Qwen3.5-9B · 400 calibration puzzles (rating 808–2776) · same 8,000 training puzzles
| Series | Training samples seen | Calibration mean reward | Interval / statistics |
|---|---|---|---|
| GRPO baseline | 0 | 5.389 | |
| GRPO baseline | 8.192 | 6.267 | |
| GRPO baseline | 16.384 | 10.014 | |
| GRPO baseline | 24.576 | 10.684 | |
| GRPO baseline | 32.768 | 15.231 | |
| GRPO baseline | 40.96 | 18.104 | |
| GRPO baseline | 49.152 | 22.576 | |
| GRPO baseline | GRPO baseline | 23.361 | |
| GRPO baseline | 65.536 | 23.349 | |
| GRPO baseline | 68 | 21.5 | |
| GRPO baseline | 70 | 20.9 | |
| Warmup | 0 | 5.389 | |
| Warmup | Warmup | 13.371 | |
| Specialists | 1.28 | 13.371 | |
| Specialists | 2.56 | 22.107 | |
| Specialists | 4.096 | 21.718 | |
| Specialists | 5.632 | 19.943 | |
| Specialists | 7.168 | 20.759 | |
| Specialists | 8.704 | 22.724 | |
| Specialists | 10.24 | 23.283 | |
| Specialists | 11.776 | 23.588 | |
| Specialists | 13.312 | 23.765 | |
| Specialists | 14.848 | 24.805 | |
| Specialists | 16.384 | 24.563 | |
| Specialists | 17.92 | 25.434 | |
| Specialists | 26.112 | 27.277 | |
| Specialists | 34.304 | 26.538 | |
| Specialists | Specialists | 27.209 | |
| Specialists | 43.52 | 26.57 | |
| Specialists | 48.128 | 26.198 | |
| Specialists | 52.736 | 27.054 | |
| Distill | 0 | 6 | |
| Distill | 1.536 | 22.312 | |
| Distill | 3.072 | 27.799 | |
| Distill | 4.608 | 30.998 | |
| Distill | 6.144 | 31.039 | |
| Distill | 7.68 | 29.831 | |
| Distill | 9.216 | 30.356 | |
| Distill | 10.752 | 30.766 | |
| Distill | 12.288 | 31.601 | |
| Distill | Distill | 31.894 |