Evidence Accumulation Before Character Change
Making large language models play novel characters sounds like a solved problem: give a character card, converse, imitate the voice. But existing systems share the same flaw—the persona is fixed once at the start. By chapter ten, Lin Daiyu is still the opening-scene Lin Daiyu; a house burned down in the story is still standing in the next scene. The literature calls this static persona imitation: the model is reciting the character, not playing it.
A team from Hong Kong University of Science and Technology (working with LIGHTSPEED, Huazhong University of Science and Technology, and independent researchers) reframed the problem: simulation of interactive literary worlds should not be a replay of scenes from the original book, but a long-horizon process—characters change through interaction, the world changes through events, and the two must stay consistent. They built this idea into a complete framework plus benchmark called EvolvingWorld.
The character side: separate fast from slow
How do real people change? Emotions shift within a day; personality takes years to accumulate. EvolvingWorld equips character agents with a multi-timescale evolution mechanism, centered on a hidden tracker.
How it works: each dimension of a character's profile (emotion, relationships, personality, ...) has its own "volatility" parameter. Weak signals emerging from interaction—an uncharacteristic remark, a hesitant gesture—do not directly rewrite the persona. Instead, they are stored in the hidden tracker as weak evidence and accumulated separately. Only when enough evidence has built up does a formal state update trigger.
The design is almost common-sensically simple, yet it is exactly what previous systems lacked: either no change was allowed at all (characters stayed static forever), or a single line could derail the persona (after one argument yesterday, "hot-tempered" appears in today's profile). Distinguishing "a momentary fluctuation" from "a trend" relies precisely on storing evidence at two levels.
The character dimensions themselves are also open: rather than a fixed template requiring every character to have personality/goal/relationship attributes, the model decides what dimensions to track based on the book's genre and setting, and even allows environments and character groups as special interaction units—necessary when writing ensemble scenes.
The world side: houses contain rooms
The other coupled module is an LLM world model. It maintains two layers of state simultaneously: global settings (historical background, social institutions) and location-level physical state. Locations can nest—a house contains rooms; the road outside is an independent atomic location. Each location stores a detailed description and tracks all significant non-character entities and their states; the paper's example is "the Christmas tree stands nearby."
The key is the update method: all state evolves automatically with character interaction, with no manual sandbox construction. Swap in a completely different book—from a period drama to interstellar sci-fi—and the framework still applies. This follows the same principle as the open schema on the character side: structure grows out of content, not from pre-set scaffolding.
Breaking the whole pipeline into seven tasks
To make all of this trainable, the team decomposed the full simulation pipeline into seven supervisable tasks: casting (the world model selects appearing characters from the full cast), scene setting (choosing location and situation), motivation updating (each participant enters with a scene-specific motivation), followed by interaction generation and state updating. The dataset was built from 57 books, yielding 138,596 supervised training samples and 222 test snapshots; the repository is open-sourced at github.com/HKUST-KnowComp/EvolvingWorld.
The evaluation is the part I most admire: a trajectory-level LLM-as-Judge protocol spanning 10 dimensions and 20 metrics. It measures not whether a single scene is well played, but whether the story remains coherent after the long run—whether the character is still that person, whether the world still remembers what happened.
The results in one sentence: on Qwen-7B and Llama-8B, agents trained with EvolvingWorld data substantially outperform the previous best role-play baselines (CoSER and Crab), and the performance decay common in long-horizon simulation is markedly reduced. One judgment from the paper deserves to be quoted on its own: long-horizon role-play requires not better dialogue imitation, but modeling the coupled evolution of character and world state.
Coda
Looking back, the smartest component in this framework is the least conspicuous one—the hidden tracker: it acknowledges that not every line counts, letting weak evidence accumulate quietly until it is ready to act. This is nearly how novelists handle foreshadowing—planted without a sound, revealed only when it ripens on its own.
---
*Source: EvolvingWorld: An Open-Schema Framework for Co-Evolving Role-Play Agents and World Model in Interactive Literary World (arXiv:2607.17250, 2026-07-19, Qing Zong / Yue Guo / Mengxin Yang / Yiwen Guo / Yangqiu Song; HKUST + LIGHTSPEED + HUST + independent researchers). Benchmark: 57 books, 138,596 training samples, 222 test snapshots. Evaluation: trajectory-level LLM-as-Judge, 10 dimensions, 20 metrics. Baselines include CoSER and Crab. Mechanism descriptions drawn from the abstract and Sections 2–3.*