@inproceedings{punglia-etal-2026-beyond,
title = "Beyond Accuracy: What Language-Subspace Interventions Capture, Remove, and Improve",
author = "Punglia, Shaanil and
Telwadkar, Dadhichi and
Rathore, Vaibhav and
Sureka, Siddharth and
Bhargava, Neha",
editor = "Wang, Xiaoman and
Wang, Qianyu and
Liu, Zhengyuan and
Chen, Nancy F. and
Yuan, Zheng and
Lan, Yunshi and
Jiang, Jing",
booktitle = "Proceedings of the 1st Workshop on Multi-modality and Multi-lingual Language Learning",
month = nov,
year = "2026",
address = "Hengqin, Zhuhai, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.multilll-1.2/",
pages = "19--32",
ISBN = "979-8-89176-449-1",
abstract = "Language-subspace interventions aim to improve multilingual reasoning by suppressing language-related components in mid-layer activations. Using MATH and MGSM across four open-weight LLMs, we ask three questions: do estimated language subspaces capture translation-associated variation? Does projecting them out remove language information? And do they improve accuracy without compromising response language? For the first question, we build subspaces using the existing Mean-SVD construction and our DiffPCA. Both capture substantial translation-associated variation and show similar patterns across model depth. For the second, we project out each subspace and train linear probes to classify language from the projected activations. The probes still identify language almost perfectly, showing that neither subspace exhaustively captures language information. For the third, we evaluate several intervention configurations on Qwen2.5-7B, by varying subspace creation methods and intervention operators. We find that the intervention that most improves MGSM accuracy sacrifices response-language fidelity, with only one intervention improving joint accuracy{---}answers that are correct and in the target language{---}slightly, by 1.60 points. Interestingly, subtle changes in the prompt alone can reproduce the accuracy gains albeit with significant damage to language fidelity. We argue that causal interventions must be judged on more than accuracy, to fully account for language substitution trade-offs."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="punglia-etal-2026-beyond">
<titleInfo>
<title>Beyond Accuracy: What Language-Subspace Interventions Capture, Remove, and Improve</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shaanil</namePart>
<namePart type="family">Punglia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dadhichi</namePart>
<namePart type="family">Telwadkar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vaibhav</namePart>
<namePart type="family">Rathore</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Siddharth</namePart>
<namePart type="family">Sureka</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Neha</namePart>
<namePart type="family">Bhargava</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 1st Workshop on Multi-modality and Multi-lingual Language Learning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Xiaoman</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qianyu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhengyuan</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nancy</namePart>
<namePart type="given">F</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zheng</namePart>
<namePart type="family">Yuan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yunshi</namePart>
<namePart type="family">Lan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jing</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, Zhuhai, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-449-1</identifier>
</relatedItem>
<abstract>Language-subspace interventions aim to improve multilingual reasoning by suppressing language-related components in mid-layer activations. Using MATH and MGSM across four open-weight LLMs, we ask three questions: do estimated language subspaces capture translation-associated variation? Does projecting them out remove language information? And do they improve accuracy without compromising response language? For the first question, we build subspaces using the existing Mean-SVD construction and our DiffPCA. Both capture substantial translation-associated variation and show similar patterns across model depth. For the second, we project out each subspace and train linear probes to classify language from the projected activations. The probes still identify language almost perfectly, showing that neither subspace exhaustively captures language information. For the third, we evaluate several intervention configurations on Qwen2.5-7B, by varying subspace creation methods and intervention operators. We find that the intervention that most improves MGSM accuracy sacrifices response-language fidelity, with only one intervention improving joint accuracy—answers that are correct and in the target language—slightly, by 1.60 points. Interestingly, subtle changes in the prompt alone can reproduce the accuracy gains albeit with significant damage to language fidelity. We argue that causal interventions must be judged on more than accuracy, to fully account for language substitution trade-offs.</abstract>
<identifier type="citekey">punglia-etal-2026-beyond</identifier>
<location>
<url>https://aclanthology.org/2026.multilll-1.2/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>19</start>
<end>32</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Beyond Accuracy: What Language-Subspace Interventions Capture, Remove, and Improve
%A Punglia, Shaanil
%A Telwadkar, Dadhichi
%A Rathore, Vaibhav
%A Sureka, Siddharth
%A Bhargava, Neha
%Y Wang, Xiaoman
%Y Wang, Qianyu
%Y Liu, Zhengyuan
%Y Chen, Nancy F.
%Y Yuan, Zheng
%Y Lan, Yunshi
%Y Jiang, Jing
%S Proceedings of the 1st Workshop on Multi-modality and Multi-lingual Language Learning
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, Zhuhai, China
%@ 979-8-89176-449-1
%F punglia-etal-2026-beyond
%X Language-subspace interventions aim to improve multilingual reasoning by suppressing language-related components in mid-layer activations. Using MATH and MGSM across four open-weight LLMs, we ask three questions: do estimated language subspaces capture translation-associated variation? Does projecting them out remove language information? And do they improve accuracy without compromising response language? For the first question, we build subspaces using the existing Mean-SVD construction and our DiffPCA. Both capture substantial translation-associated variation and show similar patterns across model depth. For the second, we project out each subspace and train linear probes to classify language from the projected activations. The probes still identify language almost perfectly, showing that neither subspace exhaustively captures language information. For the third, we evaluate several intervention configurations on Qwen2.5-7B, by varying subspace creation methods and intervention operators. We find that the intervention that most improves MGSM accuracy sacrifices response-language fidelity, with only one intervention improving joint accuracy—answers that are correct and in the target language—slightly, by 1.60 points. Interestingly, subtle changes in the prompt alone can reproduce the accuracy gains albeit with significant damage to language fidelity. We argue that causal interventions must be judged on more than accuracy, to fully account for language substitution trade-offs.
%U https://aclanthology.org/2026.multilll-1.2/
%P 19-32
Markdown (Informal)
[Beyond Accuracy: What Language-Subspace Interventions Capture, Remove, and Improve](https://aclanthology.org/2026.multilll-1.2/) (Punglia et al., Multi-LLL 2026)
ACL
- Shaanil Punglia, Dadhichi Telwadkar, Vaibhav Rathore, Siddharth Sureka, and Neha Bhargava. 2026. Beyond Accuracy: What Language-Subspace Interventions Capture, Remove, and Improve. In Proceedings of the 1st Workshop on Multi-modality and Multi-lingual Language Learning, pages 19–32, Hengqin, Zhuhai, China. Association for Computational Linguistics.