I wonder if LLMs should be trained to minimize unread text.

Unread text is probably easy to measure (e.g., the user just scrolled past).