@inproceedings{mangadoddi-2026-multi,
title = "Multi-Signal Interpretability Identifies Task-Critical Circuits for Efficient Model Compression",
author = "Mangadoddi, Srikar Vardhan",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.8/",
pages = "140--154",
ISBN = "979-8-89176-444-6",
abstract = "Post-training quantization methods such as GPTQ and AWQ minimize reconstruction error without asking which weights matter for the task. We combine six interpretability signals to select a sparse skeleton of task-critical MLP neurons that is kept at 16-bit while the rest of the model is quantized. On Gemma-2-2B fine-tuned for SQL generation, the 0.54{\%} skeleton with 3-bit MLPs and 8-bit attention retains 100{\%} of baseline-correct outputs at 4.03x compression, versus 75.2{\%} for uniform 4-bit. On Llama-3-8B fine-tuned for Cypher generation, where uniform 4-bit loses 23 points, the 0.46{\%} skeleton recovers 14.4 of them and a random skeleton of equal size recovers none. Three signals are individually effective and three are not, and the effective set differs between models, so the consensus rule provides robustness no single signal does. Selection matters where uniform quantization leaves headroom, which suggests measuring headroom, selecting by consensus, and protecting as a recipe for larger models. GPTQ error compensation degrades 3-bit layouts by 16 to 18 points regardless of protection."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="mangadoddi-2026-multi">
<titleInfo>
<title>Multi-Signal Interpretability Identifies Task-Critical Circuits for Efficient Model Compression</title>
</titleInfo>
<name type="personal">
<namePart type="given">Srikar</namePart>
<namePart type="given">Vardhan</namePart>
<namePart type="family">Mangadoddi</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>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</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">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>Post-training quantization methods such as GPTQ and AWQ minimize reconstruction error without asking which weights matter for the task. We combine six interpretability signals to select a sparse skeleton of task-critical MLP neurons that is kept at 16-bit while the rest of the model is quantized. On Gemma-2-2B fine-tuned for SQL generation, the 0.54% skeleton with 3-bit MLPs and 8-bit attention retains 100% of baseline-correct outputs at 4.03x compression, versus 75.2% for uniform 4-bit. On Llama-3-8B fine-tuned for Cypher generation, where uniform 4-bit loses 23 points, the 0.46% skeleton recovers 14.4 of them and a random skeleton of equal size recovers none. Three signals are individually effective and three are not, and the effective set differs between models, so the consensus rule provides robustness no single signal does. Selection matters where uniform quantization leaves headroom, which suggests measuring headroom, selecting by consensus, and protecting as a recipe for larger models. GPTQ error compensation degrades 3-bit layouts by 16 to 18 points regardless of protection.</abstract>
<identifier type="citekey">mangadoddi-2026-multi</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.8/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>140</start>
<end>154</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Multi-Signal Interpretability Identifies Task-Critical Circuits for Efficient Model Compression
%A Mangadoddi, Srikar Vardhan
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F mangadoddi-2026-multi
%X Post-training quantization methods such as GPTQ and AWQ minimize reconstruction error without asking which weights matter for the task. We combine six interpretability signals to select a sparse skeleton of task-critical MLP neurons that is kept at 16-bit while the rest of the model is quantized. On Gemma-2-2B fine-tuned for SQL generation, the 0.54% skeleton with 3-bit MLPs and 8-bit attention retains 100% of baseline-correct outputs at 4.03x compression, versus 75.2% for uniform 4-bit. On Llama-3-8B fine-tuned for Cypher generation, where uniform 4-bit loses 23 points, the 0.46% skeleton recovers 14.4 of them and a random skeleton of equal size recovers none. Three signals are individually effective and three are not, and the effective set differs between models, so the consensus rule provides robustness no single signal does. Selection matters where uniform quantization leaves headroom, which suggests measuring headroom, selecting by consensus, and protecting as a recipe for larger models. GPTQ error compensation degrades 3-bit layouts by 16 to 18 points regardless of protection.
%U https://aclanthology.org/2026.findings-aacl.8/
%P 140-154
Markdown (Informal)
[Multi-Signal Interpretability Identifies Task-Critical Circuits for Efficient Model Compression](https://aclanthology.org/2026.findings-aacl.8/) (Mangadoddi, Findings 2026)
ACL