A new research paper from scientists at UC Berkeley proposes an innovative approach dubbed Dynalang, which aims to design reinforcement learning agents that can learn a world model with the help of natural language.