finding
active
finding:crosscoders-give-a-dramatically-better-whole-model-mse-l0-tradeoff-than-summed-per-layer-saesCrosscoders give a dramatically better whole-model (MSE, L0) tradeoff than summed per-layer SAEs
Consolidating shared structure gives concise decomposition.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Claims (1)
claim
- Due to SAE training stochasticity.
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.
- Indicates redundant structure across layers.
- Crosscoders are about 2x less FLOP-efficient than per-layer SAEs at equal eval loss at large computefinding0.835Cost of identifying cross-layer structure.
- Motivating example with 13 vs 3 features.
- Decoder cost dominates.
- Motivation for DFC.
- Circuit simplification theory.
- Core motivation.
- Discussion.