finding
active
finding:adversarial-vulnerability-increases-3x-as-superposition-forms-and-tracks-features-per-dimensionAdversarial vulnerability increases >3x as superposition forms and tracks features per dimension
L2 attacks max norm 0.1 of average input norm.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Toy Models of Superpositionmentions
Claims (1)
claim
- Superposition may be a significant contributor to adversarial examplesassociated_withsupportsSuperposition maximalist stance.
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.
- Superposition controls robustness.
- Appendix replication.
- Unsatisfying reduction.
- Novelty claim.
- Robustness controls superposition.
- Adversarially robust models have different interpretability because adversarial training affects superpositionhypothesis0.795Could be tested by crosscoder comparisons.
- Interpretation of Figure 3.
- Adversarial example transfer may be predicted by correlated arrangement of features in superpositionhypothesis0.779Speculative.