What's this about?
A study of how reliably you can inject facts into an LLM using hypernetworks — networks that generate weight-edits — and how that reliability scales with model and edit count (11↑).
What exactly is knowledge injection?
Updating what a model knows without full retraining — like editing one encyclopedia entry instead of reprinting the book. Doing it at scale without collateral damage to unrelated knowledge is the open problem.
Why is it worth noting?
If injection scales cleanly, "continual updating" becomes an engineering knob — the capability Liang Wenfeng called the missing piece this week.