finding
active
finding:at-l0-100-sparsity-fidelity-ranking-is-batchtopk-topk-jump-gated-matryoshka-p-anneal-relu-differences-shrink-at-l0-100-500At L0<100 sparsity-fidelity ranking is BatchTopK > TopK > Jump > Gated > Matryoshka > p-anneal > ReLU; differences shrink at L0 100-500
Frontier ranking differs from other metrics.
Source paper
extracted_from(2025) · Adam Karvonen · Can Rager · Johnny Lin · Curt Tigges +11
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
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.
- Central result.
- Baseline.
- Traditional proxy evaluation criterion for SAEs.
- ReLU hidden layer model (n=10, m=5) shifts from monosemantic to polysemantic neurons as sparsity increasesfinding0.764Neuron-level phase change.
- Alternative sparsity notion proposed for predictive component.
- Comprehensive model comparison showing tuning benefit for persona fidelity
- Inverse scaling.
- Reconstruction cannot rank featurizers.