finding
active
finding:fgsm-increases-l0-by-about-1-3-at-conv1-46-2-vs-36-0FGSM increases L0 by about 1.3× at conv1 (46.2 vs 36.0).
Early layer effect.
Source paper
extracted_from(2025) · Gorton, Liv · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- If adversarial attacks exploit interference, they should activate more features than clean inputs.supportsPrediction tested via L0.
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.
- Contrast with PGD.
- Moderate mid-layer effect.
- Fig. 22.
- Table 2.
- Task accuracy on CausalGym increases consistently with model scale from 0.62 (14M) to 0.89 (6.9B)finding0.748Scaling result showing larger pythia models perform better on CausalGym linguistic tasks
- Quantitative evidence that NLA training produces increasingly informative explanations despite optimizing only for reconstruction.
- Ablation.
- Task-dependent optimal sparsity.