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

EvolvingWorld enables long-horizon literary simulation by co-evolving character profiles and world states through an open-schema framework.

How can we enable interactive literary agents to maintain consistent, evolving character profiles and world states over long, multi-character narrative sequences?

Existing role-play systems treat characters and worlds as static, failing to capture how they develop together over the course of a story. When a character changes their beliefs or a location is physically altered, these systems typically lack the mechanism to propagate those changes persistently. EvolvingWorld models literary simulation as a long-horizon process using two coupled modules: an open-schema Character Agent that tracks evolving traits via hidden evidence, and an LLM-based World Model that maintains global and entity-level physical states. The framework decomposes this into seven trainable tasks, from scene initialization to state updates, allowing the simulation to adapt to diverse literary genres without fixed schemas. Experiments show that this co-evolutionary approach significantly reduces long-horizon performance degradation compared to static persona-based baselines, maintaining higher consistency across multi-scene trajectories.

Paper Primer

The framework hinges on the "open-schema" design: instead of forcing characters into predefined slots like "BDI" or "psychological state," the system allows the LLM to infer and evolve dimensions specific to the book's genre, such as a detective's investigative habits or a Victorian orphan's social standing.

To manage multi-timescale evolution, the Character Agent uses a hidden tracker: it records emerging evidence separately from the active profile, preventing premature updates until repeated signals across scenes justify a permanent shift in personality or trait.

EvolvingWorld reduces long-horizon performance degradation in multi-agent literary simulations.

Comparison against baselines like BookWorld shows superior performance on trajectory-level metrics, particularly in character evolution smoothness and scene continuity. Across matched backbones, the framework consistently outperforms role-play-only baselines, with significant gains in character consistency and world-state maintenance.

Why is a "World Model" necessary if the LLM already knows the story?

Standard LLMs are trained for narrative continuation, not structured state tracking; the World Model explicitly maintains global and location-level physical states to ensure that events like an object being moved or a social order shifting remain consistent across scenes.

What is the scope of this framework — can it handle modern novels or user-created worlds?

The framework is domain-agnostic because it uses no predefined attribute slots, though the current benchmark is limited to public-domain classics due to copyright constraints; the authors demonstrate consistent gains on out-of-domain books, suggesting it can adapt to other narrative domains.

Researchers should shift from treating role-play as static persona imitation to modeling it as a coupled, long-horizon evolution of character and environment states.

Introduction to EvolvingWorld

We expose the need for dynamic co‑evolution of characters and worlds in interactive literary simulation.

Current interactive literary simulators either freeze character personas or generate isolated scenes, leaving the world static. This static view collapses when a story stretches across many scenes, because characters’ beliefs, motivations, and relationships must adapt while locations, objects, and social orders also shift. The central motivation is therefore to move from static persona imitation to a dynamic, open‑schema co‑evolution of characters and worlds.

It treats characters and the world as open‑schema entities that continuously update together as the story unfolds.

**Figure 1.** A simulation case from EvolvingWorld. Starting from a snapshot derived from a book, multiple characters interact as both character and world states evolve.

The shift from static persona imitation to dynamic co‑evolution unlocks long‑horizon literary simulation that remains coherent as characters and worlds grow together.

The EvolvingWorld Framework

We replace static personas with open‑schema co‑evolution of characters and worlds.

Prior role‑playing agents treat characters as immutable profiles, which quickly stalls long‑horizon storytelling. Our trick is to let both characters and worlds evolve through an open‑schema state that grows and changes as the narrative unfolds.

Think of it as a spreadsheet that can add new columns on the fly, letting any new attribute appear when the story demands it.

Step 1: The hidden tracker records the hint “inherits heirloom” as weak evidence.

Step 2: After two more scenes reinforce the inheritance, the tracker promotes the evidence.

Step 3: The open‑schema state adds a new field

Step 4: Subsequent interactions can now reference

The schema grows only when the narrative supplies reliable evidence, preventing premature or spurious attribute creation.

Why not update the profile directly instead of using a hidden tracker?

The hidden tracker separates tentative hints from confirmed changes, so a single noisy mention (e.g., a fleeting thought) does not instantly rewrite the character’s core attributes.

