An arXiv preprint examines how LLM-based planning agents decide on plan length in partially observable environments. The abstract states that agents commit to plans of different lengths before interacting with the environment.
Some agents, the abstract notes, aggressively commit to a long plan based on assumptions about unobserved states. The text is cut off after "while others…", so the alternative strategies are not described in the available source.
Because the abstract is truncated, the full methodology and conclusions are not yet accessible. The paper is a preprint and has not been peer-reviewed.