hypothesis
active
hypothesis:hard-to-steer-features-are-split-across-layers-spanning-the-sae-point-so-only-a-portion-is-intervened-uponHard-to-steer features are split across layers spanning the SAE point so only a portion is intervened upon
Cross-layer superposition conjecture.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Some features are easy to intervene on while very similar-looking ones have little or no effectsupportsObservation from intervention phenomenology.
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.
- Observed across SAE scales, e.g., 'San Francisco' split into 11 features.
- Clamping feature activations causally alters model behavior in interpretable ways.
- Core critique of sparse autoencoders: they break the geometric structure of representations, making it harder to see the big picture.
- Limitations section.
- Explains weaker datamix SAEs.
- Authors' experience-based claim.
- Surprising finding that the two evaluation methods diverge in their relationship with persistence
- Explains lower off-target effects of SAE methods.