method
active
method:optimal-transport-comparison-of-sae-featuresOptimal transport comparison of SAE features
Compares features across SAEs.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Do Sparse Autoencoders Capture Concept Manifolds?citesmentions
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.
- Optimal transport shows weak alignment of SAE decoder directions and random-data codes across architecturesfinding0.840Individual features are unstable.
- 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.
- Backprop of logit-minus-mean-logit loss to surface causal intervention candidates, summed across token positions.
- Surprising finding that the two evaluation methods diverge in their relationship with persistence
- Interpretability tool underlying top-feature, details, steering tools.
- Prediction following diversity-undertraining logic.
- Method where Kimi evaluates steered vs unsteered text samples from another instance to rate SAE feature emotionality (0-100)