claim
active
claim:sae-features-generalize-to-images-despite-training-only-on-text-indicating-out-of-distribution-robustnessSAE features generalize to images despite training only on text, indicating out-of-distribution robustness.
A promising property for interpretability analysis off-distribution.
Source paper
extracted_from(2024) ·
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.
- Out-of-distribution generalization of SAE features.
- Conclusion.
- Extension of mechanistic interpretability findings to the metacognitive domain
- Backprop of logit-minus-mean-logit loss to surface causal intervention candidates, summed across token positions.
- App. B.
- Computing attribution as the dot product of the output logit gradient with the SAE decoder weight, multiplied by feature activation.
- Surprising finding that the two evaluation methods diverge in their relationship with persistence
- Prediction following diversity-undertraining logic.