Can characters and worlds evolve together in long stories?

EvolvingWorld models interactive literary simulation as a long-horizon process in which "characters interact, scenes progress, and character and world states are persistently updated." The abstract sets this against existing systems that treat the task "as static persona imitation or isolated scene generation," which it says fail "to capture how characters and worlds evolve together over time." The introduction passage makes the same contrast. Role-playing agents can imitate fictional characters and sustain persona-grounded dialogue, but as a story unfolds "characters revise beliefs, motivations, and relationships" while the world around them changes too. The excerpt breaks off mid-list, at locations and objects. The conclusion claims the coupled design "leads to more coherent long-horizon simulations across diverse backbones."

The architecture has two coupled modules. A Character Agent handles multicharacter role-play and persistent profile evolution. An LLM-based World Model maintains global and location/entity-level state and drives scene progression. Both work over "open-schema" states instead of the fixed schemas the authors say prior systems rely on, which is what lets one framework span diverse literary worlds. Long-horizon interaction is decomposed into seven trainable tasks covering scene initialization, interaction generation and state update. The dataset comes from 57 books, with 138,596 supervised training samples and 222 test snapshots. Evaluation is a trajectory-level LLM-as-Judge protocol over 10 dimensions and 20 metrics. The conclusion credits explicit open-schema state and the seven-task decomposition as the "concrete foundation" for persistent world evolution.

This sits between two neighboring lines of work. On the persona side, Can layered persona architecture sustain coherent character behavior? deepens a persona's internal structure, while EvolvingWorld makes the profile something that changes and pairs it with a world that changes as well. Do personality-conditioned agents change like humans do? measures how far conditioned agents shift on their own. EvolvingWorld instead makes profile evolution an explicit state-update task. On the world side, Can language models learn to simulate agent environments? also treats a trained language model as the other half of the loop, and What should a world model actually be designed to do? supplies the framing this design leans toward, since state maintenance and scene progression are closer to simulation than to prediction. EvolvingWorld is the narrative-domain instance of the same coupling.

The excerpt gives no scores, baselines, ablations or backbone names, so "more coherent" is a stated conclusion rather than a visible result. It also does not say whether the LLM-as-Judge protocol was checked against human readers. The introduction is cut off, so how the authors motivate the task beyond the contrast above is not available. Only the first of the three stated limitations appears. The world is modeled as "a single objective state shared by all characters," although literary characters perceive and remember it differently: one sees a benevolent world where another sees a hostile one, or misremembers an object as on a table when it is on a chair. Separate perceived worlds per character would "substantially increase system complexity." So the co-evolution claim covers a shared ground-truth world, and whether subjective, per-character worlds can be maintained at acceptable cost is left open.