DuplexGen: Adaptive Synthesis of Human-AI Turn-Taking Dialogues

2026-07-28

AI Summary

This work addresses the challenge that existing dialogue systems struggle to adapt turn-taking behaviors to human preferences across diverse conversational contexts, often due to a lack of contextually grounded training data or reliance on non-preference-driven heuristics. The authors propose a novel approach that, for the first time, integrates slot-level human preference annotations into large language model–based dialogue generation, enabling context-adaptive turn-taking through preference calibration—without requiring large-scale corpora or fixed prompting strategies. Evaluated across six cooperative and competitive task settings, the method significantly outperforms baseline systems. Furthermore, full-duplex dialogue models trained on its generated data exhibit interaction rhythms more aligned with human preferences, demonstrating both the effectiveness and generalizability of the proposed mechanism.

Abstract

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.

Problem

• turn-taking
• full-duplex interaction
• scenario-adaptive dialogue
• human preference
• dialogue synthesis

Innovation

• turn-taking
• human-AI dialogue
• preference calibration
• full-duplex interaction
• scenario-adaptive synthesis