Code Generation on HumanEval+ (Avg@16)
88.4Avg@16 ScoreForward KL w/ Clip
Evaluation Results
| Method | Links | |
|---|---|---|
| Forward KL w/ ClipModel Backbone=Qwen3-4B-Instruct-2507, Trajectory Type (Traj.)=yo, Teacher Model=Qwen3-8B2026.06 | 88.4 | |
| Reverse KLModel Backbone=Qwen3-4B-Instruct-2507, Trajectory Type (Traj.)=yo, Teacher Model=Qwen3-8B2026.06 | 88.4 | |
| TRDModel Backbone=Qwen3-4B-Instruct-2507, Trajectory Type (Traj.)=yr, Teacher Model=Qwen3-8B2026.06 | 88.4 | |
| BaseModel Backbone=Qwen3-4B-Instruct-2507, Trajectory Type (Traj.)=-, Teacher Model=Qwen3-8B2026.06 | 87.8 | |
| Forward KLModel Backbone=Qwen3-4B-Instruct-2507, Trajectory Type (Traj.)=yo, Teacher Model=Qwen3-8B2026.06 | 87.2 | |
| Reverse KL w/ Top-KModel Backbone=Qwen3-4B-Instruct-2507, Trajectory Type (Traj.)=yo, Teacher Model=Qwen3-8B2026.06 | 87.2 | |
| TRDModel=Qwen3-4B-Instruct-2507, Traj.=yr2026.06 | 82.1 | |
| BaseModel=Qwen3-4B-Instruct-2507, Traj.=–2026.06 | 82 | |
| Forward KL w/ ClipModel=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 81.8 | |
| Reverse KL w/ Top-KModel=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 81.8 | |
| Forward KLModel=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 81.7 | |
| Reverse KLModel=Qwen3-4B-Instruct-2507, Traj.=yo2026.06 | 81.6 | |
| Reverse KL w/ Top-KModel Backbone=Qwen3-1.7B (w/ thinking), Trajectory Type (Traj.)=yo, Teacher Model=Qwen3-8B2026.06 | 78 | |
| TRDModel Backbone=Qwen3-1.7B (w/ thinking), Trajectory Type (Traj.)=yr, Teacher Model=Qwen3-8B2026.06 | 78 | |
| Forward KL w/ ClipModel Backbone=Qwen3-1.7B (w/ thinking), Trajectory Type (Traj.)=yo, Teacher Model=Qwen3-8B2026.06 | 77.4 | |
| Reverse KLModel Backbone=Qwen3-1.7B (w/ thinking), Trajectory Type (Traj.)=yo, Teacher Model=Qwen3-8B2026.06 | 76.2 | |
| BaseModel Backbone=Qwen3-1.7B (w/ thinking), Trajectory Type (Traj.)=-, Teacher Model=Qwen3-8B2026.06 | 75.6 | |
| Forward KLModel Backbone=Qwen3-1.7B (w/ thinking), Trajectory Type (Traj.)=yo, Teacher Model=Qwen3-8B2026.06 | 71.3 | |
| TRDModel=Qwen3-1.7B (w/ thinking), Traj.=yr2026.06 | 63.2 | |
| Reverse KLModel=Qwen3-1.7B (w/ thinking), Traj.=yo2026.06 | 63 | |
| Reverse KL w/ Top-KModel=Qwen3-1.7B (w/ thinking), Traj.=yo2026.06 | 62.9 | |
| Forward KL w/ ClipModel=Qwen3-1.7B (w/ thinking), Traj.=yo2026.06 | 62.6 | |
| BaseModel=Qwen3-1.7B (w/ thinking), Traj.=–2026.06 | 62.3 | |
| Forward KLModel=Qwen3-1.7B (w/ thinking), Traj.=yo2026.06 | 61.4 |