Deliberation in Silico: Validating LLM Multi-Agent Simulation Against Verbatim Council Records

Christine Sheldon, Michael Laver, Slava Jankin

Keywords: Multi-agent simulation, Large language models, Political deliberation, EU Council, Validation

Abstract:

Political deliberation is constituted in language, yet computational models of politics typically abstract away from discourse. Large language models offer a new capability: agents that participate in realistic political dialogue. Before such simulations can ground causal inference about institutional effects, we must establish that they reproduce documented patterns of real deliberation. We validate LLM-based multi-agent simulation against verbatim transcripts of the EU Council of Ministers, using the 2012 Common European Sales Law orientation debate as our
case study. Across 140 simulations examining calibration, orchestration, and coalition dynamics, we find that: (1) fully calibrated agents produce nuanced discourse that resists simple position classification—a pattern validated by comparison with real ministerial statements; (2) procedural structure reliably shapes participation patterns; (3) agents reproduce the EU Council’s documented consensus culture, converging toward compromise regardless of formal voting rules; and (4) agents show implicit bloc coordination through shared argumentative emphases, though with more explicit cross-referencing than observed in real debates. These findings establish LLM-MAS as a tool for extending simulation-based
institutional analysis to language-mediated political processes.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 7