From Kepler to Newton: Inductive Biases Guide Learned World Models in Transformers