Code Generation on MBPP (% Avg@4)
77.5Avg@4 (%)GRPO + RePro
Evaluation Results
| Method | Links | |
|---|---|---|
| GRPO + ReProBackbone=Hunyuan-1.8B-Instruct2025.12 | 77.5 | |
| RF++ B + ReProBackbone=Hunyuan-1.8B-Instruct2025.12 | 77 | |
| RF++ BBackbone=Hunyuan-1.8B-Instruct2025.12 | 76.3 | |
| PPO + ReProBackbone=Hunyuan-1.8B-Instruct2025.12 | 76.2 | |
| GRPOBackbone=Hunyuan-1.8B-Instruct2025.12 | 75.9 | |
| PPOBackbone=Hunyuan-1.8B-Instruct2025.12 | 75.8 | |
| OriginalBackbone=Hunyuan-1.8B-Instruct2025.12 | 73.1 | |
| RF++ B + ReProBackbone=Qwen3-1.7B2025.12 | 71.7 | |
| PPO + ReProBackbone=Qwen3-1.7B2025.12 | 69.1 | |
| GRPO + ReProBackbone=Qwen3-1.7B2025.12 | 68.8 | |
| PPOBackbone=Qwen3-1.7B2025.12 | 68.2 | |
| GRPOBackbone=Qwen3-1.7B2025.12 | 67.5 | |
| OriginalBackbone=Qwen3-1.7B2025.12 | 66.9 | |
| RF++ BBackbone=Qwen3-1.7B2025.12 | 66.1 | |
| PPO + ReProBackbone=MobileLLM-R1-950M2025.12 | 65.1 | |
| PPOBackbone=MobileLLM-R1-950M2025.12 | 63.6 | |
| OriginalBackbone=MobileLLM-R1-950M2025.12 | 58.2 |