EvolvingWorld uses open-schema agents to co-evolve characters and worlds

tencent · hf · 2026-07-21

**EvolvingWorld** is a new open-schema framework for co-evolving role-play agents and world models in interactive literary worlds. - The paper argues that prior systems treat literary simulation as either static persona imitation or isolated scene generation, which misses how characters and worlds evolve together over long horizons. - EvolvingWorld models simulation as a persistent process with two coupled parts: - a **Character Agent** for multi-character role-play and evolving profiles, - an **LLM-based World Model** for global and local state maintenance plus scene progression. - The authors define **7 trainable tasks** covering scene initialization, interaction generation, and state updates. - They also build a dataset from **57 books**, yielding **138,596 supervised samples** and **222 test snapshots**. - Evaluation uses a trajectory-level **LLM-as-Judge** protocol with **10 dimensions** and **20 metrics**. - Experiments show better long-horizon coherence and more persistent character/world development.

Original post →

More from Multimodal

Multimodal channel →