Saved in:
Bibliographic Details
Main Authors: Liu, Yizhou, Liu, Ziming, Gore, Jeff
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2501.12243
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916575465963520
author Liu, Yizhou
Liu, Ziming
Gore, Jeff
author_facet Liu, Yizhou
Liu, Ziming
Gore, Jeff
contents Large language models (LLMs) demonstrate remarkable performance, and improving their pre-training process appears to be key to enhancing their capabilities further. Based on the documented success of Adam, learning rate decay, and weight decay, we hypothesize that the pre-training loss landscape features a narrowing valley structure. Through experiments with synthetic loss functions, we discover that when gradient query noise is high relative to the valley's sharpness, Adam's performance falls behind that of Signum because Adam reduces the effective step size too drastically. This observation led us to develop FOCUS, an optimizer that enhances Signum by incorporating attraction toward moving averaged parameters, allowing it to handle noise better while maintaining larger step sizes. In training GPT-2, FOCUS proves to be more stable than Signum and faster than Adam. These results suggest that gradient noise may be an underappreciated limiting factor in LLM training, and FOCUS offers promising solutions.
format Preprint
id arxiv_https___arxiv_org_abs_2501_12243
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle FOCUS: First Order Concentrated Updating Scheme
Liu, Yizhou
Liu, Ziming
Gore, Jeff
Machine Learning
Computation and Language
Optimization and Control
Large language models (LLMs) demonstrate remarkable performance, and improving their pre-training process appears to be key to enhancing their capabilities further. Based on the documented success of Adam, learning rate decay, and weight decay, we hypothesize that the pre-training loss landscape features a narrowing valley structure. Through experiments with synthetic loss functions, we discover that when gradient query noise is high relative to the valley's sharpness, Adam's performance falls behind that of Signum because Adam reduces the effective step size too drastically. This observation led us to develop FOCUS, an optimizer that enhances Signum by incorporating attraction toward moving averaged parameters, allowing it to handle noise better while maintaining larger step sizes. In training GPT-2, FOCUS proves to be more stable than Signum and faster than Adam. These results suggest that gradient noise may be an underappreciated limiting factor in LLM training, and FOCUS offers promising solutions.
title FOCUS: First Order Concentrated Updating Scheme
topic Machine Learning
Computation and Language
Optimization and Control
url https://arxiv.org/abs/2501.12243