@inproceedings{su-zhang-2026-structural,
title = "Structural Collapse of Safety-Critical Weight Protection Scorers Under Group-Wise Post-Training Quantization",
author = "Su, Jianchang and
Zhang, Wei",
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.2116/",
pages = "45113--45134",
ISBN = "979-8-89176-462-0",
abstract = "Post-training quantization (PTQ) of instruction-tuned LLMs to 3 bits often destroys safety alignment, motivating safety-protection methods that keep a small budget of ``safety-critical'' weight blocks in full precision. We give a theoretical account of the many scoring rules proposed to choose those blocks. We prove that under symmetric group-wise PTQ the Taylor (SSMP) scorer concentrates to the gradient magnitude times a per-group constant; the gradient-times-weight (SNIP) scorer is empirically near-identical, so the entire compound family $\alpha|g\cdot w|+\beta|g\cdot\delta w|$ collapses to a single block ranking (worst-case Spearman 0.9971), and only pure magnitude or pure Fisher selects different blocks. We further show that recovered refusal behavior under small budgets is decoupled from the underlying safety mechanism: refusal can be restored while capability, adversarial robustness, and over-refusal balance stay at the unprotected level; at 3 bits heuristic attack metrics conflate incoherence with harm; and on weakly-aligned models protection raises classifier-judged attack success above the unprotected floor. Experiments on five open-weight 8{--}9B models across 110 conditions, including classifier-judged GCG and PAIR attacks and activation-level analysis, confirm every prediction; a four-category classification predictable from one baseline metric explains the cross-model variation. Code and configurations: https://github.com/jc-su/Safepress."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="su-zhang-2026-structural">
<titleInfo>
<title>Structural Collapse of Safety-Critical Weight Protection Scorers Under Group-Wise Post-Training Quantization</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jianchang</namePart>
<namePart type="family">Su</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Zhang</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>Post-training quantization (PTQ) of instruction-tuned LLMs to 3 bits often destroys safety alignment, motivating safety-protection methods that keep a small budget of “safety-critical” weight blocks in full precision. We give a theoretical account of the many scoring rules proposed to choose those blocks. We prove that under symmetric group-wise PTQ the Taylor (SSMP) scorer concentrates to the gradient magnitude times a per-group constant; the gradient-times-weight (SNIP) scorer is empirically near-identical, so the entire compound family α|g· w|+β|g·δ w| collapses to a single block ranking (worst-case Spearman 0.9971), and only pure magnitude or pure Fisher selects different blocks. We further show that recovered refusal behavior under small budgets is decoupled from the underlying safety mechanism: refusal can be restored while capability, adversarial robustness, and over-refusal balance stay at the unprotected level; at 3 bits heuristic attack metrics conflate incoherence with harm; and on weakly-aligned models protection raises classifier-judged attack success above the unprotected floor. Experiments on five open-weight 8–9B models across 110 conditions, including classifier-judged GCG and PAIR attacks and activation-level analysis, confirm every prediction; a four-category classification predictable from one baseline metric explains the cross-model variation. Code and configurations: https://github.com/jc-su/Safepress.</abstract>
<identifier type="citekey">su-zhang-2026-structural</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2116/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>45113</start>
<end>45134</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Structural Collapse of Safety-Critical Weight Protection Scorers Under Group-Wise Post-Training Quantization
%A Su, Jianchang
%A Zhang, Wei
%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 su-zhang-2026-structural
%X Post-training quantization (PTQ) of instruction-tuned LLMs to 3 bits often destroys safety alignment, motivating safety-protection methods that keep a small budget of “safety-critical” weight blocks in full precision. We give a theoretical account of the many scoring rules proposed to choose those blocks. We prove that under symmetric group-wise PTQ the Taylor (SSMP) scorer concentrates to the gradient magnitude times a per-group constant; the gradient-times-weight (SNIP) scorer is empirically near-identical, so the entire compound family α|g· w|+β|g·δ w| collapses to a single block ranking (worst-case Spearman 0.9971), and only pure magnitude or pure Fisher selects different blocks. We further show that recovered refusal behavior under small budgets is decoupled from the underlying safety mechanism: refusal can be restored while capability, adversarial robustness, and over-refusal balance stay at the unprotected level; at 3 bits heuristic attack metrics conflate incoherence with harm; and on weakly-aligned models protection raises classifier-judged attack success above the unprotected floor. Experiments on five open-weight 8–9B models across 110 conditions, including classifier-judged GCG and PAIR attacks and activation-level analysis, confirm every prediction; a four-category classification predictable from one baseline metric explains the cross-model variation. Code and configurations: https://github.com/jc-su/Safepress.
%U https://aclanthology.org/2026.findings-emnlp.2116/
%P 45113-45134
Markdown (Informal)
[Structural Collapse of Safety-Critical Weight Protection Scorers Under Group-Wise Post-Training Quantization](https://aclanthology.org/2026.findings-emnlp.2116/) (Su & Zhang, Findings 2026)
ACL