A study of reflective prompt evolution (GEPA) in deep-research agents found that in-distribution gains often stem from the optimizer writing answers directly into prompts. While citation rewards increased by 0.079, nearly half of the winning prompt content was identified as memorized task answers.
HOW THIS AFFECTS YOU
●
builderBe wary of prompt optimization gains that do not generalize across topic-disjoint datasets.
●
researcherThis highlights a critical failure mode where prompt evolution mimics memorization rather than reasoning.