Semantics-Behavior Coupled Bayesian Optimization for Efficient Black-Box Prompt Search

Abstract

Large language models (LLMs) are highly sensitive to prompt design, making prompt optimization important in black-box settings where models are accessible only through API calls. This thesis proposes DualBO, a semantics-behavior coupled Bayesian optimization framework for efficient prompt search. DualBO represents each prompt using two complementary views: a behavioral correctness vector computed on a controlled minibatch, and a semantic embedding of the prompt text. These views are combined in a dynamic Gaussian Process surrogate, where semantic similarity helps reduce uncertainty early in optimization and behavioral evidence provides stronger task alignment as more prompts are evaluated. DualBO also introduces staged strategy-oriented candidate generation, which first generates diverse rewriting strategies and then instantiates them into concrete prompt candidates. This improves search-space coverage compared with direct local rewriting.

Experiments on ETHOS, ARC, MMLU-Pro, and HotpotQA using GPT-5.4-mini and DeepSeek-V3.2 show that DualBO consistently improves over semantic-only and behavior-only Bayesian optimization baselines. It is also competitive with reflection-based methods such as Reflexion and ProTeGi while requiring substantially fewer model calls. Overall, this thesis demonstrates that combining semantic smoothness, behavior-grounded task alignment, and diversity-aware candidate generation improves the stability and efficiency of black-box prompt optimization under limited evaluation budgets.