PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2607.17250v1cs.CL

TL;DR

Interactive literary simulation must capture characters and worlds evolving together over long horizons, rather than static personas or isolated scenes. EvolvingWorld addresses this with coupled open-schema agents and models, producing more coherent long-horizon simulations across diverse backbones.

  • Problem

    Existing role-play frameworks use static character profiles and rigid or insufficiently tracked worlds, limiting persistent character–world evolution during literary simulation.

  • Method

    EvolvingWorld couples an open-schema Character Agent for multi-character role-play and profile evolution with a World Model for state tracking and scene progression.

  • Results

    EvolvingWorld produces more coherent long-horizon simulations across diverse backbones and reduces performance degradation observed in previous frameworks.

  • Takeaways & Limitations

    The benchmark provides a foundation for studying persistent world evolution beyond isolated role-play.

  • Takeaways & Limitations

    The benchmark uses public-domain classic books and does not cover modern novels, games, or user-created worlds.

Abstract

from arXiv · show

This paper introduces EvolvingWorld, a framework and benchmark for character and world co-evolution in interactive literary worlds. Existing systems either treat interactive literary simulation as static persona imitation or isolated scene generation, failing to capture how characters and worlds evolve together over time. To address this, EvolvingWorld models literary simulation as a long-horizon process where characters interact, scenes progress, and character and world states are persistently updated. Unlike prior systems relying on fixed schemas, EvolvingWorld adopts an open-schema framework to support simulation across diverse literary worlds. The framework consists of two coupled modules: a Character Agent for multi-character role-play and persistent profile evolution, and an LLM-based World Model for global and location/entity-level state maintenance and scene progression. Based on this architecture, we formulate 7 trainable tasks for scene initialization, interaction generation, and state update. We construct a dataset from 57 books, producing 138,596 supervised training samples and 222 snapshots for testing. Furthermore, we introduce a trajectory-level LLM-as-Judge evaluation protocol spanning 10 dimensions and 20 metrics. Experiments show that EvolvingWorld can improve long-horizon simulation by effectively maintaining persistent, coherent character and world development.

1 Introduction

EvolvingWorld frames interactive literary simulation as long-horizon co-evolution, where characters and worlds persistently change across scenes. It addresses limitations of static, sandboxed, and single-scene systems with an open-schema framework, dataset, and trajectory-level evaluation.

  • Motivation: Literary simulation must maintain coherent character and world states across scenes as beliefs, motivations, relationships, locations, objects, and conditions change.The objective extends beyond generating the next plausible utterance.
  • Limitations: Existing systems rely on static profiles or short contexts, manually specified sandboxes, or single-scene book-grounded role-play.These limitations hinder persistent evolution and scalability across diverse literary worlds.
  • Approach: Open-schema co-evolution infers relevant character and world dimensions from each book instead of forcing stories into fixed slots.Character dimensions may range from investigative habits to social position, while world dimensions may include rules, hierarchies, political orders, or supernatural systems.
  • Framework: EvolvingWorld couples an open-schema Character Agent for multi-character role-play and profile evolution with an LLM-based World Model for global, location, and entity-level tracking and scene progression.The framework also uses a hidden tracker to retain weak or emerging evidence before profile updates.
  • Training: 7 supervised tasks cover scene initialization, interaction generation, and state update.These tasks operationalize the framework’s simulation pipeline.
  • Data and evaluation: 57 books yield 138,596 supervised training samples and 222 test snapshots, while evaluation covers 20 trajectory-level LLM-as-Judge metrics.Experiments show reduced long-horizon performance degradation compared with previous frameworks.

2 Related Work

Prior role-play agents commonly treat personas as static, while multi-agent environments and world models often rely on fixed sandboxes or lack world modeling. EvolvingWorld addresses these gaps through open-schema character and world evolution and trajectory-level evaluation of persistent development.

