EvolvingWorld: An Open-Schema Framework for Co-Evolving Role-Play Agents and World Model in Interactive Literary World
Qing Zong, Yue Guo, Mengxin Yang, Yiwen Guo, Yangqiu Song
cs.CL
2026-07-19
EvolvingWorld models interactive literary worlds as a long-horizon process where open-schema character profiles and world state persistently update. With a Character Agent and a World Model covering 7 trainable tasks, it beats static persona-imitation baselines on long-range coherence.
Role-play agents in interactive fiction and narrative have two old playbooks. One is static persona imitation: slap fixed tags on a character and speak to type, leaving the character identical from start to finish. The other is isolated scene generation: each scene is produced independently and the seams between scenes never hold. Both miss what makes fiction work: characters and worlds change together over time. Characters revise beliefs, motivations and relationships; locations, objects and background conditions shift. EvolvingWorld sets out to model this long-horizon co-evolution.
The framework is open-schema: it does not force every character's profile into fixed slots, since characters vary in dimension across books. Each character gets an open-schema profile plus a hidden tracker that stores weak or emerging evidence separately, updating the profile only once enough evidence accumulates, which avoids being dragged around by noise.
Two coupled modules drive it. A Character Agent handles multi-character role-play and persistent profile evolution. An LLM-based World Model maintains global, location- and entity-level state and advances scenes.
Seven trainable tasks run in order: scenecast picks participating characters, locationscenario sets place and scenario, motivationupdate prepares per-character scene motivation, nextcharacter chooses the acting character, interactiongen produces speech, thought or action, worldupdate updates global and location state, and characterupdate revises character states. Data comes from 57 books, yielding 138,596 supervised samples and 222 test snapshots.
Evaluation uses a trajectory-level LLM-as-Judge across 10 dimensions and 20 metrics. The Character Agent covers six dimensions (consistency, evolution quality, environmental grounding, interaction quality, motivation generation, instruction compliance); the World Model covers four (scene planning, speaker management, state maintenance, instruction compliance).
The main comparison is BookWorld, which updates only predefined fields and lacks entity-level world updates. EvolvingWorld scores higher on average for both the Character Agent and the World Model, pulling ahead on character-evolution metrics (PUF, PES, MQ) and long-range scene continuity. The most telling result is behavior as trajectories lengthen: BookWorld degrades, while EvolvingWorld not only holds steady but sees its Profile Evolution Smoothness (PES) actually improve over longer trajectories.
For builders of role-play, interactive narrative and game NPCs, this paper flags an overlooked dimension: long-range coherence matters more than per-turn likeness. The open-schema design is pragmatic, not assuming all characters share a shape. Modeling character and world as coupled and co-evolving, rather than each minding its own, is what stabilizes the long horizon.
World state is single and objective, shared by all characters, with no subjective perception or imperfect memory. The world representation is bounded by LLM context length; each location tracks only "important entities," so long works lose things. The benchmark uses only public-domain classics; modern novels, games and UGC are untouched, so generalization to contemporary content is unverified. Evaluation relies heavily on LLM-as-Judge, and many of the 20 metrics are themselves model-scored; agreement with humans is not thoroughly discussed.