Your examples teach format louder than content

Few-shot learning is pattern continuation. The model is extremely sensitive to surface form, sometimes more than to the actual reasoning you intended to demonstrate.

Make every example identical in structure, pick label words that are single clean tokens, and keep lengths similar. The more rigid the template, the more reliably the model fills in the next blank.