Beyond Sample-Level Feedback: Using Reference-Level Feedback to Guide Data Synthesis