Relationship between activation rank and generalization

Problem description

Test loss and Gated MLP activation rankme seems to have a correlation

<img width="2545" height="1265" alt="Image" src="https://github.com/user-attachments/assets/f95788d5-9a7e-4579-9140-bdd0a5007dbd" />

How to reproduce
1. open comfyResearch: https://online.comfy-research.com/
2. Use Lottery mode and Type in 'him' in the input field
3. draw the random graph and run