EvolvingWorld presents an open-schema framework and benchmark for long-horizon simulation in interactive literary worlds. It couples a multi-character role-play agent, persistent character-profile evolution, and an LLM-based world model that maintains global, location-level, and entity-level state while progressing scenes. The dataset is built from 57 books, with 138,596 supervised training samples and 222 test snapshots. The authors define seven trainable tasks and evaluate trajectories using an LLM-as-Judge protocol covering 10 dimensions and 20 metrics. The abstract reports improved persistent and coherent character and world development, though detailed experimental comparisons are not provided here.
No heat snapshots are available in the last 24 hours.