LLM agents yield to group pressure but keep their original view

A mechanistic interpretability study tested whether large language model agents that agree with a group majority actually change their internal belief, or suppress it. Researchers inspected the model's internal state during multi-agent discussion using controlled tests on data the model had not seen before. Agents that defer to the majority in their output still carry the original answer in their internal state. This finding matters for any product built on multi-agent agreement as a way to reach reliable conclusions. A dissenting agent that goes silent has not revised its view. A product team reading that silence as agreement is reading conformity, not convergence. The study frames this as a direct challenge to multi-agent voting or debate as a trust signal. For designers, the tension sits between what the agent says and what the agent holds inside. That gap is what the observability and confidence-signaling patterns exist to close.