finding
active
finding:pgd-increases-l0-by-1-2-1-4-at-layers-2-3-of-resnet18PGD increases L0 by 1.2–1.4× at layers 2–3 of ResNet18.
Moderate mid-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.
- Table 2.
- Concrete numerical example showing detection and control are nearly identical at peak apparent accuracy
- Early layer effect.
- Contrast with PGD.
- Q8B most often peaks at layer 20, followed by layer 25; G20B most often peaks at layer 15finding0.757Best steering layer is model-specific; linearly accessible trait information is organized differently across models
- Forks missed with coarse spacing.
- Training configuration.
- CL training increases ⟨z,µ+⟩ from 0.64 to 0.75 and decreases ⟨z,µ−⟩ from 0.38 to 0.21 on Mistral-7Bfinding0.725Replicates CL alignment effect on second backbone, confirming generalizability