<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="cho-etal-2025-faithfulsae">
    <titleInfo>
        <title>FaithfulSAE: Towards Capturing Faithful Features with Sparse Autoencoders without External Datasets Dependency</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Seonglae</namePart>
        <namePart type="family">Cho</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Harryn</namePart>
        <namePart type="family">Oh</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Donghyun</namePart>
        <namePart type="family">Lee</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Luis</namePart>
        <namePart type="given">Rodrigues</namePart>
        <namePart type="family">Vieira</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Andrew</namePart>
        <namePart type="family">Bermingham</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Ziad</namePart>
        <namePart type="given">El</namePart>
        <namePart type="family">Sayed</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2025-07</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Jin</namePart>
            <namePart type="family">Zhao</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Mingyang</namePart>
            <namePart type="family">Wang</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Zhu</namePart>
            <namePart type="family">Liu</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Vienna, Austria</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-254-1</identifier>
    </relatedItem>
    <abstract>Sparse Autoencoders (SAEs) have emerged as a promising solution for decomposing large language model representations into interpretable features. However, Paulo &amp; Belrose (2025) have highlighted instability across different initialization seeds, and Heap et al. (2025) have pointed out that SAEs may not capture model-internal features. These problems likely stem from training SAEs on external datasets—either collected from the Web or generated by another model—which may contain out-of-distribution (OOD) data beyond the model’s generalisation capabilities. This can result in hallucinated SAE features, which we term “Fake Features”, that misrepresent the model’s internal activations. To address these issues, we propose FaithfulSAE, a method that trains SAEs on the model’s own synthetic dataset. Using FaithfulSAEs, we demonstrate that training SAEs on less-OOD instruction datasets results in SAEs being more stable across seeds. Notably, FaithfulSAEs outperform SAEs trained on webbased datasets in the SAE probing task and exhibit a lower Fake Feature Ratio in 5 out of 7 models. Overall, our approach eliminates the dependency on external datasets, advancing interpretability by better capturing model-internal features while highlighting the often neglected importance of SAE training datasets.</abstract>
    <identifier type="citekey">cho-etal-2025-faithfulsae</identifier>
    <identifier type="doi">10.18653/v1/2025.acl-srw.20</identifier>
    <location>
        <url>https://aclanthology.org/2025.acl-srw.20/</url>
    </location>
    <part>
        <date>2025-07</date>
        <extent unit="page">
            <start>297</start>
            <end>314</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T FaithfulSAE: Towards Capturing Faithful Features with Sparse Autoencoders without External Datasets Dependency
%A Cho, Seonglae
%A Oh, Harryn
%A Lee, Donghyun
%A Vieira, Luis Rodrigues
%A Bermingham, Andrew
%A Sayed, Ziad El
%Y Zhao, Jin
%Y Wang, Mingyang
%Y Liu, Zhu
%S Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)
%D 2025
%8 July
%I Association for Computational Linguistics
%C Vienna, Austria
%@ 979-8-89176-254-1
%F cho-etal-2025-faithfulsae
%X Sparse Autoencoders (SAEs) have emerged as a promising solution for decomposing large language model representations into interpretable features. However, Paulo & Belrose (2025) have highlighted instability across different initialization seeds, and Heap et al. (2025) have pointed out that SAEs may not capture model-internal features. These problems likely stem from training SAEs on external datasets—either collected from the Web or generated by another model—which may contain out-of-distribution (OOD) data beyond the model’s generalisation capabilities. This can result in hallucinated SAE features, which we term “Fake Features”, that misrepresent the model’s internal activations. To address these issues, we propose FaithfulSAE, a method that trains SAEs on the model’s own synthetic dataset. Using FaithfulSAEs, we demonstrate that training SAEs on less-OOD instruction datasets results in SAEs being more stable across seeds. Notably, FaithfulSAEs outperform SAEs trained on webbased datasets in the SAE probing task and exhibit a lower Fake Feature Ratio in 5 out of 7 models. Overall, our approach eliminates the dependency on external datasets, advancing interpretability by better capturing model-internal features while highlighting the often neglected importance of SAE training datasets.
%R 10.18653/v1/2025.acl-srw.20
%U https://aclanthology.org/2025.acl-srw.20/
%U https://doi.org/10.18653/v1/2025.acl-srw.20
%P 297-314

Markdown (Informal)

[FaithfulSAE: Towards Capturing Faithful Features with Sparse Autoencoders without External Datasets Dependency](https://aclanthology.org/2025.acl-srw.20/) (Cho et al., ACL 2025)

• FaithfulSAE: Towards Capturing Faithful Features with Sparse Autoencoders without External Datasets Dependency (Cho et al., ACL 2025)

ACL

• Seonglae Cho, Harryn Oh, Donghyun Lee, Luis Rodrigues Vieira, Andrew Bermingham, and Ziad El Sayed. 2025. FaithfulSAE: Towards Capturing Faithful Features with Sparse Autoencoders without External Datasets Dependency. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop), pages 297–314, Vienna, Austria. Association for Computational Linguistics.