AI & Computer Science · Weekly AI Papers
Duo: A Reciprocal World Model for Full-Duplex Interaction (Hengyuan Cao et al.)
LISTEN TO THIS EPISODE

Hengyuan Cao and colleagues present Duo, an audiovisual system that receives input while generating responses.
Its Intent module decides how to participate, while Realization produces coordinated audio and video.
The report explains causal timing, streaming generation, and training on the system’s own generated histories.
Its evaluation covers 30 interaction cases; content responsiveness remains weaker than speech and visual fidelity.
The discussion separates this preliminary demonstration from broader claims about reliable interaction and world modeling.
Technical clarification: the 0–3 / 4–7 GPU split applies during overlapping intent inference and decoding; denoising uses all eight GPUs, and each module’s cache has its own policy. The 41.7 ms deadline concerns the shortest aligned one-frame packet, not every four-frame packet. Synchronization and interruption handling remain empirical capabilities with failure modes, rather than guarantees.
Read about the original source: Duo: A Reciprocal World Model for Full-Duplex Interaction
Part of Marginalia: books, papers, and ideas for learning and debate.