LLM-based agents increasingly rely on persistent skills, i.e., reusable procedural prompts, to adapt without weight updates. Existing skill self-evolution me…
机构:复旦
来源:arXiv 2609.15396 | AI4Papers 论文推荐平台