@inproceedings{lee-kim-2026-auditing,
title = "Auditing {PEFT} Claims: A Stability Probe Unifying Three {L}o{RA}-Variant Fragility Patterns",
author = "Lee, Suan and
Kim, JaeSeong",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 2: Short Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-short.32/",
pages = "474--494",
ISBN = "979-8-89176-440-8",
abstract = "Most published PEFT improvements over LoRA do not survive paired multi-seed evaluation. We audit four methods {---} DoRA, PiSSA, AdaLoRA, and rsLoRA {---} against LoRA across 16 claim cells spanning 125M{--}7B, using paired seeds, bootstrap confidence intervals, sign-flip permutation tests, and Holm{--}Bonferroni correction, with adjudication rules committed before the confirmatory runs. Only 2 of 15 confirmatory cells remain supported after correction, and 4 reverse outright. Three findings stand out. (i) DoRA shows no supported gain in any of its seven audited cells. (ii) PiSSA{'}s +2.6{\%} GSM8K advantage under a common budget reverses under paper-faithful hyperparameters, including in the cleanest claim-matched 7B setting (Gemma-7B, MetaMathQA {\textrightarrow} GSM8K: {\ensuremath{\Delta}} = {\ensuremath{-}}1.58pp, p {\ensuremath{\approx}} 0.004, 1W/9L). (iii) PiSSA{'}s in-distribution NLI gain reverses on adversarial OOD yet improves on counterfactual sentiment, so in-distribution audits alone can mislead. A 303-run stability landscape then supplies a mechanism for three of the four variants. The effective adapter learning rate, lr {\texttimes} {\ensuremath{\alpha}}/{\ensuremath{\sqrt{}}}r, marks a risk boundary near 5{\texttimes}10{\ensuremath{^-}}{\ensuremath{^4}} on BERT-base/MNLI/AdamW: nothing below it collapsed (40/40 stable), while 11 of 32 configurations above it did. The boundary transfers to 7B (Mistral rsLoRA: 3/10 seeds degrade at {\ensuremath{\alpha}}/{\ensuremath{\sqrt{}}}r = 4) and predicts DoRA{'}s {\ensuremath{\alpha}}-driven failures at high rank, but it is calibrated rather than universal: a second 7B cell trains stably well above it once dtype is controlled, and we withdraw that cell as supporting evidence. AdaLoRA is excluded from this account: its failure is schedule-driven rather than driven by lr {\texttimes} {\ensuremath{\alpha}}/{\ensuremath{\sqrt{}}}r. We release all configs, seeds, and scripts, and propose a minimal reporting standard: {\ensuremath{\geq}}10 paired seeds, bootstrap CIs, a regime ablation, and disclosure of {\ensuremath{\alpha}}/{\ensuremath{\sqrt{}}}r."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="lee-kim-2026-auditing">
<titleInfo>
<title>Auditing PEFT Claims: A Stability Probe Unifying Three LoRA-Variant Fragility Patterns</title>
</titleInfo>
<name type="personal">
<namePart type="given">Suan</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">JaeSeong</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 2: Short Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-440-8</identifier>
</relatedItem>
<abstract>Most published PEFT improvements over LoRA do not survive paired multi-seed evaluation. We audit four methods — DoRA, PiSSA, AdaLoRA, and rsLoRA — against LoRA across 16 claim cells spanning 125M–7B, using paired seeds, bootstrap confidence intervals, sign-flip permutation tests, and Holm–Bonferroni correction, with adjudication rules committed before the confirmatory runs. Only 2 of 15 confirmatory cells remain supported after correction, and 4 reverse outright. Three findings stand out. (i) DoRA shows no supported gain in any of its seven audited cells. (ii) PiSSA’s +2.6% GSM8K advantage under a common budget reverses under paper-faithful hyperparameters, including in the cleanest claim-matched 7B setting (Gemma-7B, MetaMathQA → GSM8K: \ensuremathΔ = \ensuremath-1.58pp, p \ensuremath\approx 0.004, 1W/9L). (iii) PiSSA’s in-distribution NLI