Tsinghua and Huawei Propose HiSME: A Hierarchical Skill Meta-Evolving Framework for Agents

Tsinghua University and Huawei's Foundation Model Team have proposed HiSME, a lightweight hierarchical framework that enables AI agents to automatically optimize their skill generation strategies from execution experience without updating model parameters. By introducing meta-skills that dynamically adjust the skill evolution algorithm itself, HiSME improves agent adaptability in tool use and long-horizon tasks. Experiments show that HiSME effectively boosts task performance and learns better evolution strategies autonomously.