claim
active
claim:standard-crosscoders-have-an-inherent-optimization-prior-favoring-shared-features-over-exclusive-onesStandard crosscoders have an inherent optimization prior favoring shared features over exclusive ones.
Motivation for DFC.
Source paper
extracted_from(2026) · Thomas Jiralerspong · Trenton Bricken
Neighborhood — ranked by edge-count
Papers (2)
paper
Findings (1)
finding
- Footnote 2.
Claims (2)
claim
- Main architectural claim.
- Interpretation of toy dynamics.
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.
- Preliminary observation.
- Crosscoders are about 2x less FLOP-efficient than per-layer SAEs at equal eval loss at large computefinding0.826Cost of identifying cross-layer structure.
- Figure 5.
- Decoder norm profile across 18 layers.
- Compared with base SAE transfer.
- Cross-layer transcoder features form a mix of local, global and intermediate features by decoder normfinding0.801Preliminary decoder norm analysis.
- Appendix C.4.
- Discussion.