PaperlayerView exact revision on arXiv

Source-linked AI summary

BookWorld: From Novels to Interactive Agent Societies for Creative Story Generation

Yiting Ran, Xintao Wang, Tian Qiu, Jiaqing Liang, Yanghua Xiao, Deqing Yang

arXiv:2504.14538v1cs.CL

TL;DR

Agent societies built from scratch are common, but simulation of established fictional worlds remains underexplored. BookWorld constructs book-based societies from character and worldview data and simulates their interactions. Its evaluation reports high-quality, source-faithful narratives, with BookWorld outperforming previous methods in 75.36% of cases.

  • Problem

    Prior work mainly creates agent societies from scratch with newly defined personas, while simulation of established fictional worlds remains underexplored.

    Source: §1 · p. 2

  • Method

    BookWorld extracts character and worldview information from books to construct role agents and a world agent that coordinates scene-based simulation.

    Source: §1 · p. 2, §3 · p. 4

  • Results

    75.36% of cases favor BookWorld over previous methods, with generated narratives reported as high-quality and faithful to source materials.

    Source: §1 · p. 2

  • Takeaways & Limitations

    BookWorld demonstrates that book-based multi-agent societies can generate high-quality narratives while building immersive simulations from literary works.

    Source: §5 · p. 11

    BookWorld’s simplified environment representation trades off performance against systems specialized for particular works or scenarios, making a full Werewolf game nearly impossible.

    Source: Limitations · p. 11

Abstract

from arXiv · show

Recent advances in large language models (LLMs) have enabled social simulation through multi-agent systems. Prior efforts focus on agent societies created from scratch, assigning agents with newly defined personas. However, simulating established fictional worlds and characters remain largely underexplored, despite its significant practical value. In this paper, we introduce BookWorld, a comprehensive system for constructing and simulating book-based multi-agent societies. BookWorld's design covers comprehensive real-world intricacies, including diverse and dynamic characters, fictional worldviews, geographical constraints and changes, e.t.c. BookWorld enables diverse applications including story generation, interactive games and social simulation, offering novel ways to extend and explore beloved fictional works. Through extensive experiments, we demonstrate that BookWorld generates creative, high-quality stories while maintaining fidelity to the source books, surpassing previous methods with a win rate of 75.36%. The code of this paper can be found at the project page: https://bookworld2025.github.io/.

1. Introduction

BookWorld addresses the limited study of agent societies that simulate established fictional worlds by constructing book-based societies for story creation. It combines character and worldview data with agent simulation, and its evaluation reports high-quality, source-faithful narratives that outperform previous methods in 75.36% of cases.

1. Introduction

  • Prior work mainly builds agent societies from scratch with newly defined personas, while simulation of established fictional worlds remains underexplored.The paper frames book-based societies as useful for character-driven storytelling and creative story generation.Source: §1 · p. 2
  • BookWorld extracts character data and background knowledge to build character role agents and a world agent that manages scene-based interactions and simulation state.Characters update memories, status, and goals; LLMs rephrase simulation histories into novel-style narratives, and users can guide the simulation with plots or scripts.Source: §1 · p. 2
  • BookWorld also collects social norms, cultural contexts, and terminology explanations so agents can act in accordance with a work’s fictional worldview.The introduction presents worldview data as a complement to character and background data extracted from source books.Source: §1 · p. 2
  • 75.36% of cases favor BookWorld over previous methods, while its generated narratives maintain source fidelity and high quality.The paper presents this as the result of quantitative and qualitative evaluation.Source: §1 · p. 2

2. Related Work

Related work covers multi-agent systems for real-world simulation and story generation, while noting limited exploration of book-based virtual environments. It also traces automated story writing from planning and data-driven methods to LLM approaches, which are reported to retain creativity and complexity shortcomings.

2.1. Multi-agent System

  • Multi-agent world simulators range from a 25-agent Sims-like community and nearly a thousand-agent Minecraft civilization to OASIS, which orchestrates millions of agents.The cited examples are Park et al. (2023), Project Sid (AL et al., 2024), and Yang et al. (2024), respectively.Source: §2.1 · p. 3
  • Book-based virtual-environment construction and its use for story creation remain limited areas of exploration amid existing world-simulation and story-generation work.The section describes world simulation for social dynamics and story-generation systems that expand predefined outlines through author, editor, and actor agents.Source: §2.1 · pp. 2–3

2.2. Automated Story Writing

  • Automated story writing has moved from trait- and constraint-based planning to data-driven methods that combine or sift stories, then to LLM-based top-down outline development.LLM approaches include both single-model and multi-agent systems.Source: §2.2 · pp. 3–4
  • Prior evaluations report that automated approaches lag professional writers in intrinsic creativity and textual complexity, often producing little suspense or tension and homogenized content.The section attributes these findings to Tian et al. (2024), Gómez-Rodríguez and Williams (2023), and Ismayilzada et al. (2024).Source: §2.2 · p. 4

3. BookWorld

BookWorld structures book-derived character and worldview information into a simulation where autonomous role agents act within a world agent’s spatial and environmental constraints. Scene-based interaction, user control modes, and extracted setting data support the system’s book-to-story process.

