hypothesis
active
hypothesis:more-sophisticated-data-mixes-will-perform-better-as-sae-training-scalesMore sophisticated data mixes will perform better as SAE training scales
Prediction following diversity-undertraining logic.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Diverse data-mix SAEs had far more dead and very low-frequency features and were less suitable for intervention than LMSYS SAEassociated_withQualitative and quantitative observation.
Hypotheses (1)
hypothesis
- Explains weaker datamix SAEs.
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.
- SAE latents depend on training data.
- App. B.
- A promising property for interpretability analysis off-distribution.
- RM features at rank 181, outside top 100.
- Authors' experience-based claim.
- Supervised metrics discriminate.
- Feature cluster examples from the UMAP.
- Design philosophy.