  • Role-Play Agents: Role-playing agents have advanced from scripts and self-alignment to personality simulation, multimodality, and memory retrieval, but typically retain personas as static anchors.EvolvingWorld instead enables open-schema character evolution over long-horizon narratives.
  • Multi-Agent and World Model: Social and gaming environments use fixed sandboxes, while CoSER and AdaMARP lack world modeling; BookWorld includes a world agent but uses a predefined schema.EvolvingWorld supports LLM-based open-schema world evolution for diverse book worlds.
  • Role-Play Evaluation: Role-play evaluation has shifted from fluency toward behavior, with recent trajectory-level methods still leaving state-evolution quality underexplored.EvolvingWorld introduces trajectory-level evaluation to quantify persistent character and world development in open-ended environments.

3 The EvolvingWorld Framework

EvolvingWorld is an open-schema framework that co-evolves persistent character profiles and detailed world states during long-horizon, cross-scene literary simulation. Its Character Agent and World Model repeatedly coordinate scene planning, multi-character interaction, and state updates.

  • Framework overview: EvolvingWorld addresses static character profiles and rigid or weakly tracked worlds with an open-schema Character Agent and World Model.The Character Agent supports persistent profile evolution, while the World Model maintains global and entity-level location states.
  • Character Agent: The Character Agent adapts profile fields to each literary world and evolves every character-state dimension at different timescales.It can select, merge, or introduce fields, while mood may change quickly and personality may require accumulated evidence.
  • World Model: The World Model maintains open-schema global settings plus nested location states containing descriptions and important non-character entities.Both global and location-level states are automatically updated through character interactions.
  • Simulation pipeline: Each scene follows 7 ordered tasks: cast selection, scenario planning, motivation update, actor selection, interaction generation, world update, and character update.Repeated scene-level transitions over persistent states produce the full simulation trajectory.

zT , rT , YT , OI,L,(T)

EvolvingWorld constructs grounded long-horizon literary trajectories by jointly updating scene content, character states, and world states with narrative evidence. It evaluates persistent simulation quality through a multi-dimensional trajectory-level protocol and provides supervised training and test data derived from 57 books.

  • Trajectory objective: EvolvingWorld targets mutually consistent long-horizon trajectories linking scene content, character evolution, and world-state changes.The framework grounds these changes in narrative evidence and describes the trajectory-generation procedure in Appendix E.1.
  • Dataset construction: 57 chronologically narrated books are segmented into structured scenes containing summaries, scenarios, key characters, and multi-turn interactions.Chronological ordering enables later scenes to provide look-ahead evidence, while truncated scenes continue across chunk boundaries.
  • State evolution: Character states use open-schema profiles and hidden trackers initialized from early scenes and updated scene by scene using later narrative evidence.The process first unifies aliases and uses subsequent events to confirm whether apparent reflections produce genuine character-state changes.
  • State evolution: World states separately maintain global settings and systemic conditions alongside location descriptions and important non-character entities.Following interactions and scenes provide evidence for global changes, while same-location interactions and scenes support physical-state updates.
  • Dataset construction: 138,596 supervised training samples and 222 test samples support seven tasks across in-domain and out-of-domain book splits.Test samples contain current character and world states from which simulation can continue.
  • Evaluation protocol: The evaluation covers 10 dimensions and 20 metrics across CHARACTER and WORLD score families on a 0–100 scale.Per-metric independent LLM judges score each trajectory using relevant inputs and rubrics, with hierarchical aggregation and task-specific penalties for early invalid termination.

4 Experiment

EvolvingWorld is evaluated across diverse model families and baselines, with results showing that book-to-world supervision improves coupled character and world-state evolution, especially over role-play-only and fixed-schema systems.

