DuplexGen uses human calibration to adapt AI turn-taking per scenario

A new framework called DuplexGen, developed by researchers at multiple institutions including Cornell and Seoul National University, solves a key limitation of current conversational AI: turn-taking that ignores context. Today's models apply a single norm regardless of whether you're collaborating on a task or competing in a game. The paper, currently under review on arXiv, demonstrates that human calibration—not just more data or clever prompts—is what enables truly scenario-aware turn-taking.

DuplexGen works by taking a base LLM's dialogue predictions and adjusting them using a compact set of human preference annotations at the 'slot' level (each possible speaking turn). In six different cooperative and competitive tasks, the framework produced dialogues that closely matched human turn-taking preferences. When a full-duplex model was trained on DuplexGen-generated data, it exhibited distinctive behaviors that human judges preferred over those trained on generic human-human conversation corpora or uncalibrated LLM outputs. This work underscores that as AI voice interfaces move toward full-duplex (both parties speaking and listening simultaneously), adapting interaction style to the specific scenario is critical for naturalness and effectiveness.