Share your thoughts, 1 month free Claude Pro on usSee more

Scaling-law extrapolation on MNIST high-D holdout

0.122RMSE (log space)

Muennighoff

Updated 22d ago

Evaluation Results

Method	Links
Muennighoff 2026.05		0.122
Chinchilla 2026.05		0.123
Ours (Eq. (2)) 2026.05		0.137
BNSL k=2 2026.05		0.169
M4 (D-axis) 2026.05		0.192
Farseer 2026.05		1.49