claim
active
claim:dfcs-outperform-standard-crosscoders-at-isolating-model-exclusive-featuresDFCs outperform standard crosscoders at isolating model-exclusive features.
Main architectural claim.
Source paper
extracted_from(2026) · Thomas Jiralerspong · Trenton Bricken
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (8)
finding
- On the toy model (2048 concepts, 2.5% exclusive per model), DFC achieves higher exclusive concept recovery than standard and DSF crosscoders, especially undercomplete.associated_withsupportsFigure 4 left.
- Figure 6.
- Validation step 1.
- ChatGPT identity, copyright refusal and Deepseek CCP features all achieved exclusivity score 5.supports§3.3.2.
- Table 1.
- Figure 5.
- Appendix C.4.
- Qualitative comparison.
Claims (2)
claim
- Motivation for DFC.
- Model-exclusive refers to representation, not conceptual capability; Llama still knows the CCP concept.associated_withScope limitation.
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.
- Figure 15.
- Architectural variant partitioning the dictionary into A-exclusive, B-exclusive and shared sets.
- Indicates redundant structure across layers.
- Footnote 2.
- Crosscoder model diffing is best used as a high-recall pre-screening tool followed by verification.claim0.768Recommended workflow.
- Preliminary observation.
- Crosscoders give a dramatically better whole-model (MSE, L0) tradeoff than summed per-layer SAEsfinding0.765Consolidating shared structure gives concise decomposition.
- Decoder norm profile across 18 layers.