Code Reasoning on LiveCodeBench (Pass@1)
88.3Pass@1 AccuracyREBALANCE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| REBALANCEBackbone=QwQ-32B2026.03 | 88.3 | 5,000 | |
| BaselineBackbone=QwQ-32B2026.03 | 87.5 | 6,000 | |
| REBALANCEBackbone=Qwen3-14B2026.03 | 84.6 | 6,000 | |
| BaselineBackbone=Qwen3-14B2026.03 | 83.5 | 7,000 | |
| SnapStreamMaximum sequence length=128k, KV cache length=32k2025.11 | 78.08 | — | |
| DeepSeek-R1-0528Maximum sequence length=32k, KV cache length=32k2025.11 | 73.97 | — | |
| DeepSeek-R1-0528Maximum sequence length=128k, KV cache length=Full2025.11 | 71.43 | — | |
| REBALANCEBackbone=DeepSeek-R1-Distill-Qwen-7B2026.03 | 46.5 | 8,000 | |
| BaselineBackbone=DeepSeek-R1-Distill-Qwen-7B2026.03 | 44 | 9,000 | |
| REBALANCEBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.03 | 22.5 | 11,000 | |
| BaselineBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.03 | 19.5 | 12,000 |