Math on AMOBench (Avg@16)
17.3Avg@16TRD
Evaluation Results
| Method | Links | |
|---|---|---|
| TRDBackbone=Qwen3-8B (w/ thinking), Traj.=yr2026.06 | 17.3 | |
| +Reverse KLBackbone=Qwen3-8B (w/ thinking), Traj.=yo2026.06 | 16.8 | |
| BaseBackbone=Qwen3-8B (w/ thinking), Traj.=–2026.06 | 15.9 | |
| +Forward KL w/ ClipBackbone=Qwen3-8B (w/ thinking), Traj.=yo2026.06 | 15.9 | |
| +Forward KLBackbone=Qwen3-8B (w/ thinking), Traj.=yo2026.06 | 15.7 | |
| +Reverse KL w/ Top-KBackbone=Qwen3-8B (w/ thinking), Traj.=yo2026.06 | 15.2 | |
| BaseBackbone=Qwen3-4B-Instruct-2507, Traj.=–2026.06 | 10.6 | |
| TRDBackbone=Qwen3-4B-Instruct-2507, Traj.=yr2026.06 | 10.3 | |
| +Reverse KLBackbone=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 10.1 | |
| +Reverse KL w/ Top-KBackbone=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 9.6 | |
| +Forward KLBackbone=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 9.5 | |
| +Forward KL w/ ClipBackbone=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 9.3 |