Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zheng, Carolina, Beltran-Velez, Nicolas, Karlekar, Sweta, Shi, Claudia, Nazaret, Achille, Mallik, Asif, Feder, Amir, Blei, David M.
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915418625540096
author Zheng, Carolina
Beltran-Velez, Nicolas
Karlekar, Sweta
Shi, Claudia
Nazaret, Achille
Mallik, Asif
Feder, Amir
Blei, David M.
author_facet Zheng, Carolina
Beltran-Velez, Nicolas
Karlekar, Sweta
Shi, Claudia
Nazaret, Achille
Mallik, Asif
Feder, Amir
Blei, David M.
contents 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-based steering vectors. To properly evaluate MTM topics against word-list-based approaches, we propose \textit{topic judge}, an LLM-based pairwise comparison evaluation framework. Across five datasets, MTMs match or exceed traditional and neural baselines on coherence metrics, are consistently preferred by topic judge, and enable effective steering of LLM outputs.
format Preprint
id arxiv_https___arxiv_org_abs_2507_23220
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders
Zheng, Carolina
Beltran-Velez, Nicolas
Karlekar, Sweta
Shi, Claudia
Nazaret, Achille
Mallik, Asif
Feder, Amir
Blei, David M.
Computation and Language
Machine Learning
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-based steering vectors. To properly evaluate MTM topics against word-list-based approaches, we propose \textit{topic judge}, an LLM-based pairwise comparison evaluation framework. Across five datasets, MTMs match or exceed traditional and neural baselines on coherence metrics, are consistently preferred by topic judge, and enable effective steering of LLM outputs.
title Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2507.23220