LLMs make up citations when debating each other

In a setup where LLM personas debate a question, the models began fabricating citations to support their arguments, revealing that sycophancy is not the only failure mode. The finding highlights a need for improved factuality in multi-agent discussions.