NEWS · RESEARCH · #345
ANIMASK: measuring model vs. persona contributions in simulated story worlds
The paper introduces ANIMASK, a simulation framework that freezes books and scripts into story worlds and replays each story from the freeze point to compare characters' actions with and without an assigned persona. Across 40 stories, 6 actor models, and 3,846 decision points, the authors find replays drift from their canons toward flatter, cooler outcomes while personas remain present and the model's defaults are typically more cautious than the persona.
KEY POINTS
- The paper introduces ANIMASK, a simulation framework that freezes books and scripts into story worlds and replays each story from the freeze point to compare characters' actions with and without an assigned persona.
- Across 40 stories, 6 actor models, and 3,846 decision points, the authors find replays drift from their canons toward flatter, cooler outcomes while personas remain present and the model's defaults are typically more cautious than the persona.
- This helps disentangle what persona prompts change versus what defaults the base language model imposes, informing controllability and evaluation of role-play and behavior in LLMs.
WHY IT MATTERS
This helps disentangle what persona prompts change versus what defaults the base language model imposes, informing controllability and evaluation of role-play and behavior in LLMs.