Saved in:
Bibliographic Details
Main Authors: von Arx, Tobias, Dieudonné, Tanguy
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.22981
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916037853708288
author von Arx, Tobias
Dieudonné, Tanguy
author_facet von Arx, Tobias
Dieudonné, Tanguy
contents Fill-in-the-middle (FIM) is a pretraining objective widely used to equip causal language models with infilling ability, yet its effect on verbatim memorization remains underexplored. We study the memorization dynamics of FIM in a controlled setting by pretraining matched Llama 3.2 models with FIM and standard left-to-right (LTR) objectives on a FineWeb-Gutenberg corpus containing repeated Gutenberg excerpts. With prefix-based probes, FIM more often recovers short or partially matching spans, while LTR more often assigns high confidence to long exact continuations. We observe that verbatim extraction under FIM-training grows approximately linearly with repetitions over the tested range. Evaluating native FIM-format probes reveals that suffix context is not sufficient: verbatim recall under FIM-training remains strongly anchored in prefix context. Our results also show that evaluating only one span length or probing format can miss important nuances in memorization behavior.
format Preprint
id arxiv_https___arxiv_org_abs_2605_22981
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Memorization Dynamics of Fill-in-the-Middle Pretraining
von Arx, Tobias
Dieudonné, Tanguy
Computation and Language
Artificial Intelligence
Machine Learning
Fill-in-the-middle (FIM) is a pretraining objective widely used to equip causal language models with infilling ability, yet its effect on verbatim memorization remains underexplored. We study the memorization dynamics of FIM in a controlled setting by pretraining matched Llama 3.2 models with FIM and standard left-to-right (LTR) objectives on a FineWeb-Gutenberg corpus containing repeated Gutenberg excerpts. With prefix-based probes, FIM more often recovers short or partially matching spans, while LTR more often assigns high confidence to long exact continuations. We observe that verbatim extraction under FIM-training grows approximately linearly with repetitions over the tested range. Evaluating native FIM-format probes reveals that suffix context is not sufficient: verbatim recall under FIM-training remains strongly anchored in prefix context. Our results also show that evaluating only one span length or probing format can miss important nuances in memorization behavior.
title Memorization Dynamics of Fill-in-the-Middle Pretraining
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2605.22981