hypothesis
active
hypothesis:many-latents-tiling-a-low-dimensional-manifold-would-show-decoder-cosine-similarities-near-1Many latents tiling a low-dimensional manifold would show decoder cosine similarities near 1
Motivates Appendix E.
Source paper
extracted_from(2025) · Michaud, Eric J. · Gorton, Liv · McGrath, Tom
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Gemma Scope decoder nearest-neighbor cosine similarities skewed above random but few extremely highassociated_withsupportsFig 7.
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.
- Latents near digit/alphabet manifolds may have tiled a manifold with one killed to reduce L0hypothesis0.827Speculative explanation.
- Third empirical pattern.
- Detects tiling in real SAEs.
- Training artifact.
- Same concept and function in both models mostly.
- Identifying related features by cosine distance in SAE decoder space.
- Point-based landmarks tiling a joint manifold lie at distance ≥ δ−ε from factor manifold M1 (Lemma 1)finding0.760Formal result: point-based SAEs cannot recover factor manifolds under superposition.
- Experiment 4 result showing DIM captures only one facet of the multi-dimensional truth subspace