method
active
method:sparse-autoencoder-sae-featuresSparse autoencoder (SAE) features
Interpretability tool underlying top-feature, details, steering tools.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Building and evaluating alignment auditing agentscitesmentions
Methods (3)
method
- Sparse Autoencoders (SAE)related_toInterpretability method criticized in this paper for shattering manifolds into atomic pieces, obscuring overarching semantic structure.
- TopK Sparse Autoencoders (SAEs)related_toSparse dictionary learning method used to extract interpretable features from EEG transformer embeddings.
- Tool-using agent for open-ended model investigations.
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.
- Used in Anthropic welfare assessment to identify performative behavior and hidden emotional struggle co-activating features
- The primary novel framework introduced in the paper for learning facet-level personality control vectors
- Interpretability framework used to decompose layer-40 activations into sparse feature sets for studying emotional alignment and persistence
- Standard interpretability approach that VPD critiques and proposes an alternative to.
- The individual, supposedly monosemantic directions learned by SAEs; argued here to fragment manifolds into disconnected pieces.
- The specific SAE architecture trained: 100K+ possible features compressed to 64 active per token for layer-40 activations
- Sparse Autoencoders Find Highly Interpretable Features in Language Models (Cunningham et al., 2023)concept0.820Core methodology paper for SAE-based interpretable feature extraction
- Central claim of the paper: the method scales to state-of-the-art transformers.