DuplexGen: Scenario-Adaptive Dialogue Turn-Taking via Human Preference Calibration

illinois · hf · 2026-08-11

Current full-duplex models often apply a single norm for dialogue turn-taking, ignoring context-specific variations. To address this, researchers introduced DuplexGen, a framework that calibrates LLM predictions using a small set of human preference annotations to generate dialogues with scenario-adaptive turn-taking.

Across six cooperative and competitive tasks, human turn-taking preferences varied significantly. DuplexGen aligned much more closely with these preferences than uncalibrated prompting or training solely on generic human-human data. A full-duplex model trained on DuplexGen-generated data exhibited distinctive, human-preferred behaviors. This demonstrates that human calibration, rather than just corpus scale or prompt design, is the key to scenario-specific turn-taking synthesis.

Original post →

More from Research

Research channel →