Converting In-Context Learning to Weights in Linearized-Attention Transformers (arxiv.org)

is that a method to save context tokens by "baking" the attended in-context-learned into the weights? Im missing one step of so-what in the paper abstract..