DuplexSpeechBench-IFEval: Evaluating Implicit Instruction Following in Full-Duplex Voice Agents

Explain Like I'm Normal

Researchers have released DSB-IFEval, a new benchmark designed to measure how well voice agents handle real-time social cues like backchanneling and floor-taking without being told explicitly. Most agents currently rely on rigid turn-taking, but this framework tests their ability to infer conversational dynamics based on a specific persona. This moves us closer to AI that actually feels like a human on the phone rather than a walkie-talkie.