method
active
method:top-16-features-per-top-10-logits-selectionTop-16 features per top-10 logits selection
Yields 160 features of interest from a single prompt.
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- Top-16 features for each of top-10 logits yields 160 features of interest pre-deduplicationfinding0.863Analysis set size.
- Abstract parity feature.
- Greater-than feature.
- Explains nested ≈ half-capacity turn-averaged SAE.
- Open question.
- Measure of superposition in toy models.
- what is the 'correct number of features' for dictionary learning, and is this question well-posed?question0.708Open question about whether there is a true discrete feature count or a continuous splitting process
- More general greater-than feature.