Hidden Anchors in Multi-Agent LLM Deliberation
Researchers have developed a new model for multi-agent LLM deliberation, which mimics human decision-making by incorporating a hidden internal belief, or 'anchor,' for each agent. This anchor continuously influences an agent's opinion, independent of its neighbors. The study demonstrates that this anchor can be identified solely from the deliberation process and explains how an agent's confidence can surpass initial levels, moving beyond the constraints of classical consensus models. A method to test the anchor's predictive power across different model families reveals that while anchor influence is consistent, their positions vary, impacting whether deliberation escapes the initial opinion 'hull.' AI
IMPACT Provides a new framework for understanding and potentially improving multi-agent LLM reasoning by accounting for internal agent beliefs.