Skill0.5: Joint Skill Internalization and Utilization for Out-of-Distribution Generalization in Agentic Reinforcement Learning

Researchers introduced Skill0.5, a skill-based reinforcement learning framework for agentic systems.

It uses a difficulty-aware router to send hard tasks to skill internalization, medium tasks to standard RL, and easy tasks to diagnostic probing.

On ALFWorld and WebShop, Skill0.5 outperformed memory-based and prior skill-based methods in both in-distribution and out-of-distribution settings.

The approach aims to cut context costs while improving generalization by separating broad skill learning from task-specific skill use.