Planning Stories Neurally

Abstract

Symbolic planning algorithms and large language models have different strengths and weaknesses for story generation, suggesting hybrid models might leverage advantages from both. Others have proposed using a language model in combination with a partial order planning style algorithm to avoid the need for a handwritten symbolic domain of actions. This paper offers a complementary approach. We use a state space planning algorithm to plan coherent multi-agent stories in symbolic domains, with a language model acting as a guide to estimate which events are worth exploring first. We evaluate an initial implementation of this method on a set of benchmark problems and find that the LLM’s guidance is helpful to the planner in most domains.