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

Zhiyuan Jerry Lin, Benjamin Letham, Samuel Dooley, Maximilian Balandat, Eytan Bakshy

arXiv:2605.19093 · 2026-05-20 공개 · arXiv · PDF

llm prompt-optimization bayesian-optimization gaussian-process natural-language system-prompts embedding-by-elicitation re-elicitation

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. We study this aggregate feedback setting as sample-constrained black-box optimization over discrete, variable-length text. We introduce ReElicit, a Bayesian optimization framework based on \emph{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. Leveraging a probabilistic Gaussian process surrogate, an acquisition function then selects target feature vectors, which the LLM realizes and refines into deployable system prompts. Re-eliciting the feature space as new evaluations arrive lets the representation adapt to the observed prompt-score history. We evaluate the setting using offline benchmark accuracy as a controlled aggregate proxy: the optimizer observes one scalar score per prompt and no per-example labels, errors, or critiques. Across ten system prompt optimization tasks with a 30 total evaluation budget, ReElicit achieves the strongest aggregate performance profile among representative aggregate-only prompt-optimization baselines. These results suggest that LLMs can serve as adaptive semantic representation builders, not only prompt generators, for Bayesian optimization over natural-language artifacts.

한국어 요약

한 줄 요약

**[시스템 프롬프트 최적화 / Bayesian Optimization]** LLM이 동적으로 elicit한 해석 가능한 특징 공간 위에서 시스템 프롬프트를 Bayesian Optimization으로 튜닝하는 ReElicit 프레임워크 제안.

핵심 기여도

핵심 아이디어

LLM은 단순한 prompt generator가 아니라 "Bayesian Optimization을 위한 적응형 의미 표현 빌더"로 사용될 수 있다. 흐릿한 자연어 prompt 공간을 LLM이 동적으로 만들어낸 해석 가능한 feature 공간으로 옮겨 그 위에서 GP·acquisition을 돌리면, aggregate 점수만으로도 효율적으로 최적 프롬프트를 찾을 수 있다.

기술적 접근법

주요 결과

의의 및 한계

**의의**: LLM의 역할을 "출력 생성"을 넘어 "BO를 위한 의미 임베딩 구성"으로 확장, 자연어 아티팩트 최적화의 새로운 디자인 패턴 제시. **한계**: 특징 공간이 elicit 단계의 LLM 품질에 의존, 매우 좁은 budget·매우 큰 prompt 공간에서의 일반화는 추가 검증 필요.

실용적 활용