ANIMASK is a simulation framework that freezes novels or scripts into story worlds, lets characters act according to their own motivations, and replays each narrative from the freeze point. We keep the author’s continuation as a human reference and verify through in‑story interviews that each persona remains present. At every decision point we compare the character’s action when the persona is applied with the action generated by the model after removing the persona.
The study spans 40 stories, 6 language models, and a total of 3,846 decision points. All replays drift away from the canon in a shared direction: stories become flatter and cooler, leaving tensions unresolved. Personas stay obeyed and the character’s identity is preserved. In only three out of four examined choices does the model’s default already fall within what the persona accepts; otherwise the model behaves more cautiously, holding back where the persona would push forward.
Thus the persona guarantees who the character is, while the model determines how far the character will go. This provides a quantifiable benchmark for assessing a language model’s contribution to role‑play.
Review