EvolvingWorld: An Open-Schema Framework for Co-Evolving Characters and Worlds

Many interactive fiction systems are good at producing a convincing reply, but weak at preserving continuity. EvolvingWorld addresses that gap by treating literary simulation as an ongoing process: characters remember, locations change, and world facts must stay coherent across turns. Instead of assuming a fixed schema, the framework stays open enough to cover different kinds of literary worlds.

What the framework contains

• Character Agent: handles multi-character role-play and persistent profile evolution.
• World Model: maintains global state as well as location-level and entity-level updates.
• Open-schema design: avoids a rigid world template and better fits diverse narratives.
• Seven trainable tasks: cover scene initialization, interaction generation, and state updates.

Data and evaluation

The authors build a dataset from 57 books, yielding 138,596 supervised training samples and 222 test snapshots. For evaluation, they introduce a trajectory-level LLM-as-Judge protocol with 10 dimensions and 20 metrics. That matters because long-form simulation cannot be judged reliably from a single response alone; consistency has to be checked across the whole trajectory.

Why it matters

The paper sits at the intersection of role-play agents and world modeling. Its main contribution is not just richer dialogue, but a more explicit way to track how fictional worlds evolve over time. That is relevant for interactive novels, narrative agents, and any application where memory and state are as important as fluent generation.

At a broader level, EvolvingWorld reflects a shift in agent research: from isolated turns to persistent, stateful interaction. The challenge now is to make that persistence robust, general across domains, and measurable in a way that aligns with real storytelling quality.

Source: Hugging Face Daily Papers