This shouldn't be surprising. Extrapolation outside the radius of convergence has been known to be far less reliable than interpolation within. LLMs are forced to extrapolate.