ConMeZO: Adaptive Descent-Direction Sampling for Gradient-Free Finetuning of Large Language Models

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Behric, Lejs Deen, Zhang, Liang, Li, Bingcong, Thekumparampil, Kiran Koshy
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866915945260253184
author Behric, Lejs Deen
Zhang, Liang
Li, Bingcong
Thekumparampil, Kiran Koshy
author_facet Behric, Lejs Deen
Zhang, Liang
Li, Bingcong
Thekumparampil, Kiran Koshy
contents Zeroth-order or derivative-free optimization (MeZO) is an attractive strategy for finetuning large language models (LLMs) because it eliminates the memory overhead of backpropagation. However, it converges slowly due to the inherent curse of dimensionality when searching for descent directions in the high-dimensional parameter space of billion-scale LLMs. We propose ConMeZO, a novel zeroth-order optimizer that accelerates convergence by adaptive directional sampling. Instead of drawing the direction uniformly at random, ConMeZO restricts the sampling to a cone centered around a momentum estimate. This concentrates the search in directions where the true gradient is more likely to lie and thus reduces the effect of high dimensions. We prove that ConMeZO achieves the same worst-case convergence rate as MeZO. Empirically, when finetuning LLMs on natural language tasks, ConMeZO is up to 2X faster than MeZO while retaining the low-memory footprint of zeroth-order methods.
format Preprint
id arxiv_https___arxiv_org_abs_2511_02757
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle ConMeZO: Adaptive Descent-Direction Sampling for Gradient-Free Finetuning of Large Language Models
Behric, Lejs Deen
Zhang, Liang
Li, Bingcong
Thekumparampil, Kiran Koshy
Machine Learning
Optimization and Control
Zeroth-order or derivative-free optimization (MeZO) is an attractive strategy for finetuning large language models (LLMs) because it eliminates the memory overhead of backpropagation. However, it converges slowly due to the inherent curse of dimensionality when searching for descent directions in the high-dimensional parameter space of billion-scale LLMs. We propose ConMeZO, a novel zeroth-order optimizer that accelerates convergence by adaptive directional sampling. Instead of drawing the direction uniformly at random, ConMeZO restricts the sampling to a cone centered around a momentum estimate. This concentrates the search in directions where the true gradient is more likely to lie and thus reduces the effect of high dimensions. We prove that ConMeZO achieves the same worst-case convergence rate as MeZO. Empirically, when finetuning LLMs on natural language tasks, ConMeZO is up to 2X faster than MeZO while retaining the low-memory footprint of zeroth-order methods.
title ConMeZO: Adaptive Descent-Direction Sampling for Gradient-Free Finetuning of Large Language Models
topic Machine Learning
Optimization and Control
url https://arxiv.org/abs/2511.02757