Show, Don't Tell: Aligning Language Models with Demonstrated Feedback