1
GRASP: Gated Regression-Aware Skill Proposer for Self-Improving LLM Agents
LLM自我改进新方法!门控机制确保新增技能不破坏已有能力,破解程序性知识遗忘难题。
arXiv:2605.29668v1 Announce Type: new Abstract: LLM agents acting in structured environments fail in operational rather than conversational ways, and …