method
active
method:feature-attribution-via-gradient-dot-product-with-sae-decoderFeature attribution via gradient dot product with SAE decoder
Computing attribution as the dot product of the output logit gradient with the SAE decoder weight, multiplied by feature activation.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (1)
method
- Backprop of logit-minus-mean-logit loss to surface causal intervention candidates, summed across token positions.
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.
- Baseline method against which probe-based ranking is compared; more computationally expensive.
- A promising property for interpretability analysis off-distribution.
- Out-of-distribution generalization of SAE features.
- Third checkpointing strategy; adopted here.
- Interpretability tool underlying top-feature, details, steering tools.
- Used by APD; poor approximations of causal importance.
- Interpretability method criticized in this paper for shattering manifolds into atomic pieces, obscuring overarching semantic structure.
- Discrete case.