Mathematics on AIME25 (Avg@16)
69.2Avg@16TRD
Evaluation Results
| Method | Links | |
|---|---|---|
| TRDBackbone=Qwen3-8B (w/ thinking), Traj.=yr2026.06 | 69.2 | |
| +Reverse KL w/ Top-KBackbone=Qwen3-8B (w/ thinking), Traj.=yo2026.06 | 68.5 | |
| +Forward KL w/ ClipBackbone=Qwen3-8B (w/ thinking), Traj.=yo2026.06 | 68.3 | |
| +Reverse KLBackbone=Qwen3-8B (w/ thinking), Traj.=yo2026.06 | 68.2 | |
| BaseBackbone=Qwen3-8B (w/ thinking), Traj.=–2026.06 | 66.7 | |
| +Forward KLBackbone=Qwen3-8B (w/ thinking), Traj.=yo2026.06 | 65.4 | |
| TRDBackbone=Qwen3-4B-Instruct-2507, Traj.=yr2026.06 | 49.4 | |
| +Forward KL w/ ClipBackbone=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 49.2 | |
| +Reverse KL w/ Top-KBackbone=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 49 | |
| +Forward KLBackbone=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 48.8 | |
| +Reverse KLBackbone=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 48.8 | |
| BaseBackbone=Qwen3-4B-Instruct-2507, Traj.=–2026.06 | 48.2 | |
| AIPOModel=Qwen2.5-32B-Instruct2026.05 | 39.69 | |
| GRPOModel=Qwen2.5-32B-Instruct2026.05 | 34.48 | |
| SFTModel=Qwen2.5-32B-Instruct2026.05 | 31.26 | |
| OriginalModel=Qwen2.5-32B-Instruct2026.05 | 28.62 |