Code Generation on HumanEval (avg@16 accuracy)
88.43Avg Accuracy @16Qwen3-8B-Base + PASC-GRPO
Evaluation Results
| Method | Links | |
|---|---|---|
| Qwen3-8B-Base + PASC-GRPOmax_completion_tokens=8K2026.01 | 88.43 | |
| Gemma-3-27B-ITmax_completion_tokens=8K2026.01 | 87.8 | |
| Qwen3-8B-Base + GRP-SFTmax_completion_tokens=8K2026.01 | 86.79 | |
| Gemma-3-12B-ITmax_completion_tokens=8K2026.01 | 85.4 | |
| Qwen3-4B-Base + PASC-GRPOmax_completion_tokens=8K2026.01 | 82.45 | |
| Qwen3-4B-Base + GRP-SFTmax_completion_tokens=8K2026.01 | 81.67 | |
| Qwen3-8B-Basemax_completion_tokens=8K2026.01 | 80.47 | |
| Qwen3-4B-Basemax_completion_tokens=8K2026.01 | 75.6 | |
| LLaMA-3.1-8B-Instructmax_completion_tokens=8K2026.01 | 69.7 |