Use case · Idea to test
Stories and interactive fiction
Learn which style of story a reader enjoys.
- Compare short passages
- Learn reader preferences
- Rank new continuations
A reading app could learn that someone prefers short mysteries, quiet humour or a particular writing style. An interactive story could offer a few continuations and learn from the reader's choices.
Where the small model fits
A language model writes candidate passages. A small preference scorer ranks their representations. This is a ranking approach; it does not assume every language model exposes controls that can be edited into a desired story.
What would need to work
Use blind reader comparisons on new passages, checking continuity and variety as well as enjoyment. A passage can score well while contradicting the rest of the story.
Try the related work
This demo learns from simple drawings. It does not run the application described above.