标签
本文介绍了EvoSkill-GUI,一个无需训练的框架,允许GUI代理通过执行中的反思、修订和复用来改进技能,在多个基准测试中展示了无需重新训练的性能提升。
This paper introduces the Human–LLM Reflection Framework (HRF) to compare human and LLM revision behavior, finding that LLM reflection often yields zero or negative information gain and behaves more like conditioned re-generation than genuine error-driven revision.