Saved in:
Bibliographic Details
Main Authors: Cen, Zetai, Gu, Chenfei, Zhu, Jin, Li, Ting, Chen, Yunxiao, Shi, Chengchun
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.13284
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916009025208320
author Cen, Zetai
Gu, Chenfei
Zhu, Jin
Li, Ting
Chen, Yunxiao
Shi, Chengchun
author_facet Cen, Zetai
Gu, Chenfei
Zhu, Jin
Li, Ting
Chen, Yunxiao
Shi, Chengchun
contents Recent advancements in large language models demonstrate that injecting perturbations can substantially enhance extrapolation performance. However, current approaches often rely on discrete perturbations with fixed designs, which limits their flexibility. In this work, we propose a framework where token prefixes are perturbed by a learnable transformation of a continuous latent vector within an embedding space. To overcome the challenge of an intractable marginal likelihood, we derive unbiased estimating equations for model parameters and optimize them via stochastic gradient descent. We establish the statistical properties of the resulting estimator in over-parameterized regimes. Empirical evaluations on both synthetic and real-world datasets demonstrate that our proposal yields significant gains in out-of-domain settings over a range of state-of-the-art baseline methods.
format Preprint
id arxiv_https___arxiv_org_abs_2605_13284
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Learning Perturbations to Extrapolate Your LLM
Cen, Zetai
Gu, Chenfei
Zhu, Jin
Li, Ting
Chen, Yunxiao
Shi, Chengchun
Machine Learning
Statistics Theory
Recent advancements in large language models demonstrate that injecting perturbations can substantially enhance extrapolation performance. However, current approaches often rely on discrete perturbations with fixed designs, which limits their flexibility. In this work, we propose a framework where token prefixes are perturbed by a learnable transformation of a continuous latent vector within an embedding space. To overcome the challenge of an intractable marginal likelihood, we derive unbiased estimating equations for model parameters and optimize them via stochastic gradient descent. We establish the statistical properties of the resulting estimator in over-parameterized regimes. Empirical evaluations on both synthetic and real-world datasets demonstrate that our proposal yields significant gains in out-of-domain settings over a range of state-of-the-art baseline methods.
title Learning Perturbations to Extrapolate Your LLM
topic Machine Learning
Statistics Theory
url https://arxiv.org/abs/2605.13284