Looped language model training cannot control hidden-state norm growth because RMSNorm normalizes scale away before the loss sees it. A paper posted today on arXiv identifies this readout blind spot, ...
Results that may be inaccessible to you are currently showing.
Hide inaccessible results