N, not Γ
The go/no-go probe that killed a theory
Predicted window:
Main sweep — test accuracy on novel compositions, 12k steps, seed 0.
| Γ (weight decay) |
N=12 | N=24 | N=48 |
Fine sweep at N=24 — two seeds, train and test accuracy.
| Γ (weight decay) |
seed 0 train | seed 0 test |
seed 1 train | seed 1 test |