Can Language Models Serve as Text-Based World Simulators?

Broadly speaking, there are two ways to leverage LLMs in the context of world modeling and simulation. The first is neurosymbolic: a number of efforts use language models to generate code in a symbolic representation that allows for formal planning or inference (Liu et al., 2023; Nottingham et al., 2023; Wong et al., 2023; Tang et al., 2024). REASONING VIA PLANNING (RAP) (Hao et al., 2023) is one such approach – it constructs a world model using LLM priors and then uses a dedicated planning algorithm to decide on agent policies (LLMs themselves continue to struggle to act directly as planners (Valmeekam et al., 2023)). Similarly, BYTESIZED32 (Wang et al., 2023) tasks LLMs with instantiating simulations of scientific reasoning concepts in the form of large PYTHON programs. These efforts are in contrast to the second, and comparatively less studied, approach of direct simulation. For instance, AI-DUNGEON represents a game world purely through the generated output of a language model, with inconsistent results (Walton, 2020). In this work, we provide the first quantitative analysis of the abilities of LLMs to directly simulate virtual environments. We make use of structured representations in the JSON schema as a scaffold that both improves simulation accuracy and allows for us to directly probe the LLM’s abilities across a variety of conditions.

In a systematic analysis of GPT-4 (Achiam et al., 2023), we find that LLMs broadly fail to capture state transitions not directly related to agent actions, as well as transitions that require arithmetic, common-sense, or scientific reasoning. Across a variety of conditions, model accuracy does not exceed 59.9% for transitions in which a non-trivial change in the world state occurs. These results suggest that, while promising and useful for downstream tasks, LLMs are not yet ready to act as reliable world simulators without further innovation.1

We examine the abilities of LLMs to serve as world simulators in text-based virtual environments, in which an agent receives observations and proposes actions in natural language in order to complete certain objectives. Each text environment can be formally represented as a goal conditioned partially observable Markov decision process (POMDP) (Kaelbling et al., 1998) with the 7-tuple (S, A, T ,O,R,C,D), where S denotes the state space,