@article{zheng-etal-2026-model,
title = "Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders",
author = "Zheng, Carolina and
Beltran-Velez, Nicolas and
Karlekar, Sweta and
Shi, Claudia and
Nazaret, Achille and
Mallik, Asif and
Feder, Amir and
Blei, David M.",
journal = "Transactions of the Association for Computational Linguistics",
volume = "14",
year = "2026",
address = "Cambridge, MA",
publisher = "MIT Press",
url = "https://aclanthology.org/2026.tacl-1.99/",
doi = "10.1162/tacl.a.797",
pages = "2187--2212",
abstract = "Traditional topic models are effective at uncovering latent themes in large text collections. However, due to their reliance on bag-of-words representations, they struggle to capture semantically abstract features. While some neural variants use richer representations, they are similarly constrained by expressing topics as word lists, which limits their ability to articulate complex topics. We introduce Mechanistic Topic Models (MTMs), a class of topic models that operate on interpretable features learned by sparse autoencoders (SAEs). By defining topics over this semantically rich space, MTMs can reveal deeper conceptual themes with expressive feature descriptions. Moreover, uniquely among topic models, MTMs enable controllable text generation using topic steering vectors. To properly evaluate MTM topics against word list approaches, we propose topic judge, an LLM-based pairwise comparison evaluation framework. Across eight datasets, MTMs match or exceed traditional and neural baselines on coherence metrics, are consistently preferred by topic judge, and enable effective LLM steering.1"
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zheng-etal-2026-model">
<titleInfo>
<title>Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders</title>
</titleInfo>
<name type="personal">
<namePart type="given">Carolina</namePart>
<namePart type="family">Zheng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nicolas</namePart>
<namePart type="family">Beltran-Velez</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sweta</namePart>
<namePart type="family">Karlekar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Claudia</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Achille</namePart>
<namePart type="family">Nazaret</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Asif</namePart>
<namePart type="family">Mallik</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Amir</namePart>
<namePart type="family">Feder</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">David</namePart>
<namePart type="given">M</namePart>
<namePart type="family">Blei</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<genre authority="bibutilsgt">journal article</genre>
<relatedItem type="host">
<titleInfo>
<title>Transactions of the Association for Computational Linguistics</title>
</titleInfo>
<originInfo>
<issuance>continuing</issuance>
<publisher>MIT Press</publisher>
<place>
<placeTerm type="text">Cambridge, MA</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">periodical</genre>
<genre authority="bibutilsgt">academic journal</genre>
</relatedItem>
<abstract>Traditional topic models are effective at uncovering latent themes in large text collections. However, due to their reliance on bag-of-words representations, they struggle to capture semantically abstract features. While some neural variants use richer representations, they are similarly constrained by expressing topics as word lists, which limits their ability to articulate complex topics. We introduce Mechanistic Topic Models (MTMs), a class of topic models that operate on interpretable features learned by sparse autoencoders (SAEs). By defining topics over this semantically rich space, MTMs can reveal deeper conceptual themes with expressive feature descriptions. Moreover, uniquely among topic models, MTMs enable controllable text generation using topic steering vectors. To properly evaluate MTM topics against word list approaches, we propose topic judge, an LLM-based pairwise comparison evaluation framework. Across eight datasets, MTMs match or exceed traditional and neural baselines on coherence metrics, are consistently preferred by topic judge, and enable effective LLM steering.1</abstract>
<identifier type="citekey">zheng-etal-2026-model</identifier>
<identifier type="doi">10.1162/tacl.a.797</identifier>
<location>
<url>https://aclanthology.org/2026.tacl-1.99/</url>
</location>
<part>
<date>2026</date>
<detail type="volume"><number>14</number></detail>
<extent unit="page">
<start>2187</start>
<end>2212</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Journal Article
%T Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders
%A Zheng, Carolina
%A Beltran-Velez, Nicolas
%A Karlekar, Sweta
%A Shi, Claudia
%A Nazaret, Achille
%A Mallik, Asif
%A Feder, Amir
%A Blei, David M.
%J Transactions of the Association for Computational Linguistics
%D 2026
%V 14
%I MIT Press
%C Cambridge, MA
%F zheng-etal-2026-model
%X Traditional topic models are effective at uncovering latent themes in large text collections. However, due to their reliance on bag-of-words representations, they struggle to capture semantically abstract features. While some neural variants use richer representations, they are similarly constrained by expressing topics as word lists, which limits their ability to articulate complex topics. We introduce Mechanistic Topic Models (MTMs), a class of topic models that operate on interpretable features learned by sparse autoencoders (SAEs). By defining topics over this semantically rich space, MTMs can reveal deeper conceptual themes with expressive feature descriptions. Moreover, uniquely among topic models, MTMs enable controllable text generation using topic steering vectors. To properly evaluate MTM topics against word list approaches, we propose topic judge, an LLM-based pairwise comparison evaluation framework. Across eight datasets, MTMs match or exceed traditional and neural baselines on coherence metrics, are consistently preferred by topic judge, and enable effective LLM steering.1
%R 10.1162/tacl.a.797
%U https://aclanthology.org/2026.tacl-1.99/
%U https://doi.org/10.1162/tacl.a.797
%P 2187-2212
Markdown (Informal)
[Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders](https://aclanthology.org/2026.tacl-1.99/) (Zheng et al., TACL 2026)
ACL