The Deliberative Illusion: Diagnosing Factual Attrition and Stance Homogenization in Multi-Agent LLM Deliberation

Abstract

In multi-agent large language model (LLM) systems, achieving consensus is frequently mistaken for a sign of productive interaction. However, for deliberative tasks, true reliability hinges on the agents' ability to retain the specific facts and perspectives necessary to fully understand a complex issue. This study identifies a phenomenon termed the "deliberative illusion," wherein discussion leads to two detrimental outcomes: (1) factual attrition, defined as the gradual disappearance of critical information regarding the issue, and (2) stance homogenization, characterized by the convergence of varied viewpoints into a uniform consensus.

To quantify these effects, we developed DelibTrace, a novel framework that breaks down each issue into its constituent atomic facts. This system identifies which facts are essential to the issue, distributes them among agents, and monitors their persistence throughout discussion rounds. Our experiments, conducted across ethical and news-based deliberation scenarios using three distinct LLM families, reveal that multi-agent discussion can eliminate up to 72% of issue-critical facts.

This erosion of information carries significant consequences. The remaining evidence may allow for a distorted reconstruction of the issue, final decisions often remain tied to the base model’s initial priors, and a single adversarial agent can introduce misinformation into the diminishing shared context. These findings highlight a critical risk: as agents reach greater agreement, their collective knowledge base may shrink. We advocate for evaluation metrics that specifically assess which facts, uncertainties, and valid disagreements endure through the interaction process.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC