concept
active
concept:sae-featuresSAE features
The individual, supposedly monosemantic directions learned by SAEs; argued here to fragment manifolds into disconnected pieces.
Neighborhood — ranked by edge-count
Claims (2)
claim
- Core critique of sparse autoencoders: they break the geometric structure of representations, making it harder to see the big picture.
- Positive claim that geometric descriptions retain the conceptual coherence lost in atomized feature decompositions.
Concepts (2)
concept
- Features (SAE features)same_asInternal concepts a model uses, extracted as human-interpretable components.
- SAE feature splittingrelated_toExplains lower off-target effects of SAE methods.
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.
- Method of adding scaled versions of sparse autoencoder latent features during generation to causally modulate model behavior
- Interpretability tool underlying top-feature, details, steering tools.
- Explains weaker datamix SAEs.
- Compares features across SAEs.
- Method where Kimi evaluates steered vs unsteered text samples from another instance to rate SAE feature emotionality (0-100)
- Interpretable features extracted by sparse autoencoders used as steering targets in this study
- Backprop of logit-minus-mean-logit loss to surface causal intervention candidates, summed across token positions.
- The specific SAE architecture trained: 100K+ possible features compressed to 64 active per token for layer-40 activations