AGTBRIEF
A new paper tests two collaboration patterns, debate and pairwise adjudication, for inferring how two speakers are related. The methods beat single agent baselines in text and trail humans on audio only inputs.
When you hear two people talk, can you tell whether they are a parent and child, a boss and employee, friends, or strangers? A new arXiv preprint argues that teams of AI agents can make that guess reasonably well, as long as the conversation is available in text.
The paper introduces two training-free collaboration patterns. In "Multi-Role Multi-Agent Debate," agents are assigned complementary social roles, like different social-theory perspectives, and argue with each other. In "Multi-Agent Compete," answers are paired off and weaker candidates are eliminated in a tournament, leaving the most defensible guess. Both patterns outperformed single-agent and prior multi-agent baselines on the Seamless Interaction dataset, evaluated on both binary classification and detailed relationship prediction.
The methods sometimes beat human annotators, but only when text is available. In the audio-only setting, the same methods were less competitive, a limit the authors flag rather than dismiss. The cheap lever here is changing how agents interact at inference, not retraining.