Grokking delay arises from attention performing delayed Bayesian inference over task structure after MLP memorization drives loss to zero; a KL structural intervention bypasses the inverse-weight-decay scaling.
and Kucukelbir, Alp and McAuliffe, Jon D
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Grokking as Structural Inference: Transformers Need Bayesian Lottery Tickets
Grokking delay arises from attention performing delayed Bayesian inference over task structure after MLP memorization drives loss to zero; a KL structural intervention bypasses the inverse-weight-decay scaling.