HiSME: Hierarchical Skill Meta-Evolving for Agent Systems
A recent study presents HiSME, a streamlined hierarchical skill meta-evolving approach aimed at improving deployed agentic systems. The findings, available on arXiv (2605.28390), emphasize the importance of refining the skill evolving framework during testing to achieve ongoing enhancements in various downstream contexts. HiSME focuses on the simultaneous optimization of skills and the evolving strategy by extracting meta-skills from the execution traces of agents. Tests conducted on a range of agentic benchmarks reveal that meta-evolving yields superior skill libraries compared to traditional skill evolving and generates a variety of meta-skills tailored for different scenarios, thereby supporting future continual experience learning.
Key facts
- HiSME is a lightweight hierarchical skill meta-evolving solution
- It jointly optimizes skills and the skill evolving strategy
- Meta-skills are learned from agents' task execution traces
- Experiments show meta-evolving produces higher-quality skill libraries
- The paper is published on arXiv with ID 2605.28390
- Test-time skill evolving is a new paradigm for enhancing deployed agentic systems
- Existing works focus on hard-coded strategies or parametric learning
- Lightweight algorithmic adaptation is feasible for skill evolving framework refinement
Entities
Institutions
- arXiv