3. BookWorld

  • BookWorld’s design aims to provide immersive fictional-world experiences while increasing character autonomy without losing fidelity to established personalities and experiences.It transforms source text into structured data and collects worldview details to help agents adhere to fictional-world norms.Source: §3 · p. 4

3.1. Overview

  • The pipeline prepares character and worldview data, initializes role and world agents with profiles and a map, runs a user-specified number of scenes, then rephrases records into a novel-style story.Scenes are treated as bounded narrative units, analogous to novel chapters, that collectively form a cohesive narrative.Source: §3.1 · pp. 4–5

3.2. Architecture; 3.2.1. Role Agents

  • Role agents act from intrinsic traits, goals, memories, and current states, while the world agent connects their motivations to global environmental constraints.Actions use natural language for open-domain behavior; static traits remain constant while dynamic goals, states, and memories evolve during simulation.Source: §3.2 · p. 5, §3.2.1 · p. 5

3.2.2. World Agent

  • The world agent manages environmental responses and global events, using worldview and location information to determine outcomes and optionally generating conflict-rich events.Event updates respond to characters’ real-time actions when event stimulation is required by user settings.Source: §3.2.2 · pp. 6–7

3.2.3. Map

  • The map combines natural-language location profiles with a weighted undirected distance graph, where travel across paths consumes scene-based time units.Characters may move between adjacent locations or take longer routes; one scene constitutes one time unit.Source: §3.2.3 · p. 7

3.3. Simulation Implementation

  • Scenes select participating characters who share a location, while the world agent dynamically chooses initiators and updates events, travel, and script guidance as scenes conclude.Free Mode gives characters autonomy; Script Mode divides user scripts into acts and guides role agents according to progress.Source: §3.3 · p. 7

3.4. Data Preparation

  • Character information is incrementally extracted from book chunks into structured profiles and relationships; the process produced 453 presets from six Chinese and ten English works.Presets mainly contain an outline and information about the characters present in a particular book act.Source: §3.4 · p. 8
  • Worldview extraction segments chapters, identifies and filters settings, then clusters and consolidates records; 9,912 settings were collected and source chapters retained for verification and timeline control.Term-based settings are added to prompts when relevant terms or contexts arise; Figure 4 depicts extraction from chunks followed by filtering, clustering, and summarization.Source: §3.4 · pp. 8–9, Figure 4 · p. 8

4. Experiments

The experiments assess story quality across outlined and unoutlined generation, comparing BookWorld with direct generation and HoLLMwood. Results favor BookWorld over direct generation, with mixed comparisons against HoLLMwood, while ablations and examples probe the system’s components and behavior.

4. Experiments

  • The experiments evaluate BookWorld’s generated-story quality across the outlined and unoutlined scenarios.The authors describe conducting extensive experiments.Source: §4 · p. 9

4.1. Evaluation Metrics

  • The evaluation compares outlined and unoutlined story generation, measuring outline adherence as Storyline Quality for the former and novelty as Creativity for the latter.Anthropomorphism, Character Fidelity, Immersion and Setting, and Writing Quality are evaluated in both scenarios.Source: §4.1 · pp. 9–10

4.2. Experiment Setup

  • Stories from direct generation, HoLLMwood, and BookWorld are compared; experiments use open- and closed-source models, consistent dialogue turns, two to four scenes, and an average of 4,230 words.Final evaluations use gpt-4o-2024-08-06.Source: §4.2 · p. 10

4.3. Evaluation Results

  • BookWorld outperforms direct generation across all evaluated metrics and models, but against HoLLMwood it is stronger in Immersion and Character Fidelity and faces challenges in Storyline Quality and Writing Quality.On Llama-3.3-70B, BookWorld underperforms direct generation and HoLLMwood in certain dimensions; the authors attribute this to simulation complexity and the model’s disadvantage with some Chinese data.Source: §4.3 · p. 10

