@inproceedings{kong-etal-2026-gnmr,
title = "{GNMR}: Runtime Stability Control for Low-Precision Large Language Model Training",
author = "Kong, Boao and
Jia, Weichen and
Zhang, Engao and
Li, Guohong and
Dong, Yonghan and
Wang, Yao and
Wang, Yaoyuan and
Peng, Yunke and
Yuan, Kun",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1810/",
pages = "38665--38689",
ISBN = "979-8-89176-462-0",
abstract = "Training stability is a key bottleneck in low-precision language model training: efficient low-cost paths can still produce short-lived numerical risks at a small set of operators. We formulate this as runtime stability control and present **G**radient **N**orm-to-**M**ean **R**atio (**GNMR**), a lightweight controller that compares each recoverable unit{'}s current gradient norm with its historical mean. Together with $\Delta$-GNMR for abrupt short-window increases, GNMR maps local risk signals to bounded recovery actions under a hard $\text{maxO}$ budget and a short lock interval, without changing the numerical format, kernel, or backend recipe. Across activation-quantization stress, DeepSeek-style recipe-level training, and LLaMA-2 13B fine-tuning, GNMR preserves high-fidelity quality with sparse, budgeted recovery. These results support GNMR as a backend-agnostic controller to improve low-precision training stability while preserving low-cost execution."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kong-etal-2026-gnmr">
<titleInfo>
<title>GNMR: Runtime Stability Control for Low-Precision Large Language Model Training</title>
</titleInfo>
<name type="personal">
<namePart type="given">Boao</namePart>
<namePart type="family">Kong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Weichen</namePart>
<namePart type="family">Jia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Engao</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Guohong</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yonghan</namePart>
<namePart type="family">Dong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yao</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yaoyuan</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yunke</namePart>
<namePart type="family">Peng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kun</namePart>
<namePart type="family">Yuan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Training stability is a key bottleneck in low-precision language model training: efficient low-cost paths can still produce short-lived numerical risks at a small set of operators. We formulate this as runtime stability control and present **G**radient **N**orm-to-**M**ean **R**atio (**GNMR**), a lightweight controller that compares each recoverable unit’s current gradient norm with its historical mean. Together with Δ-GNMR for abrupt short-window increases, GNMR maps local risk signals to bounded recovery actions under a hard \textmaxO budget and a short lock interval, without changing the numerical format, kernel, or backend recipe. Across activation-quantization stress, DeepSeek-style recipe-level training, and LLaMA-2 13B fine-tuning, GNMR preserves high-fidelity quality with sparse, budgeted recovery. These results support GNMR as a backend-agnostic controller to improve low-precision training stability while preserving low-cost execution.</abstract>
<identifier type="citekey">kong-etal-2026-gnmr</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1810/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>38665</start>
<end>38689</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T GNMR: Runtime Stability Control for Low-Precision Large Language Model Training
%A Kong, Boao
%A Jia, Weichen
%A Zhang, Engao
%A Li, Guohong
%A Dong, Yonghan
%A Wang, Yao
%A Wang, Yaoyuan
%A Peng, Yunke
%A Yuan, Kun
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F kong-etal-2026-gnmr
%X Training stability is a key bottleneck in low-precision language model training: efficient low-cost paths can still produce short-lived numerical risks at a small set of operators. We formulate this as runtime stability control and present **G**radient **N**orm-to-**M**ean **R**atio (**GNMR**), a lightweight controller that compares each recoverable unit’s current gradient norm with its historical mean. Together with Δ-GNMR for abrupt short-window increases, GNMR maps local risk signals to bounded recovery actions under a hard \textmaxO budget and a short lock interval, without changing the numerical format, kernel, or backend recipe. Across activation-quantization stress, DeepSeek-style recipe-level training, and LLaMA-2 13B fine-tuning, GNMR preserves high-fidelity quality with sparse, budgeted recovery. These results support GNMR as a backend-agnostic controller to improve low-precision training stability while preserving low-cost execution.
%U https://aclanthology.org/2026.findings-emnlp.1810/
%P 38665-38689
Markdown (Informal)
[GNMR: Runtime Stability Control for Low-Precision Large Language Model Training](https://aclanthology.org/2026.findings-emnlp.1810/) (Kong et al., Findings 2026)
ACL
- Boao Kong, Weichen Jia, Engao Zhang, Guohong Li, Yonghan Dong, Yao Wang, Yaoyuan Wang, Yunke Peng, and Kun Yuan. 2026. GNMR: Runtime Stability Control for Low-Precision Large Language Model Training. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 38665–38689, Budapest, Hungary. Association for Computational Linguistics.