gain reverses on adversarial OOD yet improves on counterfactual sentiment, so in-distribution audits alone can mislead. A 303-run stability landscape then supplies a mechanism for three of the four variants. The effective adapter learning rate, lr × \ensuremathα/\ensuremath\sqrtr, marks a risk boundary near 5×10\ensuremath⁻\ensuremath⁴ on BERT-base/MNLI/AdamW: nothing below it collapsed (40/40 stable), while 11 of 32 configurations above it did. The boundary transfers to 7B (Mistral rsLoRA: 3/10 seeds degrade at \ensuremathα/\ensuremath\sqrtr = 4) and predicts DoRA’s \ensuremathα-driven failures at high rank, but it is calibrated rather than universal: a second 7B cell trains stably well above it once dtype is controlled, and we withdraw that cell as supporting evidence. AdaLoRA is excluded from this account: its failure is schedule-driven rather than driven by lr × \ensuremathα/\ensuremath\sqrtr. We release all configs, seeds, and scripts, and propose a minimal reporting standard: \ensuremath\geq10 paired seeds, bootstrap CIs, a regime ablation, and disclosure of \ensuremathα/\ensuremath\sqrtr.</abstract>
<identifier type="citekey">lee-kim-2026-auditing</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-short.32/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>474</start>
<end>494</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Auditing PEFT Claims: A Stability Probe Unifying Three LoRA-Variant Fragility Patterns
%A Lee, Suan
%A Kim, JaeSeong
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 2: Short Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-440-8
%F lee-kim-2026-auditing
%X Most published PEFT improvements over LoRA do not survive paired multi-seed evaluation. We audit four methods — DoRA, PiSSA, AdaLoRA, and rsLoRA — against LoRA across 16 claim cells spanning 125M–7B, using paired seeds, bootstrap confidence intervals, sign-flip permutation tests, and Holm–Bonferroni correction, with adjudication rules committed before the confirmatory runs. Only 2 of 15 confirmatory cells remain supported after correction, and 4 reverse outright. Three findings stand out. (i) DoRA shows no supported gain in any of its seven audited cells. (ii) PiSSA’s +2.6% GSM8K advantage under a common budget reverses under paper-faithful hyperparameters, including in the cleanest claim-matched 7B setting (Gemma-7B, MetaMathQA → GSM8K: \ensuremathΔ = \ensuremath-1.58pp, p \ensuremath\approx 0.004, 1W/9L). (iii) PiSSA’s in-distribution NLI gain reverses on adversarial OOD yet improves on counterfactual sentiment, so in-distribution audits alone can mislead. A 303-run stability landscape then supplies a mechanism for three of the four variants. The effective adapter learning rate, lr × \ensuremathα/\ensuremath\sqrtr, marks a risk boundary near 5×10\ensuremath⁻\ensuremath⁴ on BERT-base/MNLI/AdamW: nothing below it collapsed (40/40 stable), while 11 of 32 configurations above it did. The boundary transfers to 7B (Mistral rsLoRA: 3/10 seeds degrade at \ensuremathα/\ensuremath\sqrtr = 4) and predicts DoRA’s \ensuremathα-driven failures at high rank, but it is calibrated rather than universal: a second 7B cell trains stably well above it once dtype is controlled, and we withdraw that cell as supporting evidence. AdaLoRA is excluded from this account: its failure is schedule-driven rather than driven by lr × \ensuremathα/\ensuremath\sqrtr. We release all configs, seeds, and scripts, and propose a minimal reporting standard: \ensuremath\geq10 paired seeds, bootstrap CIs, a regime ablation, and disclosure of \ensuremathα/\ensuremath\sqrtr.
%U https://aclanthology.org/2026.aacl-short.32/
%P 474-494
Markdown (Informal)
[Auditing PEFT Claims: A Stability Probe Unifying Three LoRA-Variant Fragility Patterns](https://aclanthology.org/2026.aacl-short.32/) (Lee & Kim, AACL-IJCNLP 2026)
ACL