finding
active
finding:robust-toy-models-are-often-more-robust-than-expected-for-their-superposition-levelRobust toy models are often more robust than expected for their superposition level.
Unexpected result.
Source paper
extracted_from(2025) · Gorton, Liv · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Explains robust models being more robust than expected.
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.
- Figure 4.
- Non-robust toy models have mean off-diagonal interference about 2× that of robust models.finding0.832Figure 3B.
- Superposition controls robustness.
- Authors' hedged interpretation.
- Toy model predictions are consistent with empirical observations on polysemanticity in real networksclaim0.802Validation; geometry and dynamics less certain.
- Adversarially robust models would have worse performance and be more interpretable because robustness requires giving up superpositionhypothesis0.797Predictions of superposition maximalism.
- Main contribution for interpretability.
- Robustness influences superposition, predicting lower SAE loss in more robust models.hypothesis0.794Popperian credit for surprising prediction.