Grokking Generalization on Modular Addition (train/test)
14Grokked CountSupCon-wrong
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| SupCon-wrongn=15, lambda=all2026.07 | 14 | 25,200 | 1.25 | 12,950 | |
| Baselinen=102026.07 | 10 | 19,950 | — | 5,500 | |
| SupCon-truen=10, lambda=0.12026.07 | 8 | 28,725 | 1.17 | 5,100 | |
| SupCon-truen=10, lambda=1.02026.07 | 8 | 24,400 | 0.8 | 2,000 | |
| SupCon-truen=10, lambda=0.32026.07 | 6 | 33,325 | 1.22 | 2,850 | |
| Grokfastn=52026.07 | 5 | 16,150 | 0.77 | 10,300 | |
| SupCon-shuffledn=20, lambda=all2026.07 | 0 | 50,000 | 2.2 | — | |
| Norm-matchedn=15, lambda=all2026.07 | 0 | 50,000 | 2.2 | — |