concept
active
concept:feature-splittingFeature splitting
Phenomenon where a feature in a small SAE splits into multiple finer features in a larger SAE.
Neighborhood — ranked by edge-count
Papers (8)
paper
Claims (2)
claim
- Proposition 4.2.
- Assertion about the qualitative advantages of VPD's rank-one decomposition.
Methods (2)
method
- Supernodesassociated_withManual grouping of related features to simplify attribution graphs.
- UMAP Embedding of Featuressupports2D embedding of feature direction vectors used to visualize feature clusters and splitting geometry
Concepts (4)
concept
- Feature splitting and absorptionrelated_toProblems suggesting sparsity breaks down as interpretability proxy.
- Feature Absorptionassociated_withPhenomenon where sparsity yields gerrymandered latents for hierarchical concepts.
- monosemanticityassociated_withInterpretability property where a latent feature represents a single semantic concept; benchmarked across architectures.
- Single-Token Featuresassociated_withFeatures that fire on every instance of a single token; appear in small dictionaries as collapsed versions of many token-in-context features
Hypotheses (1)
hypothesis
- Including chat data in SAE training merges RM bias features with generic Assistant featuresassociated_withExplains poor features with three-dataset SAE.
Findings (1)
finding
- Concrete example of feature splitting revealing unexpected model structure
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.
- Explains lower off-target effects of SAE methods.
- Authors argue the absence of a fixed feature count is a property of the superposition geometry, not a failure of the method
- Observed across SAE scales, e.g., 'San Francisco' split into 11 features.
- Independent concepts are encoded by separate latents.
- Features with similar decoder magnitude across both models.
- Property that features activate on only a small fraction of inputs; enables compressed sensing and is what allows superposition to work
- Synthesizing inputs maximizing component activation.
- The subdivision of properties to create smaller, individually owned lots that support unique buildings and increased density.