Logical reasoning and game playing on GEM
99.3Sudoku pass@1Off. + On.
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Off. + On.Backbone=Qwen3-30B-A3B-Thinking-2507, Task Setup=Multi-Task2026.06 | 99.3 | 97.1 | 93.9 | 78.4 | — | — | |
| On-policy Distill.Backbone=Qwen3-30B-A3B-Thinking-2507, Task Setup=Multi-Task2026.06 | 99.1 | 96.4 | 93.4 | 77 | — | — | |
| Off-policy Distill.Backbone=Qwen3-30B-A3B-Thinking-2507, Task Setup=Single-Task2026.06 | 98.9 | 95.7 | 94.6 | 81 | — | — | |
| RLBackbone=Qwen3-30B-A3B-Thinking-2507, Task Setup=Single-Task2026.06 | 98.7 | 94.9 | 93.5 | 77.1 | — | — | |
| Off-policy Distill.Backbone=Qwen3-30B-A3B-Thinking-2507, Task Setup=Multi-Task2026.06 | 98.2 | 93 | 91.5 | 71.8 | — | — | |
| Pre-RLBackbone=Qwen3-30B-A3B-Thinking-2507, Task Setup=Single-Task2026.06 | 5.1 | 0 | 1 | 0 | — | — | |
| Off-policy Distill.Backbone=Qwen3-8B, Protocol=Single-Task2026.06 | — | — | — | — | 93.2 | 75.4 | |
| Off-policy Distill.Backbone=Qwen3-8B, Protocol=Multi-Task2026.06 | — | — | — | — | 93.8 | 77.5 | |
| Off. + On.Backbone=Qwen3-8B, Protocol=Multi-Task2026.06 | — | — | — | — | 94.8 | 81 | |
| On-policy Distill.Backbone=Qwen3-8B, Protocol=Multi-Task2026.06 | — | — | — | — | 93.9 | 77.8 | |
| Pre-RLBackbone=Qwen3-8B, Protocol=Single-Task2026.06 | — | — | — | — | 2.2 | 0 | |
| RLBackbone=Qwen3-8B, Protocol=Single-Task2026.06 | — | — | — | — | 92.8 | 73.7 |