S1's ICL advantage grows exponentially in long-horizon, out-of-distribution tasks

ZeYanjie · x · 2026-08-31

Deepak Pathak of SkildAI highlights that S1's in-context learning (ICL) excels most in super long-horizon tasks (10+ mins) and out-of-distribution scenarios. While simple tasks are solved by zero-shot prompts, ICL becomes critical for unseen scenarios, with S1's advantage over regular VLAs growing exponentially as tasks move further from the pretraining distribution.

Original post →

More from Embodied

Embodied channel →