method
active
method:analytic-optimal-l2-attackAnalytic optimal L2 attack
Used instead of gradient descent attacks due to gradient masking.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Toy Models of Superpositionmentions
Concepts (1)
concept
- Adversarial examplesaboutLinked to superposition interference.
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.
- Analytically constructed per-feature attack.
- Primary attack used for toy model robustness.
- Reduces superposition but needs 80% L2 norm.
- Sparsity trade-off.
- Security attack that bypasses LLM safety alignment by suppressing deliberation or exploiting reflection inhibition.
- Primary evaluation metric defined as the fraction of model responses classified as unsafe
- Unsatisfying reduction.
- Alternative penalty that spreads features across layers and finds only shared features.