Towards Learning High-Precision Least Squares Algorithms with Sequence Models