Duo: A Reciprocal World Model for Full-Duplex Interaction
Duo v1 is designed to keep taking in audio and video while responding, allowing users to interrupt it mid-conversation.
The model pairs “Intent,” which handles reasoning and decisions about when to participate, with “Realization,” which generates audiovisual behavior. Both modules follow a shared timeline, with attention masks blocking access to future inputs. Training uses generated reasoning tokens and synthesized audiovisual histories to reduce distribution shift. The authors describe uses beyond conversation, including reshaping virtual scenes and playing audiovisual games without a mouse or keyboard.
HF Daily Papers' note
The model pairs “Intent,” which handles reasoning and decisions about when to participate, with “Realization,” which generates audiovisual behavior. Both modules follow a shared timeline, with attention masks blocking access to future inputs. Training uses generated reasoning tokens and synthesized audiovisual histories to reduce distribution shift. The authors describe uses beyond conversation, including reshaping virtual scenes and playing audiovisual games without a mouse or keyboard.
HF Daily Papers' note
score 5