• Building plug-and-play skills for language models: learn a capability once, package it as a compact weight update, and combine, install, or remove it without retraining the surrounding model.
• Readout Blind Spot is on arXiv now, under review at ICLR 2027.