Abstract:
Generally, artificial agents can only successfully navigate a system insofar as they have some degree of understanding of the system. This understanding can be perceived as a world model. Determining whether or not language models have the ability to develop and act within a world model is significant because it determines how we react to and utilize language models. We explore ways to test the possibility of world models, and how language models may be able to establish world models via the development of natural language environments. Furthermore, we propose specific research directions for these objectives.