The agent is a per‑character controller that reads the shared open‑schema state, decides a motivation, and emits dialogue or actions.

How does the Character Agent differ from a static persona model?

A static persona has a fixed set of attributes; the Character Agent’s open‑schema profile can grow, shrink, or rename fields as the story progresses, allowing truly dynamic personalities.

The model is a global‑plus‑location state keeper that updates the world after each interaction, much like a game engine that rewrites the scene description.

Why use an LLM‑based World Model instead of a handcrafted sandbox?

The LLM can generate rich, context‑aware descriptions for any location on the fly, scaling to diverse literary worlds without hand‑authoring each environment.

It works like a notebook that collects faint clues until enough evidence accumulates to justify a profile change.

What would happen if the hidden tracker were omitted?

Without the tracker, any single mention could instantly alter a character’s profile, leading to erratic or implausible personality shifts.

Benchmark Results and Analysis

EvolvingWorld delivers higher coherence scores than static baselines across long-form simulations.

Profile Evolution Smoothness (PES) measures how consistently a character’s traits evolve over time, while Scene Continuity & Coherence (SCC) gauges the logical flow of events across consecutive scenes.

EvolvingWorld improves Profile Evolution Smoothness over the static BookWorld baseline.

Across 15‑scene simulations, PES rises from 78.36 ± 7.78 (BookWorld) to 84.86 ± 5.34 (EvolvingWorld).

**Figure 3:** Length-wise comparison between **Evolving-World** and BookWorld on Profile Evolution Smoothness (PES) and Scene Continuity & Coherence (SCC). Shading denotes mean $\pm$ std/4.

EvolvingWorld consistently outperforms static baselines in long‑form narrative coherence.

Framework Rationale and Clarifications

Appendices detail partial support rationales, training setup, data pipeline, dataset stats, and evaluation protocol.

BookWorld receives a partial‑support mark for profile updates because it only modifies a few character fields, leaving broader profile information and relational data untouched. Its global world state updates are limited to a single event, and location/entity‑level modeling is reduced to brief descriptions without detailed entity updates.

GenerativeAgents is marked partial for profile updates since it confines updates to predefined memory fields, and it is partial for scene initialization because its simulation granularity is daily rather than scene‑by‑scene, affecting only routine activities.

CharacterBox earns a partial‑support label for profile updates because it only revises fixed BDI attributes and position, ignoring broader traits, and it lacks location/entity‑level modeling as it simulates a single scene with a solitary environment description.

All EW models are fine‑tuned with LoRA (rank 64, alpha 128, dropout 0.05) in LLaMA‑Factory for two epochs at a learning rate of 2 × 10⁻⁵, using a maximum sequence length of 32,768 tokens, bf16 precision, gradient checkpointing, and FlashAttention‑2. Validation holds out 10 % of the merged data, and inference is served via vLLM with 50 concurrent requests.

The data construction pipeline consists of three main phases: book selection, scene extraction, character construction, and world construction, followed by interleaved cleaning steps that refine interactions, remove duplicates, and enhance scenes.

We select 57 books from Goodreads’ “Best Books Ever” list and retrieve their full texts from Project Gutenberg, preserving chronological order to align temporal scene progression.

Phase 1 splits each book into LLM‑processable chunks, extracting structured scenes that include summaries, scenario descriptions, key characters, and multi‑turn interactions, while preserving cross‑chunk continuity.

Phase 2 standardizes character aliases, assigns a single official name per character, and initializes open‑schema profiles from early scenes; profiles are then revised after each scene based on new events and later revelations.

Phase 3 standardizes location aliases, builds an initial global and per‑location world state, and updates these states interaction by interaction within scenes.

Interleaved cleaning refines interaction content (clarity, perspective, redundancy), merges consecutive interactions from the same actors, removes duplicate scenes caused by chunk boundaries, and enriches scenario descriptions with dramatic details.

After constructing character, global, and location states along the book timeline, we translate the timeline into task‑level supervision: scene‑level tasks (`scene_cast`, `location_scenario`, `motivation_update`) and within‑scene tasks (`next_character`, `interaction_gen`, `world_update`, `character_update`), storing multi‑turn conversations in ShareGPT format.

