finding
active
finding:adversarial-training-reduced-superposition-only-with-attacks-of-80-input-l2-normAdversarial training reduced superposition only with attacks of 80% input L2 norm
Unsatisfying reduction.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Toy Models of Superpositionmentions
Claims (1)
claim
- Superposition 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.
- Novelty claim.
- Robustness controls superposition.
- Adversarially robust models have different interpretability because adversarial training affects superpositionhypothesis0.833Could be tested by crosscoder comparisons.
- Interpretation of Figure 3.
- Adversarial vulnerability increases >3x as superposition forms and tracks features per dimensionfinding0.805L2 attacks max norm 0.1 of average input norm.
- Adversarial example transfer may be predicted by correlated arrangement of features in superpositionhypothesis0.787Speculative.
- Reduces superposition but needs 80% L2 norm.
- Trains toy models for 150,000 steps with L2 ε=0.1‖x‖.