New Model Reveals "Hidden Anchors" Influence Multi-Agent LLM Deliberation

This paper explores the dynamics of multi-agent Large Language Model deliberation, a process where multiple LLMs exchange and refine their responses over several rounds to improve accuracy. The authors propose a novel model that incorporates the concept of "hidden internal beliefs," or anchors, within each agent. These anchors are theorized to exert a constant pull on an agent's opinion, similar to how individual convictions influence human decision-making, even when interacting with group opinions. The research demonstrates that these hidden anchors can be identified solely from the deliberation process itself. Crucially, the presence of these anchors helps explain an observed phenomenon where an agent's confidence in a correct answer can exceed the initial confidence levels of any individual agent, effectively "escaping" the bounds of their starting beliefs. This behavior is not accounted for by traditional consensus models like DeGroot or Friedkin-Johnsen, which primarily focus on group influence. The study also introduces a method to test if a model is genuinely driven by such anchors by checking if the recovered anchors can predict future deliberation outcomes. Findings across several open-source LLM families suggest that while the influence of anchors is consistently strong, their specific location relative to initial opinions determines whether deliberation can break free from the initial belief space, necessitating this more comprehensive closed-loop model.