Trying to Make a VLA Its Own Reward Model

We tried replacing SRPO's 1.1B-parameter V-JEPA with the VLA's own SigLIP encoder. Here's what we learned.