finding
active
finding:allocating-10k-of-250k-features-as-shared-with-0-1-0-2x-sparsity-penalty-makes-exclusive-feature-densities-similar-to-shared-ones-while-remaining-symmetricAllocating 10k of ~250k features as shared with 0.1-0.2x sparsity penalty makes exclusive feature densities similar to shared ones, while remaining symmetric
Mitigation works on real models.
Source paper
extracted_fromSiddharth Mishra-Sharma · Trenton Bricken · Jack Lindsey · Adam Jermyn +4
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Main proposal.
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.
- Toy model in the abundant-capacity regime.
- Third empirical pattern.
- Central interpretation.
- Figure 5.
- Same concept and function in both models mostly.
- The specific SAE architecture trained: 100K+ possible features compressed to 64 active per token for layer-40 activations