question
active
question:why-does-robustness-improve-temporarily-near-antipodal-superposition-and-why-do-models-with-equal-superposition-but-different-structure-vary-in-robustnessWhy does robustness improve temporarily near antipodal superposition, and why do models with equal superposition but different structure vary in robustness?
Future work.
Source paper
extracted_from(2025) · Gorton, Liv · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Explains robust models being more robust than expected.
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.
- Adversarially robust models would have worse performance and be more interpretable because robustness requires giving up superpositionhypothesis0.814Predictions of superposition maximalism.
- Adversarially robust models have different interpretability because adversarial training affects superpositionhypothesis0.814Could be tested by crosscoder comparisons.
- Figure 4.
- Posed on SAE results.
- Discussion implication.
- Robustness influences superposition, predicting lower SAE loss in more robust models.hypothesis0.790Popperian credit for surprising prediction.
- Unexpected result.
- Superposition controls robustness.