claim
active
claim:adversarial-attacks-work-by-activating-many-interfering-features-in-superposition-adversarial-training-reorganizes-geometry-to-prevent-thisAdversarial attacks work by activating many interfering features in superposition; adversarial training reorganizes geometry to prevent this.
Interpretation of Figure 3.
Source paper
extracted_from(2025) · Gorton, Liv · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Figure 3B.
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.
- Novelty claim.
- Unsatisfying reduction.
- Robustness controls superposition.
- If adversarial attacks exploit interference, they should activate more features than clean inputs.hypothesis0.820Prediction tested via L0.
- Adversarially robust models have different interpretability because adversarial training affects superpositionhypothesis0.818Could be tested by crosscoder comparisons.
- Adversarial example transfer may be predicted by correlated arrangement of features in superpositionhypothesis0.802Speculative.
- Adversarial vulnerability increases >3x as superposition forms and tracks features per dimensionfinding0.790L2 attacks max norm 0.1 of average input norm.
- Superposition maximalist stance.