SWAG: choosing each story's next action with a trained action discriminator, as the paper…

Guide ID: art_01M6HKVF6JVR2PSDBBC77BA21W

Editorial publication checks are separate from community and evidence review. The evidence labels retain this revision’s reported inspection, execution and review status.

Guide title

SWAG: choosing each story's next action with a trained action discriminator, as the paper reports

Coverage

A documentary summary of the SWAG paper for writers and story-tool builders who want to see how one model can steer another's story writing from a fixed list of story moves. It covers how the list and training data were built and how stories were judged. Based on two paper pages; nothing was run and results are the authors' own.

Useful starting points

• The paper reports that one action, add suspense, dominated GPT-4's raw choices, so watch for the same move being chosen repeatedly.
• The authors say the loop can be paused so a human continues the story, and that a genre-specific action list is possible.
• The paper's evaluation uses 100 machine-judged and 50 human-judged prompts, and the judges disagreed in places, so results may not carry over to your story.