← Back to research

Research Brief · 2026-10-13 Adjacent

DuplexAct: How Interactive Are Frontier Full-Duplex Spoken Language Models?

The idea, in brief

The Help/Safety Interrupt subset tests taking the floor with useful help or warnings. The entire full-duplex benchmark is not a proactive-assistance benchmark.

Editorial summary · not the author abstract. Read the linked paper for the complete abstract and claims.
DuplexAct interaction actions, triggering conditions and follow-through behaviors
Figure 1 · Editorial caption: Interaction actions are evaluated through their immediate execution and subsequent follow-through. Help and safety interruptions are the proactive subsets. Original figure · paper authors ↗