method
active
method:sae-fidelity-metricSAE fidelity metric
Log-loss recovery ratio between SAE-inserted and ablated model.
Neighborhood — ranked by edge-count
Papers (1)
paper
Artifacts (1)
artifact
- ReLU SAE trained on residual stream layer 19 with MSE+L1 loss.
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.
- Open issue from Pythia vs Gemma.
- Design philosophy.
- Persistence metric for SAE features: P(fires at t+100 | fired at t) minus P(fires at t+100 | did not fire at t)
- Standard baseline SAE operating on individual token activations.
- The individual, supposedly monosemantic directions learned by SAEs; argued here to fragment manifolds into disconnected pieces.
- Internal concepts a model uses, extracted as human-interpretable components.
- Fraction of an SAE feature's length lying inside the 171-dimensional subspace spanned by emotion probes, computed via SVD orthogonalization
- Backprop of logit-minus-mean-logit loss to surface causal intervention candidates, summed across token positions.