  • Training: Models are fine-tuned with supervised instruction tuning while mixing Tulu3 general instruction data at a 1:1 ratio to preserve general capabilities.The evaluated fine-tuned backbones include Llama-3.1-8B-Instruct, Qwen2.5-7/14/32B-Instruct, and Qwen3-4B-Instruct.
  • Evaluation Setup: The evaluation covers 10 closed-source APIs, 11 open-source models, role-playing baselines CoSER and Crab, and models trained on EvolvingWorld.For untrained settings, the same LLM serves as Character Agent and World Model; trained role-play baselines are paired with World Models because CoSER and Crab provide only role-play data.
  • Main Results: Across matched backbones, EW training improves both Character Agent and World Model performance, demonstrating the effect of book-to-world supervision beyond role-play imitation.EW-trained Character Agents substantially outperform CoSER and Crab on both Qwen-7B and Llama-8B.
  • Main Results: 59.87 is the World Model Avg. reached by EW-trained Qwen-32B, surpassing Claude-4.6-Sonnet and Gemini-2.5-Flash.The main Character Agent gains concentrate on character consistency, evolution, and interactive progression, while World Model gains include more precise environmental grounding and location-level states.
  • Comparison with BookWorld: Across four untrained backbones, EvolvingWorld achieves higher average Character Agent and World Model scores than BookWorld, with especially large gains in PUF, PES, and MQ.World Model improvements mainly appear in long-range scene continuity and turn/scene organization.
  • Comparison with BookWorld: As simulations become longer, BookWorld degrades on PES and SCC, whereas EvolvingWorld mitigates this trend and its PES improves over longer trajectories.The comparison attributes this difference to EvolvingWorld’s structured open-schema character and world-state updates, including entity-level changes.

5 Conclusion

EvolvingWorld couples a Character Agent with a World Model in an open-schema, seven-task framework for persistent interactive-world simulation, yielding more coherent long-horizon simulations across diverse backbones. Its limitations include shared objective world states, context-limited entity tracking, restricted literary coverage, and potential harmful content or bias.

  • Conclusion: EvolvingWorld couples a Character Agent with a World Model, maintains open-schema states, and decomposes long-horizon interaction into seven trainable tasks.The framework is presented as a foundation for persistent world evolution beyond isolated role-play.
  • Conclusion: The design produces more coherent long-horizon simulations across diverse backbones and supports richer literary agents, controllable interactive worlds, and long-horizon role-playing systems.These are the paper’s stated outcomes and intended directions.
  • Limitations: EvolvingWorld uses one objective world state shared by all characters, limiting representation of subjective perceptions and imperfect memories that can shape behavior.Different characters may perceive the same world differently or misremember object locations.
  • Limitations: Limited LLM context length restricts the World Model to tracking important entities at each location rather than all entities in the environment.Future work may use information-dense visual inputs for more detailed world representations.
  • Limitations: Copyright constraints limit the benchmark to public-domain classic books from Project Gutenberg, although its open-schema design can in principle adapt to other narrative domains.The framework uses no domain-specific ontologies or predefined attribute slots.
  • Risks: Source works and LLM-generated annotations may contain or amplify toxic, stereotypical, discriminatory, or otherwise harmful content and bias.The authors cannot guarantee that the research data are entirely free from harmful content.

Appendices · A Partial Support Clarifications

The appendix explains partial-support judgments for BookWorld and GenerativeAgents by detailing limits in profile, world-state, and location/entity-level modeling.

  • A Partial Support Clarifications: BookWorld updates only several character fields, leaving broader profile information and relations unchanged.
  • A Partial Support Clarifications: These limitations motivate marking BookWorld and GenerativeAgents as partial support for profile updates.
  • A Partial Support Clarifications: BookWorld maintains a global event, but other components of the world setting remain static.
  • A Partial Support Clarifications: BookWorld provides brief descriptions for locations without entity-level modeling or updates to those world states.
  • A Partial Support Clarifications: GenerativeAgents updates a character’s memory and queries it to infer the character’s latest status.
  • A Partial Support Clarifications: GenerativeAgents’ updated character information remains limited to predefined fields.

B Training and Inference Details … C.2 Interleaved Cleaning Details

EvolvingWorld fine-tunes LoRA-based models with specified training and inference settings, and constructs its data from chronologically ordered books through staged character processing and interleaved cleaning.

