Code Generation on LiveCodeBench (LCB) (% Avg@4)
32.4% Avg@4RF++ B + RePro
Evaluation Results
| Method | Links | |
|---|---|---|
| RF++ B + ReProBackbone=Qwen3-1.7B2025.12 | 32.4 | |
| GRPO + ReProBackbone=Qwen3-1.7B2025.12 | 32 | |
| PPO + ReProBackbone=Qwen3-1.7B2025.12 | 31.5 | |
| PPOBackbone=Qwen3-1.7B2025.12 | 31.3 | |
| RF++ BBackbone=Qwen3-1.7B2025.12 | 31.1 | |
| OriginalBackbone=Qwen3-1.7B2025.12 | 30.6 | |
| GRPOBackbone=Qwen3-1.7B2025.12 | 30.2 | |
| GRPO + ReProBackbone=Hunyuan-1.8B-Instruct2025.12 | 27.7 | |
| GRPOBackbone=Hunyuan-1.8B-Instruct2025.12 | 27.1 | |
| RF++ BBackbone=Hunyuan-1.8B-Instruct2025.12 | 26.5 | |
| PPO + ReProBackbone=Hunyuan-1.8B-Instruct2025.12 | 26.3 | |
| RF++ B + ReProBackbone=Hunyuan-1.8B-Instruct2025.12 | 26.2 | |
| PPOBackbone=Hunyuan-1.8B-Instruct2025.12 | 25.9 | |
| OriginalBackbone=Hunyuan-1.8B-Instruct2025.12 | 24.5 | |
| PPO + ReProBackbone=MobileLLM-R1-950M2025.12 | 21.4 | |
| PPOBackbone=MobileLLM-R1-950M2025.12 | 19.8 | |
| OriginalBackbone=MobileLLM-R1-950M2025.12 | 18 |