claim
active
claim:superposition-may-be-a-significant-contributor-to-adversarial-examplesSuperposition may be a significant contributor to adversarial examples
Superposition maximalist stance.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Toy Models of Superpositionmentions
Findings (2)
finding
- Adversarial vulnerability increases >3x as superposition forms and tracks features per dimensionassociated_withsupportsL2 attacks max norm 0.1 of average input norm.
- Unsatisfying reduction.
Hypotheses (1)
hypothesis
- Adversarially robust models would have worse performance and be more interpretable because robustness requires giving up superpositionassociated_withPredictions of superposition maximalism.
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.
- Superposition may be a major contributing factor, or even the primary cause, of adversarial examples.hypothesis0.942Hypothesis under test.
- Adversarial example transfer may be predicted by correlated arrangement of features in superpositionhypothesis0.886Speculative.
- Supported by Table 1.
- Adversarially robust models have different interpretability because adversarial training affects superpositionhypothesis0.809Could be tested by crosscoder comparisons.
- Novelty claim.
- Core mechanism statement.
- superposition is deeply linked to the ability to identify and enumerate over all features in a modelquote0.800Safety rationale.
- Linked to superposition interference.
Restated by (1)
cosine ≥ 0.90Other entities that say roughly the same thing. May be merge candidates or independent restatements across papers.