Reasoning on NP-Bench (test)
93.7Graph SRGRPO
Evaluation Results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GRPO2026.05 | 93.7 | 30.5 | 69 | 49.6 | 100 | 51.8 | 63.3 | 44.3 | 91 | 24.3 | 83.4 | 40.1 | |
| GPG2026.05 | 92.7 | 23.5 | 96 | 28.5 | 100 | 51.8 | 59.3 | 39.6 | 95 | 20.3 | 88.6 | 32.7 | |
| ClipCov2026.05 | 91 | 25.5 | 93 | 52.2 | 96 | 49.8 | 73.3 | 58.1 | 82.5 | 21.2 | 87.2 | 41.3 | |
| GSPO2026.05 | 85.7 | 24.1 | 80 | 56.5 | 99 | 51.2 | 65.7 | 46.7 | 82.5 | 27.4 | 82.6 | 41.2 | |
| DMPO2026.05 | 85.3 | 25.6 | 82 | 60.3 | 100 | 52.8 | 67.7 | 55 | 81.5 | 25.6 | 83.3 | 43.9 | |
| Qwen2.5-7B-InstructRole=backbone2026.05 | 11 | 3.1 | 40 | 19.8 | 67 | 34 | 26.7 | 15.2 | 3.5 | 1 | 29.6 | 14.6 |