finding
active
finding:sae-features-trained-on-text-activations-generalize-to-image-inputs-activating-on-relevant-visual-depictionsSAE features trained on text activations generalize to image inputs, activating on relevant visual depictions.
Out-of-distribution generalization of SAE features.
Source paper
extracted_from(2024) ·
Neighborhood — ranked by edge-count
Claims (1)
claim
- Features respond to concepts across languages and in images, not just text.
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.
- A promising property for interpretability analysis off-distribution.
- Conclusion.
- Standard interpretability approach that VPD critiques and proposes an alternative to.
- Extension of mechanistic interpretability findings to the metacognitive domain
- Feature cluster examples from the UMAP.
- Chat-trained SAEs find behaviourally relevant features because many causal features occur on chat-specific special tokenshypothesis0.800Explanation for LMSYS superiority.
- Computing attribution as the dot product of the output logit gradient with the SAE decoder weight, multiplied by feature activation.
- Backprop of logit-minus-mean-logit loss to surface causal intervention candidates, summed across token positions.