Paper recorded by Signals 4 on 2026-11-02 in cs.CL. Abstract reproduced from arXiv; link to the original below.
Published 2026-11-02 on arXiv · recorded by Signals 4 on 2026-11-03
Category: cs.CL · 自然语言处理 · first seen 2026-11-03
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 build this set by gradient matching, which synthesizes examples whose training gra