claim
active
claim:saes-can-capture-manifolds-either-globally-via-a-compact-atom-group-spanning-the-manifold-or-locally-via-features-tiling-regionsSAEs can capture manifolds either globally via a compact atom group spanning the manifold or locally via features tiling regions
Theoretical framework answer.
Source paper
extracted_from(2026) · Usha Bhalla · Thomas Fel · Can Rager · Sheridan Feucht +8
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (4)
finding
- Increasing TopK sparsity drives synthetic SAEs through shattering, capture, then dilution regimesassociated_withsupportsPhase diagram on support size and receptive field spread.
- Restricted-R² variance explained on Llama3.1-8B plateaus well beyond each manifold's ambient dimension across five SAE architecturesassociated_withSAEs do not achieve compact capture; they are in dilution.
- Formal subspace recovery guarantee.
- Subspace capture on synthetic manifolds peaks near k=4 and degrades at lower and higher sparsitysupportsSparsity sweet spot in synthetic benchmark (R^128, c=512).
Artifacts (1)
artifact
- Code repository for the paper.
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.
- Central research question.
- Core concern.
- Justifies focusing on direction-based SAEs.
- Related to point-based simplicial limit.
- Core critique of sparse autoencoders: they break the geometric structure of representations, making it harder to see the big picture.
- Features act as local patch detectors.
- Restricted to specific manifolds, different SAEs show strong alignment of induced coordinate systemsfinding0.788Geometry is shared even if features are not.
- Manifold-level descriptions recover overarching semantic structure that SAE features miss.claim0.783Positive claim that geometric descriptions retain the conceptual coherence lost in atomized feature decompositions.