Conversation Is a Two-Body Problem: Dyadic Evaluation of Full-Duplex Dialogue Models
DyaFDB tests full-duplex voice models by making two of them converse and grading both sides afterward.
The paper argues that fixed or pre-recorded interlocutors miss the joint behavior that defines spoken dialogue, including turn-taking, overlap, and interruptions. Its setup assigns roles and cooperative or conflicting goals to paired models, then uses an external judge to score each participant. The authors instantiate four tasks across 140 scenarios and 7,560 recorded model-to-model conversations. They report that each model’s behavior continually reshapes its partner’s behavior, so neither side can be treated as a passive test fixture. ArXiv · AI/CL/LG's note
The paper argues that fixed or pre-recorded interlocutors miss the joint behavior that defines spoken dialogue, including turn-taking, overlap, and interruptions. Its setup assigns roles and cooperative or conflicting goals to paired models, then uses an external judge to score each participant. The authors instantiate four tasks across 140 scenarios and 7,560 recorded model-to-model conversations. They report that each model’s behavior continually reshapes its partner’s behavior, so neither side can be treated as a passive test fixture. ArXiv · AI/CL/LG's note
score 5