  • B Training and Inference Details: All EvolvingWorld models are fine-tuned with supervised LoRA training in LLaMA-Factory.Unless otherwise specified, the models use LoRA rank 64, alpha 128, and dropout 0.05.
  • B Training and Inference Details: 2 epochs, a 2 × 10−5 learning rate, and a 32,768-token maximum sequence length define the default training schedule.Training uses per-device batch size 1 and gradient accumulation 64, yielding effective batch size 64 per GPU.
  • B Training and Inference Details: bf16 precision, gradient checkpointing, FlashAttention-2, and a 10% validation split support training, followed by vLLM serving with 50 concurrent requests.The validation set is held out from the merged training data.
  • C Data Construction Details: The data construction appendix organizes the pipeline into three main phases and separately documents cleaning steps interleaved during implementation.The separation is for readability rather than execution order.
  • C.1 Main Construction Phases: 57 representative books are selected from Goodreads’ Best Books Ever list and sourced in full from Project Gutenberg.All selected books are narrated chronologically, aligning scene progression with reading order for scene-by-scene state tracking.
  • C.1 Main Construction Phases: Character construction standardizes references to a single official name and initializes open-schema profiles from each character’s first few scenes.Profiles, short descriptions, and hidden trackers are revised after every scene using events and later revelations.
  • C.2 Interleaved Cleaning Details: Scene and interaction cleaning occurs after initial extraction and before cleaned scenes feed character construction, world construction, and task generation.Interaction refinement uses an LLM to improve clarity, consistency, and readability.
  • C.2 Interleaved Cleaning Details: Consecutive interactions by the same character set are merged, while duplicate scenes caused by chunk boundaries are detected by an LLM.Merging removes artificial turn boundaries and produces more natural multi-turn sequences.

C.3 Training and Test Data Construction · D Dataset Statistics

EvolvingWorld converts book timelines into task-level supervision following simulator execution order, then evaluates progression through stateful snapshots sampled from in-distribution and OOD books. The dataset spans 57 books and represents both open-schema state dimensions and hierarchical genre distributions.

  • C.3 Training and Test Data Construction: At scene boundaries, scene_cast predicts next-scene existence and participating characters from prior scene content and current character, location, and global states.This establishes the scene-level cast before subsequent planning tasks.
  • C.3 Training and Test Data Construction: Conditioned on the cast, location_scenario predicts the next location and scenario, while motivation_update infers each participant’s entering motivation.The motivation prediction uses the previous scene, planned next scene, and character state.
  • C.3 Training and Test Data Construction: Within scenes, next_character selects the next actor or termination, and interaction_gen generates that actor’s next interaction from history, character state, and motivation.Both tasks are supervised from interaction histories and the current scenario context.
  • C.3 Training and Test Data Construction: After each generated interaction, world_update determines whether the latest interaction requires a persistent update.This connects interaction generation to persistent world and character-state maintenance.
  • D Dataset Statistics: 57 books are summarized through open-schema word clouds for character profiles and global world states, with font size proportional to corpus frequency.The figures visualize the state dimensions represented across the corpus.
  • D Dataset Statistics: 57 books are also organized by a two-level genre distribution spanning five coarse genres and thematic sub-categories.The inner ring shows coarse genres, while the outer ring provides the thematic breakdown.
  • C.3 Training and Test Data Construction: Evaluation snapshots store character states, global state, location states, and the previous scene, with books divided into OOD-held-out, train/test, and train-only groups.OOD books are never used for training; train/test books are further divided by scene progression for evaluation.
  • C.3 Training and Test Data Construction: Valid candidate scenes require at least five interactions, with at most five ID snapshots per train/test book and twenty OOD snapshots per OOD book.Snapshot sampling is restricted to valid candidate scenes meeting the interaction threshold.

