@inproceedings{teja-etal-2026-grace,
title = "{GRACE}-{SVD}: Guided Rank Allocation and Core Equalization {SVD}",
author = "Teja, Andra Siva Sai and
Sahu, Pragya Paramita and
Dave, Vibhor and
Mahadevan, Joshua and
Mahato, Ankit and
Konar, Debanjan and
Mupparthi, Vikram",
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.2136/",
doi = "10.18653/2026.findings-emnlp.2136",
pages = "45569--45587",
ISBN = "979-8-89176-462-0",
abstract = "Large Language Models (LLMs), with 7B+ parameters, remain difficult to deploy on edge devices due to memory latency constraints, despite increasing demand for personalized on-device agents with strong reasoning capabilities. Enabling such deployment, in current hardware specifications, requires aggressive compression while preserving downstream performance. Singular Value Decomposition (SVD)-based methods provide a promising reconstruction-aware compression paradigm, and recent trainable frameworks improve adaptive rank allocation under varying compression budgets. However, maintaining model quality under high compression remains challenging due to statistical mismatches between decomposed cores and compounded degradation from direct high-ratio compression.\par We propose GRACE-SVD, a trainable SVD compression framework combining Guided Rank Allocation with a novel Core Equalization method that aligns decomposed core distributions while preserving reconstruction. We further introduce a progressive compression pipeline with intermediate equalization to mitigate performance collapse under aggressive compression. Across multiple model families and compression settings, our GRACE-SVD consistently outperforms existing SVD techniques, achieving stable compression even at 0.4 parameter ratio, with lower perplexity and stronger zero-shot performance under comparable or lower compression budgets."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="teja-etal-2026-grace">
<titleInfo>
<title>GRACE-SVD: Guided Rank Allocation and Core Equalization SVD</title>
</titleInfo>
<name type="personal">
<namePart type="given">Andra</namePart>
<namePart type="given">Siva</namePart>
<namePart type="given">Sai</namePart>
<namePart type="family">Teja</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pragya</namePart>
<namePart type="given">Paramita</namePart>
<namePart type="family">Sahu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vibhor</namePart>
<namePart type="family">Dave</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Joshua</namePart>
<namePart type="family">Mahadevan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ankit</namePart>
<namePart type="family">Mahato</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Debanjan</namePart>
<namePart type="family">Konar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vikram</namePart>
<namePart type="family">Mupparthi</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>Large Language Models (LLMs), with 7B+ parameters, remain difficult to deploy on edge devices due to memory latency constraints, despite increasing demand for personalized on-device agents with strong reasoning capabilities. Enabling such deployment, in current hardware specifications, requires aggressive compression while preserving downstream performance. Singular Value Decomposition (SVD)-based methods provide a promising reconstruction-aware compression paradigm, and recent trainable frameworks improve adaptive rank allocation under varying compression budgets. However, maintaining model quality under high compression remains challenging due to statistical mismatches between decomposed cores and compounded degradation from direct high-ratio compression.\par We propose GRACE-SVD, a trainable SVD compression framework combining Guided Rank Allocation with a novel Core Equalization method that aligns decomposed core distributions while preserving reconstruction. We further introduce a progressive compression pipeline with intermediate equalization to mitigate performance collapse under aggressive compression. Across multiple model families and compression settings, our GRACE-SVD consistently outperforms existing SVD techniques, achieving stable compression even at 0.4 parameter ratio, with lower perplexity and stronger zero-shot performance under comparable or lower compression budgets.</abstract>
<identifier type="citekey">teja-etal-2026-grace</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.2136</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2136/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>45569</start>
<end>45587</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T GRACE-SVD: Guided Rank Allocation and Core Equalization SVD
%A Teja, Andra Siva Sai
%A Sahu, Pragya Paramita
%A Dave, Vibhor
%A Mahadevan, Joshua
%A Mahato, Ankit
%A Konar, Debanjan
%A Mupparthi, Vikram
%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 teja-etal-2026-grace
%X Large Language Models (LLMs), with 7B+ parameters, remain difficult to deploy on edge devices due to memory latency constraints, despite increasing demand for personalized on-device agents with strong reasoning capabilities. Enabling such deployment, in current hardware specifications, requires aggressive compression while preserving downstream performance. Singular Value Decomposition (SVD)-based methods provide a promising reconstruction-aware compression paradigm, and recent trainable frameworks improve adaptive rank allocation under varying compression budgets. However, maintaining model quality under high compression remains challenging due to statistical mismatches between decomposed cores and compounded degradation from direct high-ratio compression.\par We propose GRACE-SVD, a trainable SVD compression framework combining Guided Rank Allocation with a novel Core Equalization method that aligns decomposed core distributions while preserving reconstruction. We further introduce a progressive compression pipeline with intermediate equalization to mitigate performance collapse under aggressive compression. Across multiple model families and compression settings, our GRACE-SVD consistently outperforms existing SVD techniques, achieving stable compression even at 0.4 parameter ratio, with lower perplexity and stronger zero-shot performance under comparable or lower compression budgets.
%R 10.18653/2026.findings-emnlp.2136
%U https://aclanthology.org/2026.findings-emnlp.2136/
%U https://doi.org/10.18653/2026.findings-emnlp.2136
%P 45569-45587
Markdown (Informal)
[GRACE-SVD: Guided Rank Allocation and Core Equalization SVD](https://aclanthology.org/2026.findings-emnlp.2136/) (Teja et al., Findings 2026)
ACL
- Andra Siva Sai Teja, Pragya Paramita Sahu, Vibhor Dave, Joshua Mahadevan, Ankit Mahato, Debanjan Konar, and Vikram Mupparthi. 2026. GRACE-SVD: Guided Rank Allocation and Core Equalization SVD. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 45569–45587, Budapest, Hungary. Association for Computational Linguistics.