ReElicit: Bayesian Optimization for System Prompts
A new AI research paper introduces ReElicit, a Bayesian optimization framework for tuning system prompts using aggregate feedback. System prompts are crucial for controlling AI behavior but are hard to optimize when only scalar scores are available. ReElicit uses an LLM to elicit a compact feature space, maps prompts into it, and employs a Gaussian process surrogate to select target features. The feature space is re-elicited as new evaluations arrive, allowing dynamic adaptation. The paper is available on arXiv under ID 2605.19093.
Key facts
- ReElicit is a Bayesian optimization framework for system prompts.
- It uses embedding by elicitation to create a feature space.
- The framework handles aggregate feedback as scalar scores.
- A Gaussian process surrogate selects target feature vectors.
- The feature space is re-elicited with new evaluations.
- System prompts are discrete, variable-length text.
- The paper is on arXiv with ID 2605.19093.
- The approach is sample-constrained black-box optimization.
Entities
Institutions
- arXiv