claim
active
claim:overall-superposition-level-doesn-t-tell-the-full-story-some-superposition-structures-are-more-vulnerable-than-othersOverall superposition level doesn't tell the full story; some superposition structures are more vulnerable than others.
Explains robust models being more robust than expected.
Source paper
extracted_from(2025) · Gorton, Liv · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Unexpected result.
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.
- Confound addressed.
- Supported by Table 1.
- Superposition maximalist stance.
- Central thesis.
- superposition is deeply linked to the ability to identify and enumerate over all features in a modelquote0.788Safety rationale.
- Superposition may be a major contributing factor, or even the primary cause, of adversarial examples.hypothesis0.783Hypothesis under test.
- Authors' overall conclusion from number of interpretable features, activation-level correspondence to intensity, sensible logit weights, and interference weights
- Superposition hypothesis: neural networks represent more features than dimensions using almost-orthogonal directions.hypothesis0.778Explanation for why dictionary learning can recover many more features than dimensions.