finding
active
finding:optimal-transport-shows-weak-alignment-of-sae-decoder-directions-and-random-data-codes-across-architecturesOptimal transport shows weak alignment of SAE decoder directions and random-data codes across architectures
Individual features are unstable.
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
- Explanation of SAE instability.
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.
- Compares features across SAEs.
- Argues against the single-layer analysis approach of prior work.
- Features not merely relayed via residual connections.
- The objective function combining L2 reconstruction error and L1 penalty scaled by decoder norm, used to train the SAE.
- Core concern.
- Quantifies extreme redundancy in transformer routing; supports claim that introspection and interference patterns are architecturally permitted.
- Preference steering effective only in mid-layer range though correlations hold across layersfinding0.749Suggests mid-layer reading circuitry.
- Authors' interpretation of prompt variation results showing alignment faking disappears only when conflicting objective is removed