question
active
question:why-can-crosscoders-struggle-to-isolate-features-in-base-vs-finetune-settingsWhy can crosscoders struggle to isolate features in base vs finetune settings?
Proposed future work.
Source paper
extracted_from(2026) · Thomas Jiralerspong · Trenton Bricken
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Preliminary observation.
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.
- Near-identical numbers of exclusive features are found for both models in base vs finetuned diffsfinding0.748Symmetry of exclusive features across models.
- A priori unknown behavior discovered.
- Cross-layer transcoder features form a mix of local, global and intermediate features by decoder normfinding0.744Preliminary decoder norm analysis.
- Circuit simplification theory.
- Decoder norm profile across 18 layers.
- Decomposition into shared and model-specific features.