Stochastic Differential Equation Analysis of World-Model Generalization
A theoretical framework that models the training of a reinforcement-learning world model as a continuous-time stochastic dynamical system, using the drift and diffusion terms of a stochastic differential equation to characterise how errors in the learned latent representation affect robustness and generalization. Zero-drift errors are shown to behave as implicit regularization, while a Jacobian regularization term is introduced to control compounding error propagation under non-zero drift.
2501.00195
World models learn a latent dynamics model from observations and actions and then optimise a policy inside that learned model, forming the basis of model-based reinforcement learning methods such as …