization tasks with a total budget of 30 evaluations, ReElicit achieved the strongest aggregate performance among representative aggregate-only prompt-optimization baselines, showing that LLMs can serve as adaptive semantic representation builders, not just prompt generators, for black-box optimization over natural-language artifacts.

Paper Overview

Field: cs.AI, cs.LG Authors: Zhiyuan Jerry Lin, Benjamin Letham, Samuel Dooley Published: 2026-05-21 arXiv: 2505.01258

Abstract

System prompts are a central control mechanism in modern AI systems, shaping behavior across conversations, tasks, and user populations. Yet they are difficult to tune when feedback is available only as aggregate metrics rather than per-example labels, failures, or critiques. The authors study this aggregate feedback setting as sample-constrained black-box optimization over discrete, variable-length text.

Key Contributions

ReElicit: a Bayesian optimization framework based on *embedding by elicitation*. Given a task description, previously evaluated prompts, and scalar scores, an LLM elicits a compact, interpretable feature space and maps prompts into it.
Optimization loop: A probabilistic Gaussian process surrogate drives an acquisition function that selects target feature vectors; the LLM then realizes and refines these into deployable system prompts.
Adaptive representations: The feature space is re-elicited as new evaluations arrive, so the representation adapts to the observed prompt-score history.

Evaluation

The setting is evaluated using offline benchmark accuracy as a controlled aggregate proxy: the optimizer observes one scalar score per prompt, with no per-example labels, errors, or critiques. Across ten system prompt optimization tasks with a total budget of 30 evaluations, ReElicit achieves the strongest aggregate performance profile among representative aggregate-only prompt-optimization baselines.

Conclusion

These results suggest that LLMs can serve as adaptive semantic representation builders, not only prompt generators, for Bayesian optimization over natural-language artifacts.