Code Generation on LiveCodeBench (Pass@1, Latency)
80.97Pass@1Full Baseline
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Full BaselineBackbone=Qwen3-14B2026.05 | 80.97 | 329.3 | |
| Qwen3-8BBackbone=Qwen3-8B2026.05 | 79.69 | 267.2 | |
| FlowCompileSelection Priority=Accuracy-priority2026.05 | 78.98 | 51.5 | |
| Qwen3-4BBackbone=Qwen3-4B2026.05 | 78.69 | 186.2 | |
| KNN Router2026.05 | 50.14 | 141.3 | |
| FlowCompileSelection Priority=Latency-priority2026.05 | 41.34 | 19 | |
| MaASBackbone=Qwen3-4B2026.05 | 26.56 | 33.7 | |
| MaASBackbone=Qwen3-8B2026.05 | 15.2 | 16.3 |