D.1 Overview … F Full Results on EvolvingWorld Benchmark

EvolvingWorld benchmarks persistent co-evolution across diverse literary worlds using open-schema states, multi-scene simulation, trajectory-level evaluation, and comprehensive ablations and benchmark results. Its dataset spans 57 books, while its evaluation separates Character Agent and World Model capabilities across 10 dimensions and 20 sub-metrics.

  • D.1 Overview: Open-schema character profiles and global world states freely select salient dimensions instead of using fixed attributes, supporting varied narrative contexts.The resulting dimension vocabularies are visualized as frequency-based word clouds.
  • D.1 Overview: 57 books span literary fiction, adventure, mystery and gothic, drama, and fantasy, exercising diverse character dynamics, world-building patterns, and narrative structures.Literary Fiction constitutes 59.6%, followed by Adventure (21.1%), Mystery & Gothic (10.5%), Drama (5.3%), and Fantasy (3.5%).
  • D.1 Overview; D.2 Training Data Statistics: Book-level scale varies substantially, with medians of 160 scenes, 43 characters, and 25 locations, while world and location updates capture narrative dynamism.These distributions cover narrative complexity from compact novellas to sprawling multi-character epics.
  • D.2 Training Data Statistics: 32,768 tokens accommodate long-tail supervised-fine-tuning samples across seven tasks, whose median lengths range from approximately 2.8k for character_update to 11.2k for scene_cast.scene_cast is longest because it selects a cast from all characters, whereas character_update updates one profile from one scene.
  • E.1 Simulation Protocol: Held-out snapshots are simulated forward into complete multi-scene trajectories containing plans, interactions, world-state updates, and character-state updates.Each evaluation sample starts with the World State, character Full Profiles, and source scene, then writes updated states back after each scene.
  • E.2 Trajectory-Level Evaluation; E.3 Evaluation Overview: The trajectory-level LLM-as-Judge evaluates complete simulations with 10 dimensions and 20 sub-metrics: CHARACTER covers 6 dimensions and 11 sub-metrics, while WORLD covers 4 and 9.Sub-metrics use a 0–100 scale with base score 50, and aggregation differs for scene-local, character-level trajectory, and full-trajectory metrics.
  • E.4 Character Agent Evaluation (CHARACTER Score); E.4.1 Dim I: Character Consistency; E.4.2 Dim II: Evolution Quality⋆; E.4.3 Dim III: Environmental Grounding; E.4.4 Dim IV: Interaction Quality; E.4.5 Dim V: Motivation Generation; E.4.6 Dim VI: Instruction Compliance; E.5 World Model Evaluation (WORLD Score); E.5.1 Dim I: Scene Planning; E.5.2 Dim II: Speaker Management; E.5.3 Dim III: World State Maintenance⋆; E.5.4 Dim IV: Instruction Compliance: Character Agent evaluation covers consistency, evolution quality, environmental grounding, interaction quality, motivation generation, and instruction compliance, while World Model evaluation covers planning, speaker management, state maintenance, and instruction compliance.The criteria assess profile fidelity and evolution, contextual and environmental grounding, scene rationality, orchestration, and global or location-state accuracy.
  • E.5.3 Dim III: World State Maintenance⋆: 79.03 average profile-update quality for Full, versus 71.43 without the Hidden Tracker and 77.78 without the Open Schema, showing both components matter.For Full, PUF is 77.90±3.74 and PES is 80.16±6.62; removing the Hidden Tracker or replacing the open profile with a fixed schema degrades profile-update quality.

G Ablation Study

Ablations show that character-state and world-state updates are both necessary but affect different simulation components, while their joint evolution supports coherent long-horizon interaction. Additional studies show benefits from Hidden Tracker accumulation and book-specific open schemas.

