Research map / Multi-agent systems and games / LLM agents

StoryBox: Collaborative Multi-Agent Simulation for Hybrid Bottom-Up Long-Form Story Generation Using Large Language Models

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes StoryBox, a hybrid bottom-up long-form story generator in which LLM agents interact in a sandbox to produce emergent events that a Storyteller Agent turns into stories of about 12,000 words.

Problem

Long-form story generation with LLMs struggles with coherence, consistency, and natural plot/character development under top-down outlining approaches.

Approach

Agents with persona attributes (innate, learned, lifestyle, daily plans, abnormal behavior) interact in a tree-structured sandbox; the logged events are summarized, retrieved via keyword and embedding search, and chapter-by-chapter story generation is driven by these bottom-up events rather than a fixed outline.

Key results

• StoryBox leads on automatic and human pairwise comparisons across Plot, Creativity, Character Development, Language Use, Conflict Quality, and Overall.
• StoryBox produces stories averaging about 12,000 words, while other long-form methods average about 10,000 and vanilla LLM outputs about 1,000 words.
• Simulation duration study: Overall improves notably from 1 to 7 days, with diminishing returns beyond 7 days while token usage roughly doubles per duration step.

Surprising findings

• Human evaluators showed a stronger preference for IBSEN over GPT-4o and DeepSeek-V3 than automatic metrics suggested.
• Removing random abnormal behaviors caused the sharpest decline among ablated components.

Limitations

• Sequential character simulation slows the process; naive parallelization risks inconsistencies.
• Story quality evaluation remains difficult; human judgment is costly and automatic metrics need improvement.

Experimental setup

Datasets: DOC premises/settings/characters format 20 self-constructed story settings

Benchmarks: self-constructed 20-setting story dataset

Metrics: pairwise human preference (Plot, Creativity, Character Devel LLM-judged pairwise preference Character Behavior Consistency Average Word Count

Baselines: GPT-4o DeepSeek-V3 Re^3 DOC-V2 IBSEN

Builds on: Generative Agents DOC IBSEN Agents' Room

Closest papers by meaning

• Agents' Room: Narrative Generation through Multi-step Collaboration (ICLR 2025 · similarity 0.89)
• NarrativePlay: An Automated System for Crafting Visual Worlds in Novels for Role-Playing (AAAI 2024 · similarity 0.85)
• SceneDecorator: Towards Scene-Oriented Story Generation with Scene Planning and Scene Consistency (NeurIPS 2025 · similarity 0.84)
• TALES: Text-Adventure Learning Environment Suite (NeurIPS 2026 · similarity 0.84)
• Human Simulacra: Benchmarking the Personification of Large Language Models (ICLR 2025 · similarity 0.83)
• AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors (ICLR 2024 · similarity 0.83)