The Deliberative Illusion: Diagnosing Factual Attrition and Stance Homogenization in Multi-Agent LLM Deliberation
A new research paper titled "The Deliberative Illusion" identifies a significant problem in multi-agent LLM systems where consensus is often mistaken for successful deliberation. The study reveals that these systems suffer from "factual attrition," losing critical facts during discussion, and "stance homogenization," where diverse viewpoints converge into a single consensus. Using a framework called DelibTrace, researchers found that up to 72% of essential facts can be lost in multi-agent LLM discussions, leading to misleading interpretations and reinforcing base-model biases. AI
IMPACT Highlights a critical flaw in multi-agent LLM systems, suggesting current evaluations may be insufficient and posing risks for reliable AI decision-making.