Planning With Linguistic Hierarchical Task Networks
Proceedings of the Thirteenth Annual Conference on Advances in Cognitive Systems
2026
Abstract
Hierarchical Task Networks (HTNs) provide structured representations of procedural knowledge, but their reliance on formal symbolic representations can limit reuse when the same procedural task is expressed in other environment representations. We introduce Linguistic HTN (L-HTN) planning, which represents tasks, methods, and operators in natural language and uses a large language model to interpret environment observations, evaluate method preconditions, and ground task arguments during planning. The planning process is constrained to the L-HTN's procedural structure while the language model performs flexible pattern matching with local semantic interpretation to decompose tasks. We evaluate our approach on two intelligent tutoring tasks, measuring whether unmodified L-HTNs can remain performant across controlled changes to the environment's representation. We also compare L-HTNs to traditional HTNs and unconstrained LLMs in each case. We find that while HTNs and unconstrained LLMs exhibit highly degraded performance for many representations, L-HTNs are able to flexibly apply their task knowledge to remain robust across variations.
BibTeX
@inproceedings{smith-acs-2026,
title = {Planning With Linguistic Hierarchical Task Networks},
author = {Smith, Glen and MacLellan, Christopher J.},
booktitle = {Proceedings of the Thirteenth Annual Conference on Advances in Cognitive Systems},
year = {2026},
}