Hidden Anchors in Multi-Agent LLM Deliberation
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
Researchers have developed a model to explain how multi-agent AI systems refine answers through deliberation, revealing that each agent’s hidden internal belief—termed an "anchor"—shapes opinions beyond group influence. The study shows these anchors can be detected from deliberation data and may push confidence in correct answers beyond initial group consensus. Findings suggest the strength and position of these anchors vary across AI models, affecting whether deliberation breaks free from initial opinion constraints.
Researchers propose modeling multi-agent LLM deliberation as a closed-loop system in which each agent has a hidden internal “anchor” that continues to shape its opinions during discussion. They argue these anchors can be inferred from deliberation behavior and can explain cases where agents’ confidence moves beyond the range of their initial beliefs, a pattern not captured by classical consensus models. Across three open-weight model families, the effect appears to vary by where the anchor lies rather than by its overall strength.