question
active
question:why-do-logit-attribution-heatmaps-diverge-from-max-activating-examplesWhy do logit attribution heatmaps diverge from max-activating examples?
Unclear reason stated by authors.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
- Replicating Circuit Tracing for a Simple Known Mechanism - Goodfireassociated_withmentions
Findings (1)
finding
Claims (1)
claim
- Speculative explanation.
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.
- Input-output mapping per feature across ~100 prompts.
- Logit-minus-mean-logit loss was most effective for attribution, vs log-loss which was confusing to interpretfinding0.760Attribution loss selection.
- Feature 425104 heatmap promotes outputs above ~60 with little input dependence, e.g. promotes 62 given 90finding0.755Diverges from Hanna et al. heatmaps.
- Headline result (Figure 1).
- Design choice validated empirically.
- Abstract parity feature.
- Structured number feature.
- Caveat.