**Table 6.** Overview of EvolvingWorld dataset statistics.

**Table 7.** The 57 selected books from Goodreads' Best Books Ever list.

**Table 8.** Training data statistics per task. "Samples" = number of ShareGPT style conversations; "Asst. Turns" = total assistant responses (equals Samples for single-turn tasks).

**Figure 7.** Token length distribution per task.

**Table 9.** Task-to-metric mapping for the Metric Penalty. When a task causes simulation termination, its corresponding metrics receive the additional penalty described above.

The open‑schema design yields 581 unique character dimensions and 136 world‑state dimensions, demonstrating flexibility in capturing diverse narrative attributes.

Genre analysis shows Literary Fiction dominates (≈ 60 %), followed by Adventure, Mystery & Gothic, Drama, and Fantasy, with thematic sub‑categories further detailing narrative focus.

Extracted data statistics indicate wide variability across books (median 160 scenes, 43 characters, 25 locations), while scene‑level interactions remain modest, reflecting realistic narrative density.

Token‑length distributions per task (Figure 7) inform the 32,768‑token cutoff, ensuring that even the longest‑tail samples (`scene_cast`) fit within computational limits.

Evaluation follows a multi‑scene simulation protocol: starting from a held‑out snapshot, the system generates full trajectories—including scene casts, interactions, and persistent state updates—to assess long‑term coherence.

Trajectory‑level scoring uses a 10‑dimension, 20‑sub‑metric LLM‑as‑Judge framework, separating Character Agent (6 dimensions, 11 sub‑metrics) from World Model (4 dimensions, 9 sub‑metrics) to pinpoint planning versus embodiment errors.

Error Penalties and Simulation Robustness

Limits of the evaluation: penalties, metric aggregation, and benchmark tables.

The core premise—open‑schema states let characters and the world co‑evolve dynamically—remains the backbone of the framework.

When a simulation aborts early, the metric penalty injects a zero‑score placeholder into the averaging process so that the failure directly drags down the final score.

How does the Metric Penalty differ from simply discarding failed scenes?

Discarding a failed scene would raise the average of the remaining scenes, effectively rewarding the model for avoiding the penalty. The Metric Penalty, by inserting a zero‑score entry, forces the average to reflect the failure, ensuring that early termination cannot be hidden by selective reporting.

The error‑penalty pipeline first applies the $IC$ Penalty for format violations, then adds the Metric Penalty described above; together they lower both the $CHARACTER$ and $WORLD$ scores.

**Table 10.** Overview of the evaluation framework. * denotes evaluation perspectives introduced by this framework. Grand total: 10 dimensions, 20 sub-metrics.

**Table 11.** Full benchmark results for Character Agent evaluation. I and P denote Instruct and Preview in model names. The best performances within the same model scale are bold-faced, and the second-best are underlined.

**Table 12.** Full benchmark results for World Model evaluation. I and P denote Instruct and Preview in model names. The best performances within the same model scale are bold-faced, and the second-best are underlined.

**Table 13.** Ablation study on Character Agent performance. “Full” denotes the complete EvolvingWorld framework; “w/o Char State” removes character state updates; “w/o Both” removes both world and character state updates.

**Table 14.** Ablation study on **World Model** performance. “Full” denotes the complete EvolvingWorld framework; “w/o World State” removes world state updates; “w/o Both” removes both world and character state updates.

**Table 15.** Character-side component ablations with GPT-5.3-Chat. Removing the Hidden Tracker or replacing the open profile with a fixed schema both degrade profile-update quality.

Ablation Studies

Ablation of the Open‑Schema component quantifies its impact on world‑side metrics.

We evaluate three world‑side quality metrics—CSR (Cast Selection Rationality), LSR (Location & Scenario Rationality), GUS (Global Update Sensitivity), GSA (Global State Accuracy), and the overall Avg.—to measure how the Open‑Schema design influences simulation fidelity.

Removing the Open‑Schema component degrades most world‑side metrics while slightly improving GSA.

Table 16 shows drops of 0.94 CSR, 0.80 LSR, and 3.52 GUS points; Avg. falls by 20.33 points, whereas GSA rises from 51.11 to 72.63.

**Table 16.** World-side schema ablation with GPT-5.3-Chat. Replacing the open global state with a fixed schema reduces global-state maintenance.

