claim
active
claim:clts-discover-interpretable-and-relevant-features-for-greater-than-many-recognizably-similar-to-those-of-hanna-et-alCLTs discover interpretable and relevant features for greater-than, many recognizably similar to those of Hanna et al.
Main conclusion.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
Findings (7)
finding
- More general greater-than feature.
- Greater-than feature.
- More complex than simple comparison.
- Narrow-range greater-than feature.
- Structure of the computational pathway.
- Best CLT on GPT-2 Small achieves L0 of 79 and 59% token accuracy, consistent with Ameisen et al.supportsTraining quality of the CLT.
- Training configuration.
Claims (1)
claim
- Caveat.
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.
- Difference between sparse and neuron analyses.
- Key advantage statement.
- Closing claim.
- CLTs uncover relevant structure out of the box without labor-intensive experimental interventionsclaim0.783Advantage over original analysis.
- Inferred from Feature 461858.
- Comparison of heatmaps.
- Interpretation.
- Question about completeness of feature-based model explanation