@inproceedings{aryan-2026-concept,
title = "When Does Concept Erasure Change Behavior?",
author = "Aryan, Gagan",
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.21/",
pages = "262--280",
ISBN = "979-8-89176-440-8",
abstract = "Concept erasure can remove linearly recoverable information from a representation without showing that this information influenced the model output being evaluated. For linear edits followed by explicit linear output maps, we characterize the resulting logit change using the output map and representation covariance. We then compare downstream effects with control edits matched for validation-side damage to test whether any benefit depends on the targeted concept. On Bios, Least-squares Concept Erasure (LEACE) reduces gender-probe accuracy to chance but changes only about 0.4{\%} of occupation-logit energy and yields no detectable improvement in true positive rate (TPR) gaps beyond shuffled-label controls. Iterative Nullspace Projection (INLP) changes more occupation-logit energy at higher ranks, but its small, uncertain fairness gains are accompanied by occupation-accuracy loss. A controlled text shortcut yields a large target-specific gain when the classifier relies on the erased marker. In final-state, next-token audits of TinyLlama and Qwen2.5-7B, gender erasure substantially reduces he/she margins but leaves occupation-token choices intact. These results show that successful erasure, output sensitivity, and downstream benefit are distinct and should be evaluated separately."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="aryan-2026-concept">
<titleInfo>
<title>When Does Concept Erasure Change Behavior?</title>
</titleInfo>
<name type="personal">
<namePart type="given">Gagan</namePart>
<namePart type="family">Aryan</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>Concept erasure can remove linearly recoverable information from a representation without showing that this information influenced the model output being evaluated. For linear edits followed by explicit linear output maps, we characterize the resulting logit change using the output map and representation covariance. We then compare downstream effects with control edits matched for validation-side damage to test whether any benefit depends on the targeted concept. On Bios, Least-squares Concept Erasure (LEACE) reduces gender-probe accuracy to chance but changes only about 0.4% of occupation-logit energy and yields no detectable improvement in true positive rate (TPR) gaps beyond shuffled-label controls. Iterative Nullspace Projection (INLP) changes more occupation-logit energy at higher ranks, but its small, uncertain fairness gains are accompanied by occupation-accuracy loss. A controlled text shortcut yields a large target-specific gain when the classifier relies on the erased marker. In final-state, next-token audits of TinyLlama and Qwen2.5-7B, gender erasure substantially reduces he/she margins but leaves occupation-token choices intact. These results show that successful erasure, output sensitivity, and downstream benefit are distinct and should be evaluated separately.</abstract>
<identifier type="citekey">aryan-2026-concept</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-short.21/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>262</start>
<end>280</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T When Does Concept Erasure Change Behavior?
%A Aryan, Gagan
%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 aryan-2026-concept
%X Concept erasure can remove linearly recoverable information from a representation without showing that this information influenced the model output being evaluated. For linear edits followed by explicit linear output maps, we characterize the resulting logit change using the output map and representation covariance. We then compare downstream effects with control edits matched for validation-side damage to test whether any benefit depends on the targeted concept. On Bios, Least-squares Concept Erasure (LEACE) reduces gender-probe accuracy to chance but changes only about 0.4% of occupation-logit energy and yields no detectable improvement in true positive rate (TPR) gaps beyond shuffled-label controls. Iterative Nullspace Projection (INLP) changes more occupation-logit energy at higher ranks, but its small, uncertain fairness gains are accompanied by occupation-accuracy loss. A controlled text shortcut yields a large target-specific gain when the classifier relies on the erased marker. In final-state, next-token audits of TinyLlama and Qwen2.5-7B, gender erasure substantially reduces he/she margins but leaves occupation-token choices intact. These results show that successful erasure, output sensitivity, and downstream benefit are distinct and should be evaluated separately.
%U https://aclanthology.org/2026.aacl-short.21/
%P 262-280
Markdown (Informal)
[When Does Concept Erasure Change Behavior?](https://aclanthology.org/2026.aacl-short.21/) (Aryan, AACL-IJCNLP 2026)
ACL
- Gagan Aryan. 2026. When Does Concept Erasure Change Behavior?. In 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), pages 262–280, Hengqin, China. Association for Computational Linguistics.