**Table 17.** ID and OOD performance after full-mixture training on Character Agent and World Model. Untrained averages are included as references.

**Table 19.** Full World Model benchmark results under three judge models. Gray rows indicate the judge model; all score rows are copied from the original per-judge tables, which are retained separately.

Hidden Tracker validation shows that removing the tracker reduces PUF by 12.60 points and PES by 2.60 points, confirming its role in aggregating weak evidence before committing profile updates.

Schema Effects analysis reveals that a fixed schema lowers the focused averages by roughly 1.3 points, demonstrating that book‑specific dimensions provide a modest but consistent advantage over a single generic schema.

Model Rankings and Evaluation

Top‑6 model rankings show near‑perfect agreement across judge models.

Top‑6 model rankings are identical across the three judge models for World Model and nearly identical for Character Agent.

Table 20 shows the same ordering of the top six models for each judge, with only a minor swap between Kimi‑K2.5 and GPT‑4o in the Character Agent list for GPT‑5.1‑Chat.

Full‑mixture training consistently raises both ID and OOD scores for every backbone, confirming that the learned book‑to‑world abilities generalize beyond the training distribution.

**Table 20.** Top-6 model rankings for Character Agent and World Model evaluations under three judge models, sorted by the scores in the Average column of Tables 18 and 19.

**Table 21.** Human–judge agreement. For each pair, the model before “vs” is the one with the higher judge-assigned Average score. Winning rate is computed for this first model, with ties counted as wins. Agreement measures exact sample-level agreement between the human majority preference and the judge preference.

**Table 22.** Dimension-level agreement between human majority preferences and judge preferences.

Conclusion and Future Directions

We wrap up EvolvingWorld’s contributions, limits, and ethical considerations.

EvolvingWorld offers a concrete benchmark for interactive, persistently evolving literary worlds by coupling a Character Agent with a World Model, using explicit open‑schema states and breaking long‑horizon interaction into seven trainable tasks.

Our experiments show that this design yields more coherent long‑horizon simulations across a range of model backbones, suggesting a step toward richer literary agents and controllable interactive worlds.

Compared with BookWorld, which only updates a few predefined character fields and a single global event field, EvolvingWorld supports open‑schema character updates and entity‑level world‑state changes.

We acknowledge three primary limitations: (1) a single objective world state shared by all characters, limiting subjective perception; (2) the world representation is bounded by LLM context length, tracking only salient entities; (3) the benchmark relies on public‑domain classic books, omitting modern novels, games, and user‑created worlds.

Ethically, the source works are public domain, but they may contain toxic or biased language, and LLM‑generated annotations can amplify such biases; we therefore restrict dataset use to scientific research and advise caution in downstream applications.

Simulation Prompts and Configuration

This section details the operational prompts and evaluation criteria used for story simulation and character evolution.

The simulation framework relies on a structured sequence of prompts to manage character agency and world-state consistency. These protocols govern the transition from scene planning to interaction generation, ensuring that character motivations and environmental changes remain grounded in the established narrative arc.

**Table 1.** Simulation Prompt

**Table.** Simulation Prompt

**Simulation Prompt** | Character State Update (Character Agent) | Output | | :--- | :--- | | | • **`hidden_tracker`**: Updated tracker of accumulated events/signals (experiences, emotional pressure, unresolved tensions) that may lead to future profile changes. Overwrite the old tracker with the updated version. Null if there are no signals worth tracking. | | | • **`profile_updated`**: Boolean: true only if the scene caused a meaningful, lasting change to the character's profile. Be selective; do not update for minor or transient reactions. | | | • **`updated_profile`**: The full updated profile as a plain text string when `profile_updated` is true. Include all relevant dimensions. Null if `profile_updated` is false. | | | • **`short_description`**: A brief description (50–80 words, third person) of the character as of the end of this scene: current identity, key relationships, situation, and immediate goals/intentions. | | | **Rules** | | | • Follow the step-by-step reasoning order: dimension reasoning → hidden tracker → profile decision → description. | | | • All outputs must reflect the character's state as of the END of the current scene only. | | | • Keep `hidden_tracker` private and internal-facing — it tracks accumulated signals, not public information. | | | • Update the profile only when the scene justifies a persistent, lasting character-level change. Accumulated tracker signals combined with this scene may cross the threshold. | | | • Do NOT update the profile for minor, transient reactions that don't reflect a lasting change. | | | • Return JSON only. | | | *Simulation input: Scene Scenario, Scene Interactions, Current Profile (State At Scene Start), Hidden Tracker (State At Scene Start), and Current Motivation (State At Scene Start)* |

