LeafSkill preprint reports higher agent task completion with fewer model calls

Researchers report in a preprint that LeafSkill, a system that reuses executable actions across tasks, improved an AI agent's completion rate on 200 Robotouille tasks. It completed 137 tasks versus 121 for ReCAP, or 68.5% versus 60.5%. LeafSkill retrieves related experience, adapts action templates to available objects and permitted actions, and uses a 1.7B selector to choose an action or return the subtask to the original agent. On the 106 tasks both systems completed, it reduced Qwen3.5-27B tokens by 14.42% and calls by 14.08%. Those resource savings apply to jointly completed tasks, rather than the entire test set.