Preprint
Nov 2026
Online Label-Free Gradient Matching for Continual Language Model Training
Language models in customer-facing products must be continually updated, since the data they serve shifts over time. As retraining on the full accumulated data is impractical, each update should instead run on a small set of examples carrying the training signal of all data seen so far. The strongest current methods bu...
Sajjad Pakdamansavoji, Mohammad Reza Rezaei, Ilija Subasic et al.
· 0 citations