**Shared Evaluation Prompts** | Category | Description | | :--- | :--- | | Scoring Method | Use the following scoring method: Base score: 50. First, identify Merits (excellent aspects): each merit awards 1 to 10 points. Then, identify Demerits (problematic aspects): each demerit penalizes 1 to 10 points. Final score = min(100, max(0, 50 + $\sum$(merits) - $\sum$(demerits))). Output: (1) A list of merits with their point values; (2) A list of demerits with their point values; (3) The final score (integer 0-100). | | Output Format | {"merits": [{"description": "<what was done well>", "points": <1-10>, ...}, "demerits": [{"description": "<what was problematic>", "points": <1-10>, ...], "reasoning": "<brief explanation>"} Do NOT output a "score" field — the score is computed automatically from merits and demerits. | | Scene Summary | You are an expert summarizer for interactive fiction simulations. Your task is to produce a concise summary of a single scene that captures the key events, character actions, emotional shifts, and narrative developments. This summary will be used as context for evaluating subsequent scenes and cross-scene coherence. Aim for a concise paragraph — typically a few hundred words is sufficient. Do NOT retell every interaction turn-by-turn; focus on the overall arc and key turning points. Evaluation input: Scenario, Involved Characters, Previous Scene Summary, and Interactions. | | Per-Scene Template | You are an expert evaluator for interactive fiction simulation systems. Your task is to evaluate a single scene from a simulation on ONE specific dimension: {`Dimension_Name`}. {`Dimension_Criteria`} {`Scoring_Method`} {`Evaluation_Input`} {`Output_Format`} | | Cross-Scene Template for PES | You are an expert evaluator for interactive fiction simulation systems. Your task is to evaluate a single character's evolution trajectory across multiple scenes on ONE specific dimension: {`Dimension_Name`}. Same as Above | | Cross-Scene Template for SCC | You are an expert evaluator for interactive fiction simulation systems. Your task is to evaluate the global world state evolution and scene planning coherence across the entire simulation on ONE specific dimension: {`Dimension_Name`}. Same as Above |

**Table 34.** Per-scene evaluation criteria for Character Agent metrics. Part 2 of 6.

Evaluation Metrics and Criteria

Defines per‑scene evaluation criteria for character agents and world models, and specifies format compliance checks.

This appendix enumerates the concrete per‑scene metrics used to judge character agents and world‑model updates, and it details the required output formats for interactions, profiles, and motivations.

Measures whether characters react appropriately to the immediate conversational and narrative context on a turn‑by‑turn basis.

Assesses the suitability of the motivations generated for each character within a scene.

Ensures that each interaction turn follows the prescribed markup: inner thoughts in square brackets, spoken dialogue as plain text, and visible actions in parentheses.

Checks that updated character profiles are rendered as well‑structured Markdown prose with clear section headers.

Validates that each motivation is a concise 1‑3 sentence prose passage describing the character’s inner drive for the next scene.

Evaluates whether the characters chosen for a scene are appropriate given the global world state and narrative needs.

Checks that the chosen scene location logically fits the selected characters and the narrative context.

**Table 35.** Per-scene evaluation criteria for Character Agent metrics. Part 3 of 6.

**Table 36.** Per-scene evaluation criteria for Character Agent metrics. Part 4 of 6.

**Table 37.** Per-scene evaluation criteria for Character Agent metrics. Part 5 of 6.

**Table.** Per-scene Evaluation Criteria for World Model

**Table 40.** Per-scene evaluation criteria for World Model metrics. Part 2 of 5.

**Table 41.** Per-scene evaluation criteria for World Model metrics. Part 3 of 5.

**Table 42.** Per-scene evaluation criteria for World Model metrics. Part 4 of 5.

Read the original paper

Open the simplified reader on Paperglide

Browse all simplified papers