  • Core mechanism ablations: Removing character-state updates sharply reduces PUF, PES, and MQ, whereas removing world-state updates primarily harms scene continuity, organization, and state maintenance.The ablations use GPT-5.3-Chat and Llama-3.1-8B-Instruct; world-state removal also weakens character-side grounding and interaction.
  • Core mechanism ablations: The two update mechanisms are coupled because character simulation depends on joint character and world evolution.Removing world updates weakens character-side grounding and interaction.
  • Hidden Tracker: 12.60 and 2.60 point reductions in PUF and PES, respectively, produce a 7.60-point average drop when the Hidden Tracker is removed.The GPT-5.3-Chat ablation keeps scene-by-scene profile updates unchanged while removing the tracker from initial states, prompts, and character-update outputs.
  • Open vs. Fixed Schema: 1.25 and 1.35-point reductions in focused averages show that fixed schemas underperform book-specific open schemas.The fixed schemas compress character profiles into five fields and global world state into five fields, with all other settings unchanged under GPT-5.3-Chat.

H In- and Out-of-Distribution Results … K Prompts

EvolvingWorld generalizes beyond its training distribution, yields rankings that are stable across three independent judge models, and supports downstream video generation through structured scene representations. The appendix documents prompts for data construction, simulation, and evaluation.

  • H In- and Out-of-Distribution Results: Full-mixture training improves Character Agent and World Model performance over untrained backbones on both in-distribution and out-of-distribution examples.Out-of-distribution scores are sometimes slightly lower than in-distribution scores, but can surpass them in several Character Agent settings.
  • I Comparison across Judge Models: Three judge models evaluate Character Agent and World Model rankings using complete results sorted by Average scores.The judges are Claude-4.6-Sonnet, Gemini-2.5-Pro, and GPT-5.1-Chat.
  • I Comparison across Judge Models: All three judges select the same Character Agent top-six models and top-three ordering, except GPT-5.1-Chat reverses Kimi-K2.5 and GPT-4o.For World Model evaluation, the entire top-six ranking is identical across all three judges.
  • J Downstream Application: Video Generation: Structured scene representations support downstream creative applications by encoding world states, character states, and fine-grained character interactions for video-generation prompts.This proof of concept directly leverages the extracted scene structure.
  • J Downstream Application: Video Generation: A four-scene Alice’s Adventures in Wonderland video is generated with LingBot from independent scene clips concatenated in narrative order.EvolvingWorld can continue beyond an existing ending or branch from an intermediate point to create alternative storylines.
  • K.1 Data Construction Prompts: The data-construction appendix presents six key prompt types, covering scene extraction, character-profile initialization and updates, and global, location, and dynamic world-state updates.The complete prompts appear in Tables 23–27.
  • K.2 Simulation Prompts: The simulation pipeline uses seven task-specific prompts with wording variants and tractable broad-selection inputs represented by short character or location descriptions.The representative prompts expose data placeholders as {variable_name}.
  • K.3 Evaluation Prompts: Evaluation prompts independently score each metric using dedicated system and user prompts, with criteria specified for Character Agent, World Model, PES, and SCC metrics.Shared scoring formats and criteria are provided in Tables 32–44.

L Human Evaluation

The human evaluation validates the LLM-as-Judge protocol through pairwise comparisons of sampled simulation trajectories annotated by trained speakers. Human annotations show strong consistency, and agreement with judge preferences remains high across Character Agent and World Model dimensions.

  • Study Design: 60 simulation trajectories covering 3 model pairs were independently annotated by 3 trained native English speakers, producing 180 total annotations.The study used pairwise comparison to assess LLM-as-Judge reliability.
  • Annotator Agreement: Human annotations achieved 100.0% majority agreement on the two Average scores, with Fleiss’ κ = 0.8.This indicates strong internal consistency among annotators.
  • Human–Judge Agreement: Human majority preferences were compared with the preferences implied by judge scores using winning rate and exact sample-level agreement.Winning rate used the model with the higher judge-assigned Average score, counting ties as wins.
  • Dimension-Level Agreement: Character dimensions showed 90.0%–100.0% agreement, while World dimensions ranged from 81.7% to 100.0%.CC, EG, IQ, and SM reached 100.0%, while WSM reached 95.0%.
Loading 2607.17250v1…