Fictional Worldbuilding: Multi-Agent LLM Collaboration with Hierarchical Context Compression and Iterative Review

The paper introduces AutoWorldBuilder, a multi-agent system for constructing coherent fictional worlds. It combines a structured concept network with conflict detection, a DAG-based hybrid batch scheduler, four-layer context compression, specialized Auditor agents, and skill-driven agent extensions. The abstract reports approximately 90% token reduction and an increase in proposal pass rates from 42% to above 85%. Across 20 diverse worldbuilding tasks using GPT-OSS 120B and DeepSeek v3.2, the system achieved a reported 95.0% success rate, producing 56–103 self-consistent concepts per world in 18–31 minutes.