Hierarchical task knowledge enables more efficient problem solving by reusing previously learned skills. In this work we explore curriculum‑guided acquisition of such knowledge within an HTN (Hierarchical Task Network) learning framework. Three strategies are compared: online learning during problem solving, human‑crafted curricula, and an automated curriculum generation pipeline that leverages verified LLM‑generated solution traces.
The automated pipeline works as follows: small training problems are solved by an LLM, task accomplishment information is extracted, and this information is transformed into ordered sequences of state‑task learning problems that serve as input for HTN method acquisition.
We evaluate the three strategies across several planning domains. Results show that automatically generated curricula recover most of the benefit of expert‑designed curricula while dramatically reducing the token cost of online hierarchical knowledge acquisition.
Review