Grokking Delay Prediction on Runs (held-out)
17.7MAPEFirst-passage law of grokking delay
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| First-passage law of grokking delayOptimizer=AdamW, Quantity predicted=first-passage Tgrok - Tmem, Axes used=Vt & αt (both), Joint necessity=Thm 1 quantile-margin, Causal evidence=Block F 0/6 vs. 3/32026.05 | 17.7 | 15 |