GEPA-GAN: Teaching AI to sound human | Decagon

This post details the application of a GAN-like adversarial game to prompt optimization for simulating human-like customer support interactions. It introduces the GEPA (Reflective Prompt Evolution) framework, which uses LLM reflection to automatically improve prompts for a generator and a judge. The system achieved a 34% human fool rate, demonstrating improved realism without model fine-tuning. Key technical contributions include the adversarial setup (generator, judge, reflector), the GAN-like game mechanics (without model weight changes), the non-monotonic dynamics observed during optimization, and the identification of specific prompt improvements related to preserving user language, tone, and avoiding support-like phrasing. The post also highlights the methodological takeaway that realism must be validated outside the game and that prompt length penalty can be a useful regularization technique.