Tournament of Prompts: Evolving LLM Instructions Through Structured Debates and Elo Ratings

The paper introduces DEEVO, a debate-driven evolutionary framework for optimizing prompts for Large Language Models without relying on predefined metrics, addressing challenges in subjective and complex tasks. DEEVO uses Elo-based evaluation, intelligent genetic operations, and debate feedback to outperform manual and existing automated methods, advancing prompt optimization by enabling adaptive, metric-free improvement.