DuplexGen: Adaptive Synthesis of Human-AI Turn-Taking Dialogues
DuplexGen uses a small layer of human preference calibration to make synthetic full-duplex dialogue timing fit the task.
The paper says turn-taking norms shift across cooperative and competitive scenarios, while current models tend to apply one timing pattern everywhere. DuplexGen calibrates LLM-generated dialogue slots against human annotations, then uses that to synthesize scenario-specific turn-taking data. In the reported six-task evaluation, it matched human preferences better than uncalibrated prompting or generic human-human speech data alone. HF Daily Papers' note
The paper says turn-taking norms shift across cooperative and competitive scenarios, while current models tend to apply one timing pattern everywhere. DuplexGen calibrates LLM-generated dialogue slots against human annotations, then uses that to synthesize scenario-specific turn-taking data. In the reported six-task evaluation, it matched human preferences better than uncalibrated prompting or generic human-human speech data alone. HF Daily Papers' note
score 4