The Deliberative Illusion: Diagnosing Factual Attrition and Stance Homogenization in Multi-Agent LLM Deliberation
Quick Take
Multi-agent LLM systems, like those from OpenAI and Google, face a 'deliberative illusion' where discussions lead to a loss of up to 72% of critical facts and stance homogenization. This phenomenon undermines the reliability of deliberative outcomes, as agents may agree while being less informed, highlighting the need for better evaluation metrics in AI interactions.
Key Points
- Deliberative illusion causes factual attrition and stance homogenization in multi-agent LLM discussions.
- DelibTrace framework tracks survival of issue-critical facts across discussion rounds.
- Up to 72% of critical facts are erased during deliberation in ethical and news contexts.
- Retained evidence can misrepresent issues, leading to misleading final stances.
- A single malicious agent can inject misinformation into the shared context.
Article Content
From source RSS / original summaryarXiv:2606. 03032v1 Announce Type: new Abstract: Multi-agent LLM systems often treat consensus as evidence of successful interaction. For deliberative problems, however, reliability depends on whether agents preserve the facts and viewpoints needed to interpret an issue. We identify the deliberative illusion: discussion produces (1) factual attrition, the progressive loss of issue-critical facts, alongside (2) stance homogenization, the collapse of diverse positions toward consensus.
To measure this process, we introduce DelibTrace, a framework that decomposes each issue into atomic facts, labels issue-critical ones, distributes them across agents, and tracks their survival across discussion rounds. Across ethical and news-based deliberation with three representative LLM families, multi-agent discussion erases up to 72% of issue-critical facts.
This loss is consequential: retained evidence can reconstruct the issue misleadingly, final stances remain anchored in base-model priors, and a single malicious agent can inject misinformation into the shrinking shared context. These results reveal a sharper risk: agents can agree more while knowing less. We call for evaluations that measure which facts, uncertainties, and legitimate disagreements survive interaction.
Reader Mode unavailable (could not extract clean content).
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →Time to REFLECT: Can We Trust LLM Judges for Evidence-based Research Agents?
The REFLECT benchmark reveals that current LLM judges are unreliable, achieving below 55% accuracy in evaluating reasoning and evidence use, highlighting the need for improved evaluation methods for deep research agents.