method
active
method:years-by-years-logit-attribution-heatmapsYears-by-years logit attribution heatmaps
Input-output mapping per feature across ~100 prompts.
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.
- Logit attribution heatmaps and max-activating examples disagree for Features 425104 and 349410finding0.810Limitation.
- Dot product of summed decoder vector with unembedding vector, following Hanna et al.
- Unclear reason stated by authors.
- Speculative explanation.
- Logit-minus-mean-logit loss was most effective for attribution, vs log-loss which was confusing to interpretfinding0.750Attribution loss selection.
- Cross-model generalization.
- Baseline method against which probe-based ranking is compared; more computationally expensive.
- Some gradient-based attribution methods give attributions independent of model and data-generating processfinding0.735Adebayo et al. sanity checks.