Antipodal Pairing and Mechanistic Signals in Dense SAE Latents

Alessandro Stolfo, Ben Peng Wu, Mrinmaya Sachan

Track: long paper (up to 4 pages)

Keywords: mechanistic interpretability, SAE, sparse autoencoder

Abstract:

Sparse autoencoders (SAEs) are designed to extract interpretable features from language models, yet they often yield frequently activating latents that remain difficult to interpret.
It is still an open question whether these \textit{dense} latents are an undesired training artifact or whether they represent fundamentally dense signals in the model's activations.
dense latents capture fundamental signals which (1) align with principal directions of variance in the model's residual stream and (2) reconstruct a subspace of the unembedding matrix that was linked by previous work to internal model computation.
Furthermore, we show that these latents typically emerge as nearly antipodal pairs that collaboratively reconstruct specific residual stream directions.
These findings reveal a mechanistic role for dense latents in language model behavior and suggest avenues for refining SAE training strategies.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Presenter: ~Alessandro_Stolfo1

Format: Yes, the presenting author will definitely attend in person because they attending ICLR for other complementary reasons.

Funding: No, the presenting author of this submission does *not* fall under ICLR’s funding aims, or has sufficient alternate funding.

Submission Number: 46