claim
active
claim:the-sparsity-fidelity-curve-is-necessary-but-not-sufficient-for-evaluating-saesThe sparsity-fidelity curve is necessary but not sufficient for evaluating SAEs
Motivates multi-metric evaluation.
Source paper
extracted_from(2025) · Adam Karvonen · Can Rager · Johnny Lin · Curt Tigges +11
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Central result.
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.
- SAEs assume sparsity (number of concepts) is time-invariant, drawn i.i.d. from a fixed distribution.claim0.804Corollary 4.1.1.
- Lower SAE loss at fixed sparsity likely indicates less interference and less superposition.claim0.798Justifies SAE proxy.
- Better bases may give sparsity.
- Open question.
- Traditional proxy evaluation criterion for SAEs.
- Proposition 4.2.
- Robustness influences superposition, predicting lower SAE loss in more robust models.hypothesis0.768Popperian credit for surprising prediction.
- Authors' experience-based claim.