StraGo: Harnessing Strategic Guidance for Prompt Optimization

Yurong Wu

Abstract:

Prompt engineering is pivotal in harnessing the
capabilities of large language models (LLMs)
for diverse applications. While existing prompt
optimization methods enhance prompt effec-
tiveness, they often induce prompt drifting,
where newly generated prompts may detrimen-
tally affect previously successful cases while
correcting failures. Moreover, these methods
heavily rely on LLMs’ intrinsic capabilities
for prompt optimization tasks. In this paper,
we introduce STRAGO (Strategic-Guided Op-
timization), a novel approach that mitigates
prompt drifting by leveraging both success-
ful and failed cases to identify critical fac-
tors to achieve the objectives. Furthermore,
STRAGO adopts a how-to-do approach, inte-
grating in-context learning to develop specific,
actionable strategies that offer detailed, step-
by-step guidance for prompt optimization. Ex-
tensive experiments across various tasks—such
as reasoning, natural language understanding,
domain-specific knowledge, and industrial ap-
plications—demonstrate STRAGO’s superior
performance. It sets a new state-of-the-art in
prompt optimization, showcasing its capability
to deliver stable and effective prompt improve-
ments