claim
active
claim:adversarial-training-reduces-superposition-these-are-the-first-results-demonstrating-causality-from-robustness-to-superpositionAdversarial training reduces superposition; these are the first results demonstrating causality from robustness to superposition.
Novelty claim.
Source paper
extracted_from(2025) · Gorton, Liv · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (2)
finding
- Robustness controls superposition.
- Figure 4.
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.
- Unsatisfying reduction.
- Adversarially robust models have different interpretability because adversarial training affects superpositionhypothesis0.869Could be tested by crosscoder comparisons.
- Interpretation of Figure 3.
- Central thesis.
- Superposition maximalist stance.
- Supported by Table 1.
- Adversarially robust models would have worse performance and be more interpretable because robustness requires giving up superpositionhypothesis0.801Predictions of superposition maximalism.
- Adversarial vulnerability increases >3x as superposition forms and tracks features per dimensionfinding0.797L2 attacks max norm 0.1 of average input norm.