LLMs Generate Heuristics for Hierarchical Task Network Planning

A recent investigation examines the capability of large language models (LLMs) to create effective search heuristics for Hierarchical Task Network (HTN) planning, building on previous research in classical planning. HTN planning breaks down high-level tasks through a library of methods until only executable actions are left, utilizing domain knowledge while encountering difficulties that exceed classical state-space search. Conducted by Corrêa, Pereira, and Seipp (2025), this study modifies their approach from classical to hierarchical planning. They assess heuristics produced by LLMs using the Pytrich planner across six standard total-order HTN benchmark domains. The findings are published on arXiv under ID 2605.07707.