EvolvingWorld: An Open-Schema Framework for Co-Evolving Role-Play Agents and World Model in Interactive Literary World
EvolvingWorld tests whether role-play agents can keep characters and settings changing coherently over long literary simulations.
The paper frames interactive fiction as a persistent process, not one-off persona imitation or scene generation. Its system pairs character agents with an LLM-based world model that tracks global, location, entity, and profile state. The authors build training tasks around scene setup, interaction generation, and state updates, using data from 57 books. They report improvements on long-horizon coherence under a trajectory-level LLM-as-judge evaluation. HF Daily Papers' note
The paper frames interactive fiction as a persistent process, not one-off persona imitation or scene generation. Its system pairs character agents with an LLM-based world model that tracks global, location, entity, and profile state. The authors build training tasks around scene setup, interaction generation, and state updates, using data from 57 books. They report improvements on long-horizon coherence under a trajectory-level LLM-as-judge evaluation. HF Daily Papers' note
score 4