hypothesis
active
hypothesis:sae-features-become-interpretable-before-they-become-suitable-for-intervening-uponSAE features become interpretable before they become suitable for intervening upon
Explains weaker datamix SAEs.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Qualitative and quantitative observation.
Hypotheses (1)
hypothesis
- Prediction following diversity-undertraining logic.
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.
- Authors' hedged position on SAEs.
- Automated interpretability and specificity ratings show SAE features are clearer than MLP neurons.
- Surprising finding that the two evaluation methods diverge in their relationship with persistence
- Internal concepts a model uses, extracted as human-interpretable components.
- Policy recommendation based on latent #10 activating at 5% incorrect data before behavioral misalignment emerges
- Extension of mechanistic interpretability findings to the metacognitive domain
- Out-of-distribution generalization of SAE features.
- The individual, supposedly monosemantic directions learned by SAEs; argued here to fragment manifolds into disconnected pieces.