Long-WAM: 10-second video context cuts robot latency to 22 ms

arXiv preprint 2610.10528 presents Long-WAM, a world-action model with an autoregressive video backbone that reasons over 10 seconds of visual history while issuing actions in 22 ms. On the MRMS suite, success rate rose to 84% versus 61% for a 2-second-context baseline.

• 10-second context: 300 tokens, router keeps 64 relevant ones
• Decision latency cut from 150 ms to 22 ms
• MRMS success rate up from 61% to 84%
• Forecasts future embeddings at 0.5, 2 and 5 seconds ahead