Multi-agent deliberation paper probes hidden anchoring effects in LLM groups

A new arXiv paper titled “Hidden Anchors in Multi-Agent LLM Deliberation” examines how anchor information can shape the behavior of LLM groups. The topic is important because many AI systems now use multiple agents or model calls to debate, critique, or refine answers.

If hidden anchors bias the group, multi-agent deliberation may produce confidence without genuine independence. That would weaken one of the main arguments for using model ensembles or debate-style workflows in high-stakes reasoning.

The paper adds to a growing research thread: agentic systems need evaluation methods that test coordination failures and social dynamics, not just single-model accuracy.