You Live More Than Once: Towards Hierarchical Skill Meta-Evolving

Researchers propose HiSME, a hierarchical skill meta-evolving framework that enables AI agents to continuously improve both their skills and the strategies used to evolve those skills at test-time, without expensive model parameter updates. The approach learns meta-skills from task execution traces and demonstrates higher-quality skill libraries compared to static skill evolving approaches.

This research addresses a fundamental challenge in deploying agentic AI systems: the ability to improve performance over time in diverse downstream scenarios without retraining underlying large language models. Traditional approaches either rely on hard-coded skill strategies or computationally expensive parameter updates, both limiting practical deployment flexibility.

HiSME introduces a two-level optimization framework where agents not only refine individual skills but also learn meta-level strategies about how to evolve those skills. By analyzing task execution traces, the system extracts patterns about what works in different contexts, enabling adaptive behavior without model retraining. This lightweight algorithmic approach represents a meaningful shift toward more sustainable agent development.

For the AI development community, this work has practical implications. The ability to meta-learn skill evolution strategies means deployed systems can adapt to new domains more efficiently, reducing operational costs and deployment friction. The framework's hierarchical design suggests scalability advantages—organizations could deploy agents that improve autonomously rather than requiring continuous model fine-tuning cycles.

The research validates these improvements across multiple benchmarks, demonstrating that meta-evolving produces qualitatively different and superior skill libraries compared to baseline approaches. The anonymous code release suggests active research momentum in this area. Future development hinges on whether meta-skills derived in controlled benchmark settings transfer effectively to real-world agent deployments and whether the approach scales to more complex task distributions.

• →HiSME enables test-time optimization of both skills and skill evolution strategies without expensive LLM parameter updates.
• →Meta-skills learned from execution traces allow AI agents to adapt their evolution strategies to different scenarios automatically.
• →The approach produces higher-quality skill libraries than conventional skill evolving methods across diverse benchmarks.
• →Lightweight algorithmic adaptation reduces deployment friction compared to traditional parametric learning approaches.
• →The framework supports continual experience learning by deriving scenario-specific meta-skills from agent behavior data.