Digital Pantheon: Simulating and Auditing Coalition Formation with LLM Agents

2026-07-20 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Social Sciences & Behavioral Studies · Depth: Expert, extended

Summary

The Digital Pantheon framework introduces a multi-agent system designed to simulate political coalition formation, addressing the inherent neutrality biases of Large Language Models (LLMs) trained with Reinforcement Learning from Human Feedback (RLHF). Developed by Ghent University researchers, this framework combines Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Retrieval-Augmented Generation (RAG) to instill aggressive, party-specific personas grounded in official manifestos. Operationalized on the 2019 Flemish election using Gemma3 (27B) quantized to 4-bit via QLoRA, the system employs a hub-and-spoke negotiation arbitrated by a formateur. It features a Multi-Layered Information Lineage Topology (MILT) for tracing policy clauses and a Coalition Influence Score (CIS) to quantify party contributions. Across three independent simulations, the framework consistently identified N-VA as the winner with 40.0% ± 3.3% attributable influence, followed by CD&V and Open Vld. A real-world grounding pass demonstrated that manifesto-anchored lineage reliably predicts materialization in the historical 2019 agreement, with 45.8% ± 1.5% of simulated provisions finding a real-world counterpart.

Key takeaway

For AI Scientists developing multi-agent systems for political or complex negotiations, this framework offers a robust approach to overcome LLM neutrality biases. You should integrate SFT, DPO, and RAG to create genuinely partisan agents grounded in specific policy documents. Utilize MILT and CIS to ensure transparency and quantify influence, providing auditable insights into negotiation dynamics and policy outcomes. This enables more realistic and explainable simulations.

Key insights

LLM agents can simulate partisan political coalition formation with auditable lineage and real-world grounding.

Principles

• RLHF biases prevent sustained partisan LLM behavior.
• Combining SFT, DPO, and RAG enables faithful partisan agents.
• Policy lineage tracing enhances multi-agent simulation explainability.

Method

A three-stage pipeline: RAG knowledge base, two-stage SFT+DPO ideological alignment, and a hub-and-spoke multi-agent negotiation arena.

In practice

• Use ASH chunking for structured document partitioning.
• Apply DPO to instill aggressive, party-specific personas.
• Trace policy clauses with MILT for negotiation transparency.

Topics

• LLM Agents
• Political Simulation
• Coalition Formation
• AI Alignment
• Retrieval-Augmented Generation
• Explainable AI

Code references

• chroma-core/chroma
• ml-explore/mlx
• pymupdf/PyMuPDF
• ollama/ollama
• qdrant/fastembed

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Digital Pantheon: Simulating and Auditing Coalition Formation with LLM Agents — Digital Pantheon uses SFT, DPO, and RAG to create partisan LLM agents for simulating political coalition formation and auditing outcomes.
• MARCH: Multi-Agent Reinforced Self-Check for LLM Hallucination — MARCH uses multi-agent reinforcement learning and information asymmetry to reduce LLM hallucination in RAG systems.
• Beyond Static Responses: Multi-Agent LLM Systems as a New Paradigm for Social Science Research — A six-tier framework maps LLM agent evolution from static tools to complex adaptive systems for social science research.
• Point of Order: Action-Aware LLM Persona Modeling for Data-Grounded Civic Deliberation — Speaker-attributed, action-aware data significantly enhances LLM persona modeling for realistic multi-party civic simulations.
• LLM Consortium for Software Design Refinement: A Controlled Experiment on Multi-Agent Collaboration Topologies — Structural adversarial and cross-model review topologies significantly enhance multi-agent LLM software design.
• Network Effects and Agreement Drift in LLM Debates — LLM agent opinion dynamics are shaped by network structure, group imbalance, and a systematic "agreement drift" bias.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.