LLM-negotiated joint programmes of government

Presented by Slava Jankin.

After an election, parties with different manifesto commitments must negotiate a common programme of government -- a governing mandate that no party offered voters on its own. We develop a computational model of this process in which LLM agents represent political parties and negotiate substantial policy texts in natural language. Parties choose prospective partners, draft and amend a joint programme, and seek cabinet approval and parliamentary support, with a specified policy outcome if bargaining fails.

Across 900 simulations of Germany, Slovenia and Denmark, we examine routes to agreement, the programmes produced and the distribution of policy accommodation among parties. Agreement frequently requires amendments or follows unsuccessful formation attempts, with substantial differences across countries. Negotiation records reveal concrete policy compromises, while comparisons with observed cabinets and programmes identify both strengths and limits of the model.

The approach makes it possible to study bargaining over rich policy texts under interpretive uncertainty: parties must assess commitments whose meaning and acceptability are not given by a shared numerical representation. We also discuss extension experiments examining information provision, procedural reminders and direct drafting, using them to clarify which findings depend on the negotiation process.