“The Self-Taught Optimizer (STOP) framework improved performance when using GPT-4 as the base model, but degraded performance when using weaker models like GPT-3.5 and Mixtral.”