finding
active
finding:point-based-landmarks-tiling-a-joint-manifold-lie-at-distance-from-factor-manifold-m1-lemma-1Point-based landmarks tiling a joint manifold lie at distance ≥ δ−ε from factor manifold M1 (Lemma 1)
Formal result: point-based SAEs cannot recover factor manifolds under superposition.
Source paper
extracted_from(2026) · Usha Bhalla · Thomas Fel · Can Rager · Sheridan Feucht +8
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Justifies focusing on direction-based SAEs.
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 L0hypothesis0.761Speculative explanation.
- Many latents tiling a low-dimensional manifold would show decoder cosine similarities near 1hypothesis0.760Motivates Appendix E.
- Appendix claim.
- Evidence that the weekday cyclic structure is not anomalous but reflects broader principle of concept geometry.
- Cross-modality result from the full paper demonstrating that representation-behavior geometry alignment is not limited to language models.
- Mere correlation of feature directions, as used in prior work, need not suffice to find manifoldsclaim0.740Contrast with Engels et al.
- Future-work perspective.
- Latents tile a manifold to reduce loss.