New AI Can Guess Who's Talking — Boss, Sibling, or Stranger

Researchers introduced a training-free multi-agent reasoning framework for inferring speaker relationships from spoken conversations, a task they describe as underexplored and costly to tackle with supervised modeling. The framework organizes inference through structured interaction among LLM agents, letting relationship judgments be proposed, challenged, and adjudicated without task-specific training. Two designs are tested: Multi-Role Multi-Agent Debate, which assigns agents complementary roles or social-theory-grounded perspectives, and Multi-Agent Compete, a competition-based protocol that compares judgments through pairwise adjudication, eliminates weaker candidates, and keeps the most defensible one.

Evaluated on the Seamless Interaction dataset across different modality settings, covering binary classification and fine-grained relationship-detail prediction, the methods improved over zero-shot and existing multi-agent baselines in most cases. Human evaluation suggested the task is challenging even for people: LLM methods sometimes outperformed human annotators in text-included settings but were less competitive in the audio setting. The authors conclude that relationship inference benefits from structured inference-time interaction among agents, while acoustic cues are not yet fully captured by current models.