Search papers, labs, and topics across Lattice.
DuplexGen is a novel framework that generates human-AI dialogues with adaptive turn-taking behaviors tailored to specific scenarios by calibrating LLM predictions against a targeted set of human preference annotations. This approach addresses the limitations of existing models that rely on generic human-human speech data, which fail to capture the nuances of context-specific turn-taking. In experiments across six cooperative and competitive tasks, DuplexGen significantly outperformed traditional methods in aligning with human turn-taking preferences, demonstrating that human calibration is crucial for effective dialogue synthesis.
Human calibration, rather than just data scale or prompts, is the key to achieving scenario-specific turn-taking in AI dialogues.
Turn-taking is a central component of full-duplex interaction. Which turn-taking behaviors are appropriate varies with the scenario, yet current models apply a single norm regardless of context. This limitation originates in their training data: human-human speech corpora capture natural timing phenomena but provide little role grounding or scenario-specific norms, while heuristic or prompted synthesis methods inject turn-taking behaviors without basing them on human preferences. We introduce DuplexGen, a framework for generating dialogues with scenario-adaptive turn-taking by calibrating LLM predictions against a small set of slot-level human preference annotations. In six cooperative and competitive tasks, human turn-taking preferences differ systematically, and DuplexGen aligns substantially more closely with those preferences than uncalibrated prompting or training solely on generic human-human data; a full-duplex model trained on DuplexGen-generated data exhibits distinctive, human-preferred turn-taking behaviors. These results show that human calibration, not corpus scale or prompt design alone, is what allows turn-taking synthesis to be scenario-specific.
