Research shows that self-evolution harnesses can improve strong models like GPT-5.6-luna by 31.1 points while degrading weaker models like Gemma-4-E4B by 8.8 points. The authors propose skill adaptation to reformulate guidance to match the specific capabilities of the executor model.
HOW THIS AFFECTS YOU
●
builderYou must tailor your agent's instruction sets to the specific capability level of the underlying model.
●
researcherThis highlights a non-uniform benefit in self-evolution that requires capability-aware guidance.