finding
active
finding:in-toy-models-adversarial-vulnerability-increases-with-feature-sparsity-and-superposition-with-a-dip-at-antipodal-superpositionIn toy models, adversarial vulnerability increases with feature sparsity and superposition, with a dip at antipodal superposition.
Superposition controls robustness.
Source paper
extracted_from(2025) · Gorton, Liv · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Central thesis.
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.
- Robustness controls superposition.
- Adversarial vulnerability increases >3x as superposition forms and tracks features per dimensionfinding0.830L2 attacks max norm 0.1 of average input norm.
- Figure 4.
- Appendix replication.
- Adversarially robust models have different interpretability because adversarial training affects superpositionhypothesis0.815Could be tested by crosscoder comparisons.
- Unexpected result.
- Confound addressed.
- Core demonstration.