LeafSkill: Efficient Action Grounding for LLM Agents
October 26, 2026
LeafSkill uses a 1.7B parameter selector to decouple high-level planning from repetitive leaf-level action grounding. This external reuse layer retrieves and binds action templates to environments, allowing agents to execute tasks without constant large-model invocation.
HOW THIS AFFECTS YOU
●
builderYou can reduce the computational cost and latency of recursive agentic workflows by offloading grounding to a lightweight selector.