concept
active
concept:sae-latents

SAE Latents

Interpretable features extracted by sparse autoencoders used as steering targets in this study

Neighborhood — ranked by edge-count

Concepts (1)

concept

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

  • Using GPT-4o or o3 to automatically generate interpretations of SAE latents from top-activating examples
  • Adding a multiple of the SAE latent decoder vector to token activations to causally test each latent's role in misalignment
  • Reduces SAE optimization to allocating N latents across features.
  • Pre-filtering step excluding abstract latents where off-topic detection is harder
  • Interpretability method criticized in this paper for shattering manifolds into atomic pieces, obscuring overarching semantic structure.
  • Per-Token SAEframework0.781
    Standard baseline SAE operating on individual token activations.
  • SAE featuresconcept0.776
    The individual, supposedly monosemantic directions learned by SAEs; argued here to fragment manifolds into disconnected pieces.
  • Internal concepts a model uses, extracted as human-interpretable components.