hypothesis
active
hypothesis:crosscoders-struggle-on-base-vs-finetune-comparisons-due-to-mirror-featuresCrosscoders struggle on base vs finetune comparisons due to mirror features.
Preliminary observation.
Source paper
extracted_from(2026) · Thomas Jiralerspong · Trenton Bricken
Neighborhood — ranked by edge-count
Papers (1)
paper
Questions (1)
question
- Proposed future work.
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.
- Motivation for DFC.
- Compared with base SAE transfer.
- Decoder norm profile across 18 layers.
- Decomposition into shared and model-specific features.
- Near-identical numbers of exclusive features are found for both models in base vs finetuned diffsfinding0.780Symmetry of exclusive features across models.
- Appendix C.4.
- Cross-layer transcoder features form a mix of local, global and intermediate features by decoder normfinding0.779Preliminary decoder norm analysis.
- Crosscoders are about 2x less FLOP-efficient than per-layer SAEs at equal eval loss at large computefinding0.775Cost of identifying cross-layer structure.