Hidden Anchors in Multi-Agent LLM Deliberation

The paper introduces a model for multi-agent LLM deliberation, conceptualizing it as a closed-loop dynamical system where each agent possesses a hidden internal belief, termed an anchor, that influences its opinion independently of others. The study demonstrates that this anchor can be identified through deliberation processes and allows agents to exceed their initial confidence levels, thus escaping conventional consensus limits. This finding is significant for practitioners as it offers insights into enhancing reasoning and accuracy in LLMs by understanding the dynamics of individual beliefs within collaborative environments.