A Unifying Framework for Unsupervised Concept Extraction

2026-04-27

AI Summary

Existing unsupervised concept extraction methods lack a unified theoretical framework and identifiability guarantees, making it difficult to assess their reliability in downstream tasks such as model steering and unlearning. This work formulates concept extraction as an identifiability problem in generative models and introduces the first general theoretical framework for identifiability analysis. By establishing a meta-theorem, the framework reduces complex identifiability proofs to characterizing the intersection of two sets, thereby substantially simplifying theoretical analysis. It unifies the treatment of prominent approaches—including sparse autoencoders and transcoders—under a common theoretical lens and provides principled guidance for designing new algorithms. Empirical evaluations across multiple benchmark methods demonstrate both the theoretical soundness and practical utility of the proposed framework.

Abstract

Techniques for concept extraction, such as sparse autoencoders and transcoders, aim to extract high-level symbolic concepts from low-level nonsymbolic representations. When these extracted concepts are used for downstream tasks such as model steering and unlearning, it is essential to understand their guarantees, or lack thereof. In this work, we present a unified theoretical framework for unsupervised concept extraction, in which we frame the task of concept extraction as identifying a generative model. We present a general meta-theorem for identifiability, which reduces the problem of establishing identifiability guarantees to the problem of characterizing the intersection of two sets. As we demonstrate on a range of widely-used approaches, this meta-theorem substantially simplifies the task of proving such guarantees, thus paving the way for the development of new, principled approaches for concept extraction.

Problem

• unsupervised concept extraction
• identifiability
• generative model
• concept representation
• theoretical framework

Innovation

• unsupervised concept extraction
• generative model
• identifiability
• sparse autoencoders
• theoretical framework