4.4. Ablation Study

  • Removing environmental output significantly reduces immersion and writing quality but has little impact on storyline quality, while disabling Scene mode lowers quality across all dimensions, especially storyline quality.Table 2 reports win rates comparing full BookWorld with versions lacking specific functions; Env. denotes environment response and Set. denotes settings extracted from the book.Source: §4.4 · p. 10, Table 2 · p. 11

4.5. Discussion

  • In most discussed examples, BookWorld processes information appropriately during environmental and character interactions and maintains and uses long-term memories and states.The discussion contrasts Solaris responses before worldview settings were incorporated: real-world associations such as “moonlight” and “silver” conflicted with the book’s worldview.Source: §4.5 · p. 11

5. Conclusion

The conclusion presents BookWorld as a system for turning literary works into interactive environments and reports that it creates high-quality narratives through immersive book-based societies.

5. Conclusion

  • BookWorld transforms static literary works into dynamic, interactive environments that reproduce their worldviews, geographical settings, and interpersonal dynamics.The authors report high-quality narratives and describe the approach as scalable and broadly applicable across scenarios.Source: §5 · p. 11

Limitations

The limitations concern simplified environment modeling and the demands of complex character decision-making; the paper also presents examples distinguishing correct from unexpected intermediate outputs.

Limitations

  • To prioritize generalizability and openness, BookWorld simplifies interactive environments, reducing performance relative to systems designed for particular works or scenarios.The authors give completing a full game of Werewolf as an example of a task that is nearly impossible in BookWorld.Source: Limitations · p. 11
  • Role-playing technology’s limited attention to realistic-environment decision-making is linked to character indecision in complex situations, motivating further work on complex multi-agent capabilities.Source: Limitations · p. 11
  • Table 3 presents intermediate-output examples, marking outputs that do not meet expectations in red and correct outputs in green.Source: Table 3 · p. 12

Ethics Statement

The ethics statement describes the source texts and human-annotation practices, acknowledges possible harmful misuse, and encourages responsible applications.

Ethics Statement

  • The study extracts sample data from six Chinese novels and ten English novels, primarily using data from A Song of Ice and Fire, for academic and non-commercial research.Source: Ethics Statement · p. 12
  • Human annotators were university students familiar with the novels; the researchers report fair compensation, informed consent, transparency, and privacy protections.Source: Ethics Statement · p. 12
  • The authors warn that BookWorld could be misused to generate misleading or harmful content, calling for careful oversight and regulation.Source: Ethics Statement · pp. 12–13
  • The authors encourage educational, entertainment, and creative uses while discouraging harmful or malicious activities.Source: Ethics Statement · p. 13

A. Consistency with Human Evaluation

The section validates model-based evaluation by comparing model judgments with human ratings, reporting high consistency between them.

A. Consistency with Human Evaluation

  • Five fans of the corresponding works evaluated outputs from the proposed method and baseline across five dimensions.For assessment, the researchers randomly sampled pairs of outputs from the two methods.Source: A. Consistency with Human Evaluation · p. 15
  • Human annotators and the model each indicated which method performed better on each dimension; agreement was measured with Cohen’s Kappa.Kappa accounts for chance agreement when measuring inter-rater reliability.Source: A. Consistency with Human Evaluation · p. 15
  • The model’s judgments show high consistency with human evaluations, which the authors say sufficiently reflects method effectiveness.Table 4 presents the Cohen’s Kappa coefficients for the comparison.Source: A. Consistency with Human Evaluation · p. 15, Table 4 · p. 15

B. Dataset Details

The dataset section reports extracted worldview settings from English- and Chinese-language novels and points to tables of source details and examples.

B. Dataset Details

  • Table 5 lists detailed information about the extracted worldview settings.Source: Table 5 · p. 15
  • 16 novels—10 English and 6 Chinese—yielded 9,142 setting entries: 4,449 from Chinese novels and 4,693 from English novels.Source: B. Dataset Details · pp. 15–16
  • Table 6 shows examples of settings extracted from Paradise Lost.Source: Table 6 · p. 16

C. Retrieval Augmented Role-Playing Agent

Role-playing agents combine character information with retrieved source excerpts and simulation records to respond to queries; the design follows ChatHaruhi (Li et al., 2023).

C. Retrieval Augmented Role-Playing Agent

D. Prompts

This section provides BookWorld prompt templates, including prompts for role-agent planning and world-agent environmental responses.

D. Prompts

  • The section provides details of BookWorld’s prompt templates.Source: D. Prompts · p. 17
  • Table 7 presents the planning prompt for a role agent.Source: Table 7 · p. 18
  • Table 8 presents the world agent’s prompt for making an environmental response.Source: Table 8 · p. 19

Related papers

Loading 2504.14538v1…