Task Knowledge for LLMs: Optimizing Behavior Tree Representations

A behavior tree (BT) is a compact and interpretable way to represent tasks and their structure. When a language model has to reason about those tasks, how the tree is written to it matters a lot. This thesis studies how to represent a BT to an LLM: which serialization to use (plain text, structured or abstracted summaries, key-value or pruned views), how much of the tree to include, and what makes the model’s decisions more accurate, consistent, and faster. The goal is a compact, effective task-knowledge representation that a model can read and work with reliably.