finding
active
finding:adversarial-training-reduces-features-per-dimension-and-adversarial-vulnerability-in-toy-models-at-given-sparsity-levelsAdversarial training reduces features per dimension and adversarial vulnerability in toy models at given sparsity levels.
Robustness controls superposition.
Source paper
extracted_from(2025) · Gorton, Liv · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (2)
claim
- Central thesis.
- Novelty claim.
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.
- Superposition controls robustness.
- Unsatisfying reduction.
- Adversarially robust models have different interpretability because adversarial training affects superpositionhypothesis0.836Could be tested by crosscoder comparisons.
- Appendix replication.
- Interpretation of Figure 3.
- Adversarial vulnerability increases >3x as superposition forms and tracks features per dimensionfinding0.797L2 attacks max norm 0.1 of average input norm.
- Figure 4.
- Training that suppresses representations of undesirable concepts may incentivize new representations, removing the detection signalhypothesis0.794Tension for monitorability.