Math Reasoning on MATH 500 (BA @20, BA @40, BA @60)
86.2BA @20LRM-only
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LRM-onlySetting=Same-family, Models=Qwen3-1.7B / Qwen3-14B2026.05 | 86.2 | 86.2 | 86.2 | |
| LRM-onlySetting=Cross-family, Models=Qwen3-1.7B / DeepSeek-R1-Distill-Qwen-14B2026.05 | 82.6 | 82.6 | 82.6 | |
| GlimpRouterSetting=Same-family, Models=Qwen3-1.7B / Qwen3-14B2026.05 | 76 | 78 | 80.1 | |
| RoRoSetting=Same-family, Models=Qwen3-1.7B / Qwen3-14B2026.05 | 75.5 | 79.8 | 81.9 | |
| TRIMSetting=Same-family, Models=Qwen3-1.7B / Qwen3-14B2026.05 | 75.1 | 77.5 | 79.8 | |
| STEERSetting=Same-family, Models=Qwen3-1.7B / Qwen3-14B2026.05 | 74.3 | 77.4 | 79.9 | |
| RoRoSetting=Cross-family, Models=Qwen3-1.7B / DeepSeek-R1-Distill-Qwen-14B2026.05 | 74.2 | 78.2 | 80.6 | |
| SpecReasonSetting=Same-family, Models=Qwen3-1.7B / Qwen3-14B2026.05 | 73.9 | 76.9 | 79.4 | |
| TRIMSetting=Cross-family, Models=Qwen3-1.7B / DeepSeek-R1-Distill-Qwen-14B2026.05 | 73.8 | 76.2 | 78.5 | |
| SpecCoTSetting=Same-family, Models=Qwen3-1.7B / Qwen3-14B2026.05 | 73.4 | 77.8 | 78.8 | |
| GlimpRouterSetting=Cross-family, Models=Qwen3-1.7B / DeepSeek-R1-Distill-Qwen-14B2026.05 | 73.4 | 78.8 | 79.2 | |
| STEERSetting=Cross-family, Models=Qwen3-1.7B / DeepSeek-R1-Distill-Qwen-14B2026.05 | 73.1 | 75.8 | 78.8 | |
| RSDSetting=Same-family, Models=Qwen3-1.7B / Qwen3-14B2026.05 | 72.8 | 75.8 | 79.2 | |
| SpecReasonSetting=Cross-family, Models=Qwen3-1.7B / DeepSeek-R1-Distill-Qwen-14B2026.05 | 72.6 | 75.2 | 78.4 | |
| SpecCoTSetting=Cross-family, Models=Qwen3-1.7B / DeepSeek-R1-Distill-Qwen-14B2026.05 | 72.1 | 75.8 | 78 | |
| RSDSetting=Cross-family, Models=Qwen3-1.7B / DeepSeek-R1-Distill-Qwen-14B2026.05 | 71.5 | 74.3 | 77.8 | |
| RandomSetting=Same-family, Models=Qwen3-1.7B / Qwen3-14B2026.05 | 70.4 | 74.2 | 76.1 | |
| RandomSetting=Cross-family, Models=Qwen3-1.7B / DeepSeek-R1-Distill-Qwen-14B2026.05 | 69.8 | 73.5 | 75.4 | |
| SRM-onlySetting=Same-family, Models=Qwen3-1.7B / Qwen3-14B2026.05 | 68.6 | 68.6 | 68.6 | |
| SRM-onlySetting=Cross-family, Models=Qwen3-1.7B / DeepSeek-R1-Distill-Qwen-14B2026.05 | 68.6 | 68.6 | 68.6 |