Researchers model multi-agent LLM deliberation as a dynamical system where each agent maintains a hidden internal belief (anchor) that influences its opinions across discussion rounds. The study reveals that agents can escape the convex hull of initial beliefs through deliberation, a behavior unexplained by classical consensus models, and demonstrates that these anchors can be recovered and validated across open-weight model families.
This research addresses a critical gap in understanding how multi-agent language model systems reach improved reasoning through iterative deliberation. Traditional opinion-dynamics models like DeGroot and Friedkin-Johnsen capture social consensus effects but fail to explain why individual agents can develop confidence in correct answers beyond their starting positions. By introducing the concept of hidden anchors—persistent internal beliefs that continuously influence each agent's contributions—the researchers provide a more complete mathematical framework for modeling LLM collaboration. The work bridges artificial and human decision-making by recognizing that both involve tension between external social influence and internal conviction. The ability to recover anchors from deliberation traces alone offers practical validation: when recovered anchors predict held-out experimental runs, the model is genuinely capturing the underlying mechanism rather than fitting noise. Testing across three open-weight model families reveals important heterogeneity—not all models rely equally on anchors, and deliberation escapes the convex hull only when anchors sit sufficiently far from initial opinions. This spectrum of behavior suggests that different model architectures or training approaches produce varying degrees of opinion persistence. For developers building multi-agent AI systems, these findings clarify when deliberation genuinely improves reasoning versus when it merely reweights initial beliefs. Understanding anchor strength could guide prompt engineering, agent diversity requirements, and performance expectations in collaborative reasoning tasks.
- →Multi-agent LLM deliberation can be modeled as closed-loop dynamics driven by hidden internal beliefs (anchors) that persist across discussion rounds.
- →Agents can escape the convex hull of initial beliefs through deliberation, behavior classical consensus models cannot explain.
- →Recovered anchors can be validated by testing their predictive power on held-out deliberation runs across model families.
- →Anchor influence varies across model families in both strength and positioning relative to initial opinions.
- →Deliberation only escapes the convex hull when anchors are positioned far from initial collective beliefs.