EvolvingWorld: An Open Schema Framework for Co Evolving Role Play Agents and World Model in Interactive Literary World
Masonry and Eyre dig into EvolvingWorld, a new framework that lets fictional characters and their world co-evolve across long stories. Eyre walks through the open-schema architecture and seven supervised tasks; Masonry sizes up who would actually build with this and where product pain lives. The hosts end up excited about the open-schema premise but skeptical of the benchmark’s generality.
Transcript
Masonry —okay, but hear me out, a framework where the world and the characters evolve together like an actual book, not just a chatbot playing dress-up—
Eyre Right, and you’re already imagining the indie studio buying it on day one.
Masonry Ha! yeah, exactly—
Eyre Yeah, no, I mean the premise is honestly clever …
Masonry I know, I know—
Eyre … but the devil’s in the details: how do you decide which dimensions even matter when Sherlock Holmes starts noticing the tobacco smell on the wallpaper and suddenly the whole study’s state changes?
Masonry Okay, fair—
Eyre And the paper admits that—open-schema means the system infers relevant character and world dimensions per book instead of shoving every Victorian orphan into the same fixed slots—
Masonry Which, to me, is the whole ballgame—
Eyre Mm-hm.
Masonry —because right now every role-play framework is stuck doing static persona imitation or isolated scene generation. EvolvingWorld tries to model the whole damn story as a long-horizon process where characters revise beliefs, motivations, relationships, and the world itself mutates—locations, objects, background conditions—all while the dialogue keeps flowing.
Eyre Right, and the paper’s framing is that the goal isn’t to reproduce the original book verbatim—
Masonry —it’s to sustain a grounded, continuously developing world.
Eyre …which is the part that usually collapses after the first few scenes.
Masonry I see your point—and that’s why the open-schema angle matters. Existing systems either hard-code profiles or update only a handful of predefined fields. EvolvingWorld’s Character Agent keeps hidden trackers for weak evidence—emotions might flip fast, but personality traits need a stack of scenes—
Eyre So you’re delaying the profile update until the evidence crosses a threshold.
Masonry Exactly—and the World Model maintains global state, location state, entity state, all with open schema so the detective’s investigative habits and the Victorian orphan’s social standing can evolve on their own timelines.
Eyre And it’s all supervised via seven tasks: scene initialization, interaction generation, state update across character and world—
Masonry —and the repo’s right there, so someone will.
Eyre Fine. If you ignore the data bottleneck—
Masonry —the mechanism itself is kind of elegant. Hidden trackers for soft evidence. World Model that keeps global and location/entity states open and mutable—
Eyre And the seven tasks stitch it together—
Masonry —which, in practice, means you could plug this into an interactive fiction engine and have actual evolution instead of reruns.
Eyre …assuming you’re okay with the eval being 222 snapshots from 57 books.
Masonry Okay, okay—
Eyre I’m not saying it’s bad. I’m saying it’s research-first.
Masonry And that’s the part people actually ship.
Eyre So you’re telling me you’d build a product on this?
Masonry No, I’m telling you someone will build a product on top of this framework once the dataset grows and the eval gap closes.
Eyre …which could be never.
Masonry Okay, but—
Eyre —or it could be the next CharacterLLM in two years.
Masonry Exactly—so that’s why we’re watching this space.
Eyre Fair.
Masonry And if anyone’s curious, the repo’s on GitHub under HKUST-KnowComp/EvolvingWorld—link in the show notes—and the paper’s got the dataset and eval details.