Mathematical Reasoning on AMO-Bench (average@16)
14.8Average@16GRPO
Evaluation Results
| Method | Links | |
|---|---|---|
| GRPOModel Scale=8B, Backbone=Qwen3, Temperature=1.2, Generation Length=38k2026.01 | 14.8 | |
| OPSDModel Scale=8B, Backbone=Qwen3, Temperature=1.2, Generation Length=38k2026.01 | 14.3 | |
| OPSDModel Scale=4B, Backbone=Qwen3, Temperature=1.2, Generation Length=38k2026.01 | 13.5 | |
| Qwen3 (Base Instruct)Model Scale=8B, Temperature=1.2, Generation Length=38k2026.01 | 13.4 | |
| SFTModel Scale=8B, Backbone=Qwen3, Temperature=1.2, Generation Length=38k2026.01 | 12.9 | |
| GRPOModel Scale=4B, Backbone=Qwen3, Temperature=1.2, Generation Length=38k2026.01 | 12.8 | |
| SFTModel Scale=4B, Backbone=Qwen3, Temperature=1.2, Generation Length=38k2026.01 | 12.5 | |
| Qwen3 (Base Instruct)Model Scale=4B, Temperature=1.2, Generation Length=38k2026.01 | 12.4 | |
| TRDModel=Qwen3-4B-Instruct-2507, Traj.=yr2026.06 | 10.3 | |
| BaseModel=Qwen3-4B-Instruct-2507, Traj.=–2026.06 | 10.1 | |
| Forward KL w/ ClipModel=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 10.1 | |
| Forward KLModel=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 9.9 | |
| Reverse KL w/ Top-KModel=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 9.8 | |
| Reverse KLModel=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 8.2 | |
| OPSDModel Scale=1.7B, Backbone=Qwen3, Temperature=1.2, Generation Length=38k2026.01 | 5 | |
| GRPOModel Scale=1.7B, Backbone=Qwen3, Temperature=1.2, Generation Length=38k2026.01 | 4.5 | |
| Qwen3 (Base Instruct)Model Scale=1.7B, Temperature=1.2, Generation Length=38k2026.01 | 4.3 | |
| SFTModel Scale=1.7B, Backbone=Qwen3, Temperature=1.2, Generation Length=38k2026.01 | 3.9 | |
| Reverse KL w/ Top-KModel=Qwen3-1.7B (w/ thinking), Traj.=yo2026.06 | 3 | |
| TRDModel=Qwen3-1.7B (w/ thinking), Traj.=yr2026.06 | 3 | |
| Forward KLModel=Qwen3-1.7B (w/ thinking), Traj.=yo2026.06 | 2.4 | |
| Reverse KLModel=Qwen3-1.7B (w/ thinking), Traj.=yo2026.06 | 2.4 | |
| BaseModel=Qwen3-1.7B (w/ thinking), Traj.=–2026.06 | 2.1 | |
| Forward KL w/ ClipModel=Qwen3-1.7B (w/ thinking), Traj.=yo2026.06 | 2.1 |