Digital Pantheon: Simulating and Auditing Coalition Formation with LLM Agents¶

🕒 Published (v1): 2026-07-16 15:08 UTC · Source: Arxiv · link

Why this paper was selected

Multi-LLM agent coalition simulation + auditing framework; practical multi-agent behavior evaluation

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

Digital Pantheon is a multi-agent LLM framework for simulating political coalition formation, combining SFT, DPO, and RAG to produce partisan agents that resist RLHF-induced neutrality bias. Applied to the 2019 Flemish election, it introduces MILT (Multi-Layered Information Lineage Topology) and a Coalition Influence Score (CIS) to make emergent negotiation auditable. Across three independent simulations, manifesto-anchored provisions reliably match the historically adopted coalition agreement while hallucinated content does not.

Problem¶

RLHF-trained LLMs are tuned for neutrality and helpfulness, making them unable to sustain steadfast partisan behavior across extended multi-agent negotiation—prompt-based personas degrade reasoning and induce conformity. Existing political simulators either predict terminal outcomes without simulating deliberation, or simulate deliberation without attributing provisions to verifiable ideological origins, leaving negotiation a black box.

Method¶

Data preparation: Party manifestos are chunked via an Agentic–Structural Hybrid (ASH) procedure combining a deterministic Hard-Gate (regex on structural markers) with a probabilistic Soft-Gate (Gemma3-27B binary same-topic judgement on trailing 800 characters + candidate block, temperature 0.1).

Preference-pair generation: Each chunk is passed to Gemma3-27B (temperature 0.7) to produce (Prompt, Chosen, Rejected) triplets: Chosen rewrites manifesto content in first-person assistant register; Rejected argues from an opposing ideology or takes a non-committal stance.

Two-stage alignment: Gemma-3 (27B) quantized to 4-bit via QLoRA on Apple MLX. Stage 1 (SFT) establishes a "Method Actor" first-person persona ($B=2, E=5$). Stage 2 (DPO) maximizes the implicit reward gap between Chosen and Rejected with KL regularization strength $\beta=0.1$ against the SFT reference policy ($B=1, E=3$, learning rate $5\times10^{-6}$), penalizing encyclopedic neutrality. Adapter steps are scaled as $I = \frac{N}{B} \times E$ to equalize alignment depth across manifestos of varying length.

RAG: Each party receives an isolated ChromaDB vector collection (paraphrase-multilingual-MiniLM-L12-v2 embeddings); at inference, top-$k$ chunks with cosine similarity $< 0.51$ ($k \leq 5$) are injected as hard ideological constraints.

Negotiation arena (Digital Pantheon): Hub-and-spoke topology over 4 rounds across 22 policy dossiers. Spoke agents are party-aligned models; the formateur hub is instantiated from the largest-party adapter but system-prompted for neutrality. The formateur is given Banzhaf power indices to enforce realistic leverage. Adapter hot-swapping keeps base weights static while party-specific DPO adapters load per turn.

Explainability (MILT): A backward NLI classifier (Qwen3.6-27B) traces every final-agreement clause ($L_3$) through agents' opening standpoints ($L_1$) to retrieved manifesto chunks ($L_0$), assigning one of five provenance states: (1) Direct Lineage, (2) Diluted Lineage, (3) Synthesized Lineage, (4) Pipeline Artifact, (5) Orphan (hallucination).

Coalition Influence Score: $$\text{CIS}(\pi) = \sum_{p:\,\pi(p)=\pi} w(\tau(p)), \quad w = \begin{cases} 3 & \tau = \text{Direct} \\ 2 & \tau = \text{Diluted} \\ 1 & \tau = \text{Synthesized} \\ 0 & \text{otherwise} \end{cases}$$

Real-world grounding ($L_\text{Real}$): Every $L_3$ provision is graded Present / Partially Present / Absent against the historically adopted 2019 Flemish coalition agreement via the same Qwen3.6 NLI pipeline; the realization index is $G = (n_\text{present} + 0.5\,n_\text{partial})/N$.

Key Contributions¶

SFT+DPO alignment pipeline that embeds party-specific partisanship at the parameter level, overcoming RLHF neutrality without relying on fragile prompt personas.
Per-party isolated RAG (separate ChromaDB collections) preventing cross-party knowledge contamination and bounding hallucinations to manifesto content.
MILT: a five-state provenance topology that traces final agreement clauses backward through negotiation layers to source manifesto chunks using NLI inference.
Coalition Influence Score (CIS): an ordinal, hallucination-discounting metric for attributing negotiating power.
Empirical external validation showing manifesto-anchored lineage predicts real-world policy materialization while ungrounded content does not.
ASH hybrid chunking strategy (Hard-Gate + Soft-Gate) that preserves document hierarchy while handling fluid thematic transitions cost-efficiently.

Results¶

Three independent simulations; $N = 2{,}629$ total classified provenance points (~876 per run), each evaluated twice to isolate evaluator noise from run variation.
Stable coalition influence ranking across all runs: N-VA $>$ CD&V $>$ Open Vld.
Manifesto-anchored MILT states (Direct, Diluted, Synthesized) account for nearly all "Present" and most "Partial" grounding judgements against the 2019 coalition agreement.
Systemic-error classes (Orphan, Pipeline Artifact) are overwhelmingly "Absent" in the real-world grounding pass.
Ordering of manifesto-anchored vs. ungrounded provisions' materialization rates is stable across all simulations and evaluation passes.

(Note: per-class quantitative breakdowns in Table 1 are in the truncated portion of the paper.)

Limitations¶

Four-round negotiation horizon is a design choice validated qualitatively in preliminary runs, not an empirically measured convergence criterion.
Evaluated on a single historical case (2019 Flemish election, three-party coalition); generalizability to other electoral systems or party counts is undemonstrated.
CIS weights (3/2/1) are an ordinal encoding, not calibrated magnitudes; any monotonically decreasing assignment over {Direct, Diluted, Synthesized} preserves the intended ordering.
The formateur is instantiated from the largest coalition party's adapter, introducing structural bias toward that party's perspective in the hub role.
$N = 3$ simulation runs is a small sample for statistical confidence in cross-dossier aggregates.
All models run locally (consumer hardware via QLoRA/MLX); scalability to larger party sets or longer negotiation horizons is uncharacterized.

Relevance to Agentic AI / LLM Agents¶

This paper directly addresses one of the core unsolved problems in multi-agent LLM systems: how to produce agents that maintain divergent, stable identities under sustained peer interaction when their base models are trained toward conformity. The SFT+DPO approach to persona persistence is more robust than prompt-only methods and generalizes to any domain requiring agents with distinct, irreconcilable objectives. The MILT framework is a concrete contribution to the open problem of process transparency in emergent multi-agent behavior—tracing causal lineage from final outputs back to grounding sources is directly applicable to auditing agentic pipelines beyond the political domain. The hub-and-spoke orchestration with a formateur as a neutral arbitrator offers a reusable protocol for structured multi-agent deliberation toward consensus.