Fictional Worldbuilding: Multi-Agent LLM Collaboration with Hierarchical Context Compression and Iterative Review 虚构世界构建：具有分层上下文压缩和迭代审查的多智能体LLM协作

TL;DR

• AutoWorldBuilder introduces a multi-agent system for automated fictional worldbuilding that resolves context explosion, consistency conflicts, and lack of quality assurance.
• The architecture features a four-layer context compression mechanism reducing tokens by ~90%, a DAG-based scheduler grouping tasks by semantic locality, and specialized Auditor agents.
• Experiments using GPT-OSS 120B and DeepSeek v3.2 achieved a 95.0% success rate, generating 56-103 self-consistent concepts per world in 18-31 minutes with zero conflicts.
• The system improves proposal pass rates from 42% to over 85% through iterative review and supports zero-code extension via a skill-driven agent architecture.

Why It Matters

This research provides a scalable blueprint for managing long-horizon, knowledge-intensive generative tasks where maintaining global consistency is critical. By demonstrating effective context compression and semantic-aware scheduling, it offers practical solutions for reducing latency and token costs in complex multi-agent workflows. The separation of generation and review mechanisms serves as a robust template for improving reliability in automated content creation pipelines beyond just worldbuilding.

Technical Details

• Context Compression: A four-layer mechanism achieves approximately 90% token reduction, addressing the linear growth of context during extensive building processes.
• Task Scheduling: Utilizes a Directed Acyclic Graph (DAG)-based hybrid batch scheduler that groups tasks based on semantic locality to optimize processing efficiency.
• Quality Assurance: Implements an iterative review system with specialized "Auditor" agents, increasing the proposal pass rate from 42% to over 85%.
• Agent Architecture: Features a skill-driven design allowing zero-code extension and differentiated temperature configurations to balance creativity and consistency.
• Performance Metrics: Tested on 20 diverse tasks with backends GPT-OSS 120B and DeepSeek v3.2, resulting in a 95.0% success rate and zero-conflict delivery within 18-31 minutes per world.

Industry Insight

The "layer-as-budget" compression strategy and semantic-locality scheduling are highly transferable to other knowledge-intensive domains like legal document generation or complex software engineering assistance. Separating generation from rigorous review via specialized agents can significantly enhance the reliability of autonomous AI systems, reducing the need for human-in-the-loop correction. Developers should consider adopting modular, skill-driven agent architectures to enable easier customization and scalability without extensive code modifications.

TL;DR

• 提出AutoWorldBuilder多智能体系统，通过分层上下文压缩解决长程世界构建中的上下文爆炸问题。
• 引入基于DAG的混合批调度器和语义局部性分组，优化任务执行效率并降低冲突率。
• 采用四层上下文压缩机制实现约90%的Token减少，配合迭代审查机制将提案通过率从42%提升至85%以上。
• 在GPT-OSS 120B和DeepSeek v3.2后端验证中，20个任务达到95%成功率，生成56-103个自洽概念且零冲突。
• 架构模式（如层预算压缩、生成与审查分离）可迁移至其他知识密集型多智能体LLM应用。

为什么值得看

本文针对多智能体协作中常见的上下文溢出和一致性难题提供了具体的工程化解决方案，特别是分层压缩和迭代审查机制，对提升复杂生成任务的可靠性具有参考价值。其验证的架构模式为开发高复杂度、长周期的AI辅助创作工具提供了可复用的设计范式。

技术解析

• 核心架构：AutoWorldBuilder包含五个组件：带冲突检测的结构化概念网络、基于DAG的混合批调度器、四层上下文压缩机制、带专用Auditor智能体的迭代审查系统，以及支持零代码扩展的技能驱动智能体架构。
• 上下文压缩：采用四层上下文压缩机制，旨在解决随着构建过程线性增长的上下文爆炸问题，实验显示可实现约90%的Token减少，同时保持内容连贯性。
• 调度与优化：通过基于DAG的混合批调度器，根据语义局部性对任务进行分组，优化资源分配；技能驱动架构允许差异化温度配置，平衡创造性与一致性。
• 性能指标：在20个多样化世界构建任务中，系统能在18-31分钟内生成56-103个自洽概念，交付零冲突内容，整体成功率达95.0%。

行业启示

• 多智能体协作标准化：生成与审查分离（Separation of Generation and Review）及专用审计员角色的引入，证明了在复杂任务中引入自动化质量闭环的重要性，可作为多智能体系统的标准实践。
• 上下文管理策略：分层压缩和语义局部性调度为解决LLM长上下文限制提供了有效路径，建议在开发知识密集型应用时优先考虑此类架构以减少推理成本并提升稳定性。
• 可扩展性设计：技能驱动且支持零代码扩展的架构表明，未来的AI工具应注重模块化设计，以便快速适配不同领域的特定需求，降低部署和维护门槛。