AI Trends

Ant Technology Research Institute and Westlake Xinchen Release and Open-Source Bidirectional World Model Duo v1

VR GyroscopeUpdated Nov 19 11:06

AI summary

Ant Technology Research Institute and Westlake Xinchen released and open-sourced the bidirectional world model Duo v1 on November 17, enabling virtual characters to continuously perceive input and autonomously decide whether to speak or listen. The model uses a dual-module architecture comprising Intent and Realization, makes an interaction decision approximately every 0.7 seconds, and compresses audiovisual generation to four steps. Across 180 checks in 30 real-world interactions, Duo v1 passed 168, achieving an overall pass rate of 93.3% and a content responsiveness rate of 73.3%.

Why read this

AI score 60

Duo v1 separates and coordinates interactive decision-making and audiovisual generation, providing a concrete architectural reference for enabling virtual characters to shift from passive responses to continuous perception and proactive interaction.

Topics

Tags

Body · AI translation

On November 17, Ant Technology Research Institute and Hangzhou-based large-model company Westlake Xinchen jointly released and open-sourced Duo v1, a bidirectional world model. The model introduces a new paradigm called the Reciprocal World Model (RWM), which incorporates intelligent agents into the world model, enabling virtual characters to continuously perceive user input, autonomously decide when to speak and when to listen, and generate corresponding audio and visuals for bidirectional audiovisual interaction.

Unlike conventional world models, which focus on simulating physical environments and predicting dynamics, Duo v1 seeks to shift virtual characters from passively executing instructions to actively participating in interactions. In an official demonstration, a virtual character said they were thirsty after exercising. When the user offered ice water, the character was able to take the bottle and drink from it.

Technically, Duo v1 uses a dual-module architecture comprising Intent and Realization. Intent handles interaction reasoning and decision-making, determining approximately every 0.7 seconds whether to speak or wait. Realization converts those decisions into corresponding speech, facial expressions, and video imagery. The two modules operate collaboratively along a unified timeline, supporting continuous conversations and mid-conversation interruptions.

To improve generation efficiency and continuity, the research team adopted a three-stage training scheme and used dual-branch rollout distillation to compress audiovisual generation to four steps, mitigating visual degradation and distribution shift during long generation sequences.

In terms of performance, Duo v1 passed 168 of 180 checks in a human evaluation covering 30 sets of real-world interactions, achieving an overall pass rate of 93.3%. Language instruction adherence, speech intelligibility, and visual identity stability all reached 100%, but content responsiveness was only 73.3%, indicating that its understanding of complex intents and long-term consistency still need improvement.

Submissions/News Tips: tougao@youxituoluo.com

Editorial/Business Cooperation: Lin Nan (WeChat: 19250561593)

Join the Industry Discussion Group: Lin Nan (WeChat: 19250561593)

A Digital Technology Industry Information and Services Platform

Download the “Tuoluo Technology” app for cutting-edge, in-depth news and insights on the digital technology industry

Source

VR Gyroscope