claim
active
claim:point-based-saes-tile-the-joint-manifold-as-a-single-object-and-cannot-recover-factor-manifolds-under-superpositionPoint-based SAEs tile the joint manifold as a single object and cannot recover factor manifolds under superposition
Justifies focusing on direction-based SAEs.
Source paper
extracted_from(2026) · Usha Bhalla · Thomas Fel · Can Rager · Sheridan Feucht +8
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Formal result: point-based SAEs cannot recover factor manifolds under superposition.
Claims (1)
claim
- Related to point-based simplicial limit.
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.
- Theoretical framework answer.
- Core concern.
- Features act as local patch detectors.
- Central research question.
- Restricted to specific manifolds, different SAEs show strong alignment of induced coordinate systemsfinding0.782Geometry is shared even if features are not.
- Manifold-level descriptions recover overarching semantic structure that SAE features miss.claim0.771Positive claim that geometric descriptions retain the conceptual coherence lost in atomized feature decompositions.
- Authors' hedged interpretation.
- Top SAE features on the years manifold show localized smooth tuning curves with overlapping supportfinding0.754Reminiscent of population coding.