claim
active
claim:designating-a-small-set-of-shared-features-with-reduced-sparsity-penalty-renders-exclusive-features-more-interpretable-and-monosemanticDesignating a small set of shared features with reduced sparsity penalty renders exclusive features more interpretable and monosemantic
Main proposal.
Source paper
extracted_fromSiddharth Mishra-Sharma · Trenton Bricken · Jack Lindsey · Adam Jermyn +4
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (3)
finding
- Mitigation isolates expected sleeper agent behavior.
- Mitigation works on real models.
- E.g. 'Here is a draft', apology, and 'explain like I'm in sixth grade' features.
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.
- Proposed variation designating ~10k of ~250k features as shared with penalty 0.1-0.2x baseline to soak up shared variance.
- Mechanistic argument.
- Better bases may give sparsity.
- Third empirical pattern.
- Central interpretation.
- Major prediction checked against InceptionV1.
- Discovery story.
- Central claim of the paper, supported by detailed feature analysis, human evaluation, automated interpretability of activations, and automated interpretability of logit weights