finding
active
finding:most-saes-learn-years-features-selective-to-the-ones-digit-period-10-years-alongside-decade-encoding-featuresMost SAEs learn years features selective to the ones digit (period 10 years) alongside decade-encoding features
Specific tuning curve structure.
Source paper
extracted_from(2026) · Usha Bhalla · Thomas Fel · Can Rager · Sheridan Feucht +8
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Central empirical interpretation.
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.
- Discrete case.
- Automated interpretability and specificity ratings show SAE features are clearer than MLP neurons.
- SAEs assume sparsity (number of concepts) is time-invariant, drawn i.i.d. from a fixed distribution.claim0.765Corollary 4.1.1.
- Extension of mechanistic interpretability findings to the metacognitive domain
- Claim that feature grounding enables interpretability metrics.
- Proposition 4.1 central claim.
- Out-of-distribution generalization of SAE features.
- Distillation of why learning generalises.