This paper addresses the challenge of prompt engineering in maximizing the potential of large-scale language models (LLMs), especially in tasks that require subjective quality assessments, where explicit optimization objectives are difficult to define. While existing automatic prompt optimization methods are not effective for such problems, in this paper we present DEEVO, a novel prompt optimization framework that leverages discussion-based evaluation and Elo-based selection. DEEVO explores the discrete prompt space while maintaining semantic consistency through intelligent crossover and strategic mutation operations. It simultaneously pursues prompt improvement and diversity using Elo ratings as a relevance metric, and outperforms existing methods on both open and closed problems without correct-answer feedback. Combining the inference capabilities of LLMs and adaptive optimization, it contributes to continuously improving AI systems without a predefined metric.
