Out-of-Distribution Generalization on MATH-500 OOD (test)
67Score (Sav.)TTT QK (dh=128)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| TTT QK (dh=128)Labeling protocol=Supervised labels, delta=0.1, zero-shot=true2026.04 | 67 | 2.1 | |
| TTT no-QKLabeling protocol=Supervised labels, delta=0.1, zero-shot=true2026.04 | 63.7 | 2.3 | |
| TTT QK (dh=128)Labeling protocol=Consistent labels (no ground truth), delta=0.1, zero-shot=true2026.04 | 63.7 | 1.6 | |
| TTT no-QKLabeling protocol=Consistent labels (no ground truth), delta=0.1, zero-shot=true2026.04 | 55.5 | 1.2 | |
| Static ProbeLabeling protocol=Supervised labels, delta=0.1, zero-shot=true2026.04 | 24.8 | 0.8 | |
| Static ProbeLabeling protocol=Consistent labels (no ground truth), delta=0.1, zero-shot=true2026.04 | 23.9 | 0.4 |