@inproceedings{zeinali-fu-2026-distribution,
title = "From Distribution to Decision: A Diagnostic for Composable {PTQ} Primitives",
author = "Zeinali, Dorsa and
Fu, Yun",
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.2105/",
pages = "44825--44854",
ISBN = "979-8-89176-462-0",
abstract = "Post-training quantization (PTQ) methods are typically architecture-specific. Yet two distributional pathologies they target, dense cores with sparse outliers and input-conditioned activation clusters, recur across model families. We argue that for these pathologies the unit of selection should be the tensor distribution, not the architecture. We propose two composable primitives, dispatched from one calibration pass by a two-statistic diagnostic: Dual-Band Affine Folding (DBAF) where a per-tensor outlier gate fires, and Prompt-Conditioned Scale Anchoring (PCSA) where a per-prompt compactness ratio does. The diagnostic is a selection heuristic, not a calibrated predictor. Four prediction-then-outcome checks are consistent with the dispatch: a rotation cliff on a six-host LLaMA-3-8B W4A4 matrix, cross-family transfer to Qwen-2.5-7B, per-site gating on the KV-cache and SAM-B (+5pp NIAH at 8k, +4.8 mAP), and a negative SwinIR-{\texttimes}3 prediction at training-free W4A16. Both primitives are training-free and calibrate in seconds. The gate{'}s boundary is learnable, not merely hand-set: a logistic dispatcher trained with each model family held out reproduces it at mean AUC 0.962. Calibrated once on LLaMA-3-8B, the same thresholds dispatch across seven model families in six domains; off distribution the gate errs toward skipping."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zeinali-fu-2026-distribution">
<titleInfo>
<title>From Distribution to Decision: A Diagnostic for Composable PTQ Primitives</title>
</titleInfo>
<name type="personal">
<namePart type="given">Dorsa</namePart>
<namePart type="family">Zeinali</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yun</namePart>
<namePart type="family">Fu</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) methods are typically architecture-specific. Yet two distributional pathologies they target, dense cores with sparse outliers and input-conditioned activation clusters, recur across model families. We argue that for these pathologies the unit of selection should be the tensor distribution, not the architecture. We propose two composable primitives, dispatched from one calibration pass by a two-statistic diagnostic: Dual-Band Affine Folding (DBAF) where a per-tensor outlier gate fires, and Prompt-Conditioned Scale Anchoring (PCSA) where a per-prompt compactness ratio does. The diagnostic is a selection heuristic, not a calibrated predictor. Four prediction-then-outcome checks are consistent with the dispatch: a rotation cliff on a six-host LLaMA-3-8B W4A4 matrix, cross-family transfer to Qwen-2.5-7B, per-site gating on the KV-cache and SAM-B (+5pp NIAH at 8k, +4.8 mAP), and a negative SwinIR-×3 prediction at training-free W4A16. Both primitives are training-free and calibrate in seconds. The gate’s boundary is learnable, not merely hand-set: a logistic dispatcher trained with each model family held out reproduces it at mean AUC 0.962. Calibrated once on LLaMA-3-8B, the same thresholds dispatch across seven model families in six domains; off distribution the gate errs toward skipping.</abstract>
<identifier type="citekey">zeinali-fu-2026-distribution</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2105/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>44825</start>
<end>44854</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T From Distribution to Decision: A Diagnostic for Composable PTQ Primitives
%A Zeinali, Dorsa
%A Fu, Yun
%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 zeinali-fu-2026-distribution
%X Post-training quantization (PTQ) methods are typically architecture-specific. Yet two distributional pathologies they target, dense cores with sparse outliers and input-conditioned activation clusters, recur across model families. We argue that for these pathologies the unit of selection should be the tensor distribution, not the architecture. We propose two composable primitives, dispatched from one calibration pass by a two-statistic diagnostic: Dual-Band Affine Folding (DBAF) where a per-tensor outlier gate fires, and Prompt-Conditioned Scale Anchoring (PCSA) where a per-prompt compactness ratio does. The diagnostic is a selection heuristic, not a calibrated predictor. Four prediction-then-outcome checks are consistent with the dispatch: a rotation cliff on a six-host LLaMA-3-8B W4A4 matrix, cross-family transfer to Qwen-2.5-7B, per-site gating on the KV-cache and SAM-B (+5pp NIAH at 8k, +4.8 mAP), and a negative SwinIR-×3 prediction at training-free W4A16. Both primitives are training-free and calibrate in seconds. The gate’s boundary is learnable, not merely hand-set: a logistic dispatcher trained with each model family held out reproduces it at mean AUC 0.962. Calibrated once on LLaMA-3-8B, the same thresholds dispatch across seven model families in six domains; off distribution the gate errs toward skipping.
%U https://aclanthology.org/2026.findings-emnlp.2105/
%P 44825-44854
Markdown (Informal)
[From Distribution to Decision: A Diagnostic for Composable PTQ Primitives](https://aclanthology.org/2026.findings-emnlp.2105/) (Zeinali & Fu, Findings 2026)
ACL