Yeah, but it's likely you can RLVR on this capability. Extrapolate from there.
But why would you train on this capability versus, say, optimizing RL training runs?
But why would you train on this capability versus, say, optimizing RL training runs?