Deterministic Continuous Replacement: Fast and Stable Module Replacement in Pretrained Transformers

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Bradbury, Rowan, Ashok, Aniket Srinivasan, Kasanagottu, Sai Ram, Jhingran, Gunmay, Meng, Shuai
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866914168579293184
author Bradbury, Rowan
Ashok, Aniket Srinivasan
Kasanagottu, Sai Ram
Jhingran, Gunmay
Meng, Shuai
author_facet Bradbury, Rowan
Ashok, Aniket Srinivasan
Kasanagottu, Sai Ram
Jhingran, Gunmay
Meng, Shuai
contents Replacing modules in pretrained models, especially swapping quadratic self-attention for efficient attention alternatives, poses a hard optimization problem: cold-start reinitialization destabilizes frozen backbones. We isolate this core stability challenge in a controlled study. Deterministic Continuous Replacement (DCR) blends teacher and student outputs with a deterministic, annealed weight. Theoretically, DCR eliminates gate-induced gradient variance inherent to stochastic replacement. In a single-seed study, DCR attains faster convergence and stronger alignment than stochastic gating and distillation baselines on controlled attention replacement, establishing a foundation for heterogeneous operator swaps.
format Preprint
id arxiv_https___arxiv_org_abs_2511_18670
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Deterministic Continuous Replacement: Fast and Stable Module Replacement in Pretrained Transformers
Bradbury, Rowan
Ashok, Aniket Srinivasan
Kasanagottu, Sai Ram
Jhingran, Gunmay
Meng, Shuai
Machine Learning
Artificial Intelligence
Computer Vision and Pattern Recognition
Replacing modules in pretrained models, especially swapping quadratic self-attention for efficient attention alternatives, poses a hard optimization problem: cold-start reinitialization destabilizes frozen backbones. We isolate this core stability challenge in a controlled study. Deterministic Continuous Replacement (DCR) blends teacher and student outputs with a deterministic, annealed weight. Theoretically, DCR eliminates gate-induced gradient variance inherent to stochastic replacement. In a single-seed study, DCR attains faster convergence and stronger alignment than stochastic gating and distillation baselines on controlled attention replacement, establishing a foundation for heterogeneous operator swaps.
title Deterministic Continuous Replacement: Fast and Stable Module Replacement in Pretrained Transformers
topic Machine Learning
Artificial Intelligence
Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2511.18670