finding
active
finding:nested-sae-s-top-128-features-are-99-inner-turn-level-featuresNested SAE's top-128 features are ~99% inner (turn-level) features
Explains nested ≈ half-capacity turn-averaged SAE.
Source paper
extracted_from(2026) · Kevin Der · Harish Kamath · Thompson, Ben
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Nested sum 70.6% and nested max 57.3% on 10-way matchingassociated_withFigure 1.
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.
- Table 1.
- The specific SAE architecture trained: 100K+ possible features compressed to 64 active per token for layer-40 activations
- Fig 2 right.
- Most features dead in largest SAE, indicating room for improvement.
- Figure 7.
- Moderation statistics.
- Hybrid SAE jointly training turn-averaged inner and per-token outer partitions.
- Top SAE features on the years manifold show localized smooth tuning curves with overlapping supportfinding0.746Reminiscent of population coding.