CASSANDRA: Programmatic and Probabilistic Learning and Inference for Stochastic World Modeling

Keywords: world model, llm, neurosymbolic

TL;DR: We propose a neurosymbolic approach to world modeling that leverages Large Language Models to write code and design probabilistic models to build state transition models for planning in stochastic domains.

Abstract: Building world models from limited data is crucial for planning in real-world environments such as businesses. Such environments have semantically rich relationships between variables and stochastic dynamics arising from unobservable factors. Consequently, these environments require world knowledge to efficiently model complex action-effects and the causal relationships between variables. In this work, we propose CASSANDRA, a neurosymbolic approach to world modeling that leverages a Large Language Model's (LLM) general knowledge to create lightweight state-transition models for model-based planning in stochastic domains. We achieve this through a novel, two-part process: 1) using an LLM to synthesize executable code that models deterministic environment features, and 2) leveraging the LLM as a structural prior to discover a probabilistic graphical model for the stochastic features that captures causal relationships between variables. By integrating these two components, we create a single, factored world model where the symbolic code computes the deterministic action effects and the graphical model captures the stochastic effects that propagate through the variables. We apply our method to two environments: i) a small-scale coffee-shop simulator, in which we perform a detailed examination of our method's operation, and ii) a complex theme park business simulator, where we demonstrate significant improvements in planning and modeling transition dynamics over baselines.

Submission Number: 713