Layer Importance for Mathematical Reasoning is Forged in Pre-Training and Invariant after Post-Training

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Nepal, Aadim, Shrestha, Safal, Shrestha, Anubhav, Kim, Minwu, Naghiyev, Jalal, Shwartz-Ziv, Ravid, Ross, Keith
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866915599479734272
author Nepal, Aadim
Shrestha, Safal
Shrestha, Anubhav
Kim, Minwu
Naghiyev, Jalal
Shwartz-Ziv, Ravid
Ross, Keith
author_facet Nepal, Aadim
Shrestha, Safal
Shrestha, Anubhav
Kim, Minwu
Naghiyev, Jalal
Shwartz-Ziv, Ravid
Ross, Keith
contents Large language models improve at math after instruction tuning, reinforcement learning, or knowledge distillation. We ask whether these gains come from major changes in the transformer layers or from smaller adjustments that keep the original structure. Using layer-wise ablation on base and trained variants, we find that math reasoning depends on a few critical layers, which stay important across all post-training methods. Removing these layers reduces math accuracy by as much as 80%, whereas factual recall tasks only show relatively smaller drops. This suggests that specialized layers for mathematical tasks form during pre-training and remain stable afterward. As measured by Normalized Mutual Information (NMI), we find that near these critical layers, tokens drift from their original syntactic clusters toward representations aligned with tokens less syntactically related but potentially more useful for downstream task.
format Preprint
id arxiv_https___arxiv_org_abs_2506_22638
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Layer Importance for Mathematical Reasoning is Forged in Pre-Training and Invariant after Post-Training
Nepal, Aadim
Shrestha, Safal
Shrestha, Anubhav
Kim, Minwu
Naghiyev, Jalal
Shwartz-Ziv, Ravid
Ross, Keith
Machine Learning
Artificial Intelligence
Large language models improve at math after instruction tuning, reinforcement learning, or knowledge distillation. We ask whether these gains come from major changes in the transformer layers or from smaller adjustments that keep the original structure. Using layer-wise ablation on base and trained variants, we find that math reasoning depends on a few critical layers, which stay important across all post-training methods. Removing these layers reduces math accuracy by as much as 80%, whereas factual recall tasks only show relatively smaller drops. This suggests that specialized layers for mathematical tasks form during pre-training and remain stable afterward. As measured by Normalized Mutual Information (NMI), we find that near these critical layers, tokens drift from their original syntactic clusters toward representations aligned with tokens less syntactically related but potentially more useful for downstream task.
title Layer Importance for Mathematical Reasoning is Forged in Pre-Training and Invariant after Post-Training
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2506.22638