finding
active
finding:features-with-highest-virtual-weights-to-a-years-feature-are-all-about-yearsFeatures with highest virtual weights to a years feature are all about years
Contrast with answer selection.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Supports monosemantic reading.
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 to project features through QK weights and find aligned features.
- Naive weight magnitude works here.
- Large virtual weights are not inherently misleading but not guaranteed to be helpful or effectiveclaim0.761Caution for practitioners.
- Implicit weights directly connecting any pair of layers computed by multiplying output weights of one layer with input weights of another through the residual stream
- Second of three speculative claims asserting that subgraphs of neural networks are tractable and meaningful objects of study
- what metrics can reliably tell us if dictionary learning has successfully extracted high quality features?question0.736Central methodological gap: current metrics (loss, density histograms, manual inspection) are inadequate
- More general greater-than feature.
- Interpretive claim that circuits render raw weights interpretable as algorithmic structures