source: arxiv artificial intelligence: hidden anchors in multi-agent llm deliberation
level: research
multi-agent llm deliberation lets agents exchange and revise answers over rounds to improve reasoning. this process mirrors human decision-making, where people are influenced by both the group and their own internal beliefs. classical opinion-dynamics models like degroot and friedkin-johnsen capture the herd effect but ignore personal anchors. a new model treats deliberation as a closed-loop system where each agent has a hidden internal belief, or anchor, that constantly pulls its opinion.
the anchor can be recovered from the deliberation alone. it explains behavior that classical consensus rules forbid: an agent's confidence in the correct answer can rise beyond any initial belief, escaping the convex hull formed by starting opinions. this means the group can reach a level of certainty that no single agent had at the start. the model shows how internal anchors drive this effect, offering a mechanistic understanding of why multi-agent deliberation often outperforms individual agents.
this work provides a formal framework for analyzing multi-agent llm systems. by modeling hidden anchors, researchers can better predict and control deliberation outcomes. it also connects llm behavior to social dynamics, suggesting that internal representations act like persistent beliefs. the findings could improve the design of collaborative ai systems, making them more reliable and interpretable.
why it matters: understanding hidden anchors helps build more predictable multi-agent ai systems, improving reliability in tasks like collective reasoning and decision-making.
source: arxiv artificial intelligence: hidden anchors in multi-agent llm deliberation