Sharkey et al. (2023) interim research report
Sparse Autoencoders Find Highly Interpretable Features in Language Models — inherited
An unpublished interim research report ('Taking features out of superposition with sparse autoencoders,' Sharkey, Braun & Millidge, 2023) that first proposed training sparse autoencoders on language model activations to address superposition. Cited as the direct motivation and foundation this paper builds on and scales up.
The core idea behind this whole method didn't debut in a peer-reviewed paper at all. It first appeared months earlier in an informal research report, tested only on toy problems with features whose ground truth was already known, posted by researchers betting the idea would scale further than that.
What comes next covers what that earlier report actually showed and left untested, and how its relationship to this paper differs from an ordinary citation: the same lead researcher carried the idea from a toy demonstration through to validation on a real language model's activations.
The unpublished report that started this paper
This concept names a specific piece of literature: Sharkey, Braun & Millidge's 2023 interim research report, posted to the AlignmentForum under the title 'Taking features out of superposition with sparse autoencoders,' the exact phrase this paper's own Section 2 heading borrows. It was the first work to propose training sparse autoencoders on language model activations specifically to address Superposition, predating this paper by several months and appearing outside the peer-reviewed venues typical of the wider machine-learning literature. The paper names it as its direct point of departure: 'Building on Sharkey et al. (2023), we train sparse autoencoders to learn these sets of directions' (sparse-autoencoders, §"1 INTRODUCTION", p. 2).
From toy demonstration to validated method
The interim report's original contribution was narrower than this paper's: it demonstrated, in toy settings with known ground-truth features, that $\ell_1$-penalized reconstruction could recover those features empirically, evidence this paper cites directly when justifying its own sparsity penalty (sparse-autoencoders, §"2 TAKING FEATURES OUT OF SUPERPOSITION WITH SPARSE DICTIONARY LEARNING", p. 3). It did not run the method on a real language model's activations at scale, nor test the results with Autointerpretability score, causal Activation patching against a task like Indirect Object Identification (IOI) task, or the hand-run monosemanticity case studies this paper contributes. Lee Sharkey, the interim report's lead author, is also a co-author of this paper, making the relationship one of direct continuation by the same researcher rather than independent replication.
Two kinds of prior-work relationship in the same paper
Both the related-work section and the introduction single this report out, alongside Transformer visualization via dictionary learning (Yun et al., 2021), as the closest prior work, but the paper frames the interim report specifically as having 'motivated this work' (sparse-autoencoders, §"6.1 RELATED WORK", p. 8), a stronger relationship than the parallel citation to Yun et al. (2021), whose granularity and evaluation differ more substantially. The dictionary-learning lineage running through this corpus therefore converges two distinct kinds of prior-work relationship in this one paper: a classical mathematical framework, Sparse dictionary learning / sparse coding, inherited from outside interpretability entirely, and an unpublished but methodologically identical alignment-community report scaled up by its own original author.