Math Reasoning on Math500, AIME24, AIME25, Minerva, and OlympiadBench (test)
92.2Math500 AccuracyOpenThinker3-7B
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| OpenThinker3-7B2026.03 | 92.2 | 53.3 | 40 | 39 | 55.6 | 56 | |
| teacher top-K local support matchingmasking=false2026.03 | 82 | 33.3 | 16.7 | 32.7 | 44 | 41.7 | |
| teacher top-K local support matchingmasking=true2026.03 | 79 | 20 | 16.7 | 34.6 | 42.5 | 38.6 | |
| Sampled-token OPDmasking=true2026.03 | 76 | 20 | 13.3 | 33.5 | 40.4 | 36.6 | |
| Sampled-token OPDmasking=false2026.03 | 74.8 | 13.3 | 13.3 | 32.1 | 40.5 | 34.8 | |
| Qwen2.5-7B-It2026.03 | 68.2 | 13.3 | 0 | 26.5 | 32.9 | 28.2 |