Just speculating but proximity to a reference answer is a much denser reward signal. In contrast, parsing out a final answer into a pass/fail only provides a sparse reward signal.