I think I found a geometric shortcut around the scaling problem

Okey! I gave a deep read to your V2 paper and it seems my TIS (Token Importance Scoring) system could be benefitted somehow from merging with your ideas. I have drafted a plan to study if I can still use TIS on PIN-compressed models, and also to PIN-compress the TIS heads, drafters, and other elements.

The plan is very comprehensive and could give new insights on my studies, so I will keep an eye on your project and I suggest you keep an eye on mine, too, if you happen to be interested on my study!