LLM‑based planning agents must commit to a plan before acting in a partially observable environment. Depending on assumptions about hidden states, two opposite commitment styles emerge: an aggressive style that issues a long plan based on inferred unobserved facts, and a conservative style that plans only over what has been observed. Both appear implicitly in pretrained models, whether the LLM outputs a plan directly or produces a PDDL description for a classical solver. Aggressive commitment can cause costly failures, while conservative commitment increases the number of LLM calls. We evaluate aggressive and conservative planners on two partially observable domains, supplying priors as either explicit statistics or common‑sense knowledge. Across three LLMs the final accuracy of both strategies is comparable. Conservative methods achieve higher safety with fewer failed attempts but require more API calls; aggressive methods fail more often but can replan efficiently when priors are informative. Analysis shows LLMs treat supplied priors almost as facts, which hurts performance when priors are unreliable. We also propose hybrid strategies with tunable hyperparameters that let users trade safety against efficiency.
Review