finding
active
finding:relu-saes-reduce-loss-slightly-by-tiling-circle-s-1-with-4-8-24-latentsReLU SAEs reduce loss slightly by tiling circle S^1 with 4, 8, 24 latents
Fig 1.
Source paper
extracted_from(2025) · Michaud, Eric J. · Gorton, Liv · McGrath, Tom
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Core concern.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- ReLU underperforms.
- Exception to ReLU underperformance.
- Evidence in real models.
- Training artifact.
- Fig 2 right.
- Lower SAE loss at fixed sparsity likely indicates less interference and less superposition.claim0.745Justifies SAE proxy.
- Strongest argument against pathology.
- Unlearning not discriminative.