Three seconds to learn: Gen1.5's few-shot skill learning may hinge on one attention masking rule

KyeGomezB · x · 2026-08-20

GEN is the first model line where one-shot and few-shot learning of physical skills appears to emerge from pretraining alone — three seconds of demonstration and it performs the task with no gradient updates. The author hypothesizes the key is a simple masking rule: an action token at time t may only attend to sensor tokens at t−δ, where δ is the inference latency sampled during training. This could explain how GEN avoids a System 1/System 2 split and inference-time guidance while hitting 100 Hz with a 7B model.

Original post →

More from Embodied

Embodied channel →