According to arxiv.org, researchers have published two papers demonstrating multi-agent systems powered by large language models for specialized creative and technical tasks.
The first paper introduces AutoWorldBuilder, a multi-agent collaborative system designed for fictional worldbuilding in game design and literary creation. According to the research, the system addresses key challenges including context explosion, balancing creative diversity with content consistency, and quality assurance. The system features five integrated components: a structured concept network with conflict detection, a DAG-based hybrid batch scheduler, a four-layer context compression mechanism achieving approximately 90% token reduction, an iterative review system with specialized Auditor agents, and a skill-driven agent architecture. In experiments across 20 worldbuilding tasks using GPT-OSS 120B and DeepSeek v3.2, the system demonstrated a 95.0% success rate, generating 56-103 self-consistent concepts per world in 18-31 minutes with zero-conflict delivery. The research states that the Auditor agents improved proposal pass rates from 42% to over 85%.
The second paper presents QAgent, described as “the first autonomous multi-agent framework for end-to-end OpenQASM code generation” for quantum circuit programming. According to arxiv.org, QAgent integrates schema-aware task planning, code synthesis, and hardware-aware calibration. Evaluated on 12 quantum kernels across five LLMs, QAgent improved Pass@1 accuracy by 47-70% on single-kernel tasks and achieved over 88% accuracy on multi-kernel workflows for large models.