Digital Pantheon: Simulating and Auditing Coalition Formation with LLM Agents

arXiv cs.CL · 6 hours ago

Researchers developed a multi-agent framework using large language models to simulate political coalition formation, tested on the 2019 Flemish election with partisan agents aligned to party manifestos through fine-tuning and retrieval techniques. The framework deployed three independent simulations that produced stable rankings with N-VA emerging as the strongest coalition partner, and introduced traceable provenance methods to validate which proposals originated from party manifestos versus model hallucinations. The approach enables computational political science to predict real-world coalition agreements while maintaining interpretability by grounding simulated negotiations in actual party policy documents.