hypothesis
active
hypothesis:adversarially-robust-models-have-different-interpretability-because-adversarial-training-affects-superpositionAdversarially robust models have different interpretability because adversarial training affects superposition
Could be tested by crosscoder comparisons.
Source paper
extracted_fromRelated 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.
- Adversarially robust models would have worse performance and be more interpretable because robustness requires giving up superpositionhypothesis0.905Predictions of superposition maximalism.
- Novelty claim.
- Robustness controls superposition.
- Unsatisfying reduction.
- Figure 4.
- Interpretation of Figure 3.
- Superposition controls robustness.
- Future work.