DIGITAL PANTHEON: SIMULATING AND AUDITING COALITION FORMATION WITH LLM AGENTS

Political coalition formation is traditionally studied through game-theoretic models that often abstract away the language-mediated deliberation and ideological nuances of real-world politics. The authors address the tendency of standard LLMs to default to neutral, agreeable, or conformist behavior when tasked with multi-agent negotiation. To solve this, they developed the 'Digital Pantheon' framework, which forces agents to remain faithful to their party's official 2019 Flemish election manifestos.

The methodology employs a three-stage pipeline:

1. Agentic–Structural Hybrid (ASH) Chunking: A deterministic and AI-assisted process to partition manifestos into coherent policy blocks.
2. Two-Stage Alignment: Using Supervised Fine-Tuning (SFT) to adopt a 'method actor' persona and Direct Preference Optimization (DPO) to penalize neutral or non-partisan responses.
3. RAG-Grounded Negotiation: A hub-and-spoke negotiation arena where agents retrieve specific manifesto chunks to justify their positions, preventing them from hallucinating compromises outside their party's core commitments.

The framework introduces the Multi-Layered Information Lineage Topology (MILT), which allows researchers to trace every clause in a simulated coalition agreement back to the specific manifesto chunk that inspired it. By applying a Coalition Influence Score (CIS), the authors can quantify which party drove specific policy outcomes. When tested against the historical 2019 Flemish coalition, the model produced stable rankings and agreements that closely mirrored real-world outcomes, demonstrating that manifesto-anchored lineage is a reliable predictor of policy materialization.

This research provides a scalable, transparent testbed for political scientists to explore 'what-if' scenarios in coalition formation. By moving away from black-box LLM-judge evaluations and toward a system where every negotiation outcome is auditable and grounded in verifiable ideological text, the authors offer a tool that bridges the gap between computational agent-based modeling and the qualitative realities of political deliberation.