Introduction: The first system for constructing fiction-based multi-agent societies

Recent advances in 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 our paper titled “BookWorld: From Novels to Interactive Agent Societies for Creative Story Generation”, we introduce BookWorld, a comprehensive system for constructing and simulating book-based multi-agent societies.

The Overview of BookWorld

BookWorld provides a complete pipeline from data collection to social simulation, and finally, rephrasing into novelistic text.

Data Extraction

Before the simulation starts, we extract character and worldview data from source materials, used for book-to-system construction.

Simulation

The simulation begins by initializing role agents and the world agent, loading character profiles, the geospatial map, the worldview data and other necessary information.

Rephrasing

After the simulation ends, we collect the simulation records and apply LLMs to rephrase the records into the final, novel-style story.

While some existing work has implemented “chatroom” systems with multi-role-playing agents, they still fall short of the envisioned “virtual world.” Through the experimental process of this work, we have distilled the following insights to further advance toward a realistic virtual world:

1. Map: We present spatial relationships through a discretely represented map, supplemented with scene mode, which naturally creates narrative space.

2. Environmental Interaction: Character-to-character interaction alone easily creates a “chatroom-like” feeling. In realistic scenarios, many interactions involve the environment. We address this by implementing environmental responses.

3. Reinforcing Worldview with Setting Data: Furthermore, by extracting implicit worldview information from the source material (such as technological development progress in science fiction works, or magic systems in fantasy works) and equipping both role agents and world agents with retrieval modules, we enable their behaviors to remain synchronized with the original worldview.

4. Character Long-term and Short-term Goal Setting: Without clear purposes, characters easily fall into repetitive behavior and lack autonomy. To address this issue, our characters maintain both long-term motivations and short-term goals—the former being lifelong objectives such as defending the nation, while the latter are short-term goals formulated by characters to achieve their long-term motivations, such as reaching reconciliation with another character, which are frequently updated as the story progresses.

5. Controlling the Narrative Chain: When implementing designated plots, the world agent guides each character’s behavior at the beginning of each turn, specifically by modifying their short-term goals in conjunction with the designated plot.

If you find this work useful to your research, please kindly cite our paper:

@misc{ran2025bookworldnovelsinteractiveagent,
      title={BookWorld: From Novels to Interactive Agent Societies for Creative Story Generation}, 
      author={Yiting Ran and Xintao Wang and Tian Qiu and Jiaqing Liang and Yanghua Xiao and Deqing Yang},
      year={2025},
      eprint={2504.14538},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.14538}, 
}