FIM pretraining yields linear growth of verbatim extraction with data repetitions and stronger prefix dependence for recall than left-to-right training in matched Llama models.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Memorization Dynamics of Fill-in-the-Middle Pretraining
FIM pretraining yields linear growth of verbatim extraction with data repetitions and stronger prefix dependence for recall than left-to-right training in matched Llama models.