finding
active
finding:features-jump-between-dimensionalities-during-training-with-sudden-loss-drops-in-digon-modelsFeatures jump between dimensionalities during training with sudden loss drops in digon models
Learning dynamics.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Toy Models of Superpositionmentions
Claims (1)
claim
- Suspicion from energy level jumps.
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.
- Perturbing a single feature.
- Primary empirical claim of the paper
- Quantitative relationship between concept frequency and feature presence.
- How do features change over model training; do they form abruptly or gradually, and do their directions drift?question0.728Fresh-traction question.
- MLP and attention analysis.
- Core demonstration.
- Controls for dataset structure, showing trained model activations have richer structure than data distribution alone
- The central hypothesis of the paper; the platonic representation hypothesis itself