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

Analysis

Why This Paper Matters

Turn-taking is a fundamental aspect of full-duplex human-AI interaction, yet current models apply a single norm regardless of context. This paper identifies a critical gap: existing training data either captures natural timing but lacks scenario-specific norms, or injects behaviors without human grounding. DuplexGen addresses this by introducing a calibration framework that aligns LLM predictions with human preferences, enabling scenario-adaptive turn-taking.

The significance lies in its demonstration that human calibration, not just data scale or prompt design, is the key to generating contextually appropriate turn-taking. This challenges the prevailing assumption that more data or better prompts alone suffice, and provides a practical method for injecting human preferences into dialogue generation.

Technical Contributions

• DuplexGen Framework: A novel approach that calibrates LLM predictions using slot-level human preference annotations, allowing for fine-grained control over turn-taking behaviors.
• Preference Annotation: Collects a small set of human preference annotations across six cooperative and competitive tasks, revealing systematic differences in turn-taking norms.
• Comparison with Baselines: Evaluates against uncalibrated prompting and training on generic human-human data, showing DuplexGen's superior alignment with human preferences.
• Full-Duplex Model Training: Demonstrates that a full-duplex model trained on DuplexGen-generated data exhibits distinctive, human-preferred turn-taking behaviors, validating the framework's utility.

Results

The paper reports that human turn-taking preferences differ systematically across the six tasks, and DuplexGen aligns substantially more closely with those preferences than uncalibrated prompting or training solely on generic human-human data. While specific metrics are not provided in the abstract, the qualitative and comparative results indicate a significant improvement. The trained full-duplex model shows distinctive behaviors that match human preferences, suggesting that the generated data effectively transfers turn-taking norms.

Significance

This work has broader implications for the field of human-AI interaction, particularly for full-duplex systems like voice assistants and robots. By emphasizing human calibration, it provides a pathway to create more natural and context-aware interactions. The framework could be extended to other aspects of dialogue beyond turn-taking, such as politeness or emotional tone, and may influence how future models are trained with human feedback. The findings also contribute to the ongoing discussion on the role of human preferences in AI alignment, reinforcing the need for human-in-the-loop approaches in dialogue generation.