method
active
method:nearest-neighbour-feature-discrimination-testNearest-neighbour feature discrimination test
Stringent test distinguishing a feature from its 10 nearest embedding neighbours with one true and three distractor examples, repeated ten times.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Understanding and Steering Llama 3 with Sparse Autoencoders - Goodfireintroducesmentions
Methods (1)
method
- Claude-based explanation of features from max-activating examples.
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.
- Retrieves similar features by decoder cosine similarity.
- Detects tiling in real SAEs.
- Primary alignment metric used in experiments; measures mean intersection of k-nearest neighbor sets between two kernels
- Property that features activate on only a small fraction of inputs; enables compressed sensing and is what allows superposition to work
- Gemma Scope decoder nearest-neighbor cosine similarities skewed above random but few extremely highfinding0.720Fig 7.
- Identifying related features by cosine distance in SAE decoder space.
- Method to select d_steer top-activated SAE features for constructing control vectors
- Proposed variation designating ~10k of ~250k features as shared with penalty 0.1-0.2x baseline to soak up shared variance.