The goal of the paper was not to get the best pattern fitter, but to see if it's even possible to do symbolic regression with a simple token-to-token prediction method.