Kernel Code Generation on KernelBench Level-3
100AccuracyIR
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| IRPass Strategy=Multi-Pass, Steps=50, Skill Memory=false, Base Model=GPT-5-mini2026.04 | 100 | 1.31 | 24 | 12 | |
| IRPass Strategy=Multi-Pass, Steps=50, Skill Memory=true, Base Model=GPT-5-mini2026.04 | 100 | 1.16 | 10 | 4 | |
| OpenEvolvePass Strategy=Multi-Pass, Steps=50, Skill Memory=true, Base Model=GPT-5-mini2026.04 | 100 | 1.47 | 28 | 10 | |
| AdaExplorePass Strategy=Multi-Pass, Steps=50, Base Model=GPT-5-mini2026.04 | 100 | 1.55 | 28 | 16 | |
| AdaExplorePass Strategy=Multi-Pass, Steps=100, Base Model=GPT-5-mini2026.04 | 100 | 1.72 | 32 | 20 | |
| AdaExplorePass Strategy=Multi-Pass, Steps=200, Base Model=GPT-5-mini2026.04 | 100 | 1.78 | 36 | 22 | |
| PSPass Strategy=Multi-Pass, Steps=50, Skill Memory=true, Base Model=GPT-5-mini2026.04 | 98 | 1.12 | 28 | 4 | |
| PSPass Strategy=Multi-Pass, Steps=50, Skill Memory=false, Base Model=GPT-5-mini2026.04 | 92 | 0.97 | 18 | 4 | |
| DR. KernelPass Strategy=Multi-Pass, Steps=4 x 14, Base Model=GPT-5-mini2026.04 | 84 | 0.97 | 16 | 8 | |
| Claude-4.6-OpusPass Strategy=Single-Pass2026.04 | 72 | 0.84 | 8 | 6 | |
| GPT-5Pass Strategy=Single-Pass2026.04 | 44 | 0.55 | 12 | 4 | |
| AutoTritonPass Strategy=Single-Pass2026.04 | 32 | 0.3 | 0 | 0 | |
| GPT-5-miniPass Strategy=Single-Pass2026.04 | 22 | 0.21 | 2 | 0 | |
| DR. KernelPass Strategy=Single-Pass2026.04 | 10 | 0.21 | 2 | 0 |