OOD Generalization on ARC-C, GPQA*, and MMLU-Pro
83.1ARC-C ScoreBV-Blend
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| BV-BlendBackbone=Qwen2.5-Math-7B, Training Protocol=On-Policy RLVR2026.06 | 83.1 | 52.6 | 56.5 | 64.1 | |
| GRPOBackbone=Qwen2.5-Math-7B, Training Protocol=On-Policy RLVR, Variation=Our replication2026.06 | 82.3 | 40.4 | 49.3 | 57.3 | |
| LUFFY†Backbone=Qwen2.5-Math-7B, Training Protocol=Hybrid & Off-Policy Methods, Training Variant=Extended-training2026.06 | 81.8 | 49 | 54.7 | 61.8 | |
| LUFFYBackbone=Qwen2.5-Math-7B, Training Protocol=Hybrid & Off-Policy Methods, Training Variant=Standard2026.06 | 80.5 | 39.9 | 53 | 57.8 | |
| ReLIFTBackbone=Qwen2.5-Math-7B, Training Protocol=Hybrid & Off-Policy Methods2026.06 | 76.2 | 37.9 | 52.5 | 55.5 | |
| SFTBackbone=Qwen2.5-Math-7B, Training Protocol=Hybrid & Off-Policy Methods2026.06 | 75.2 | 24.7 | 42.7 | 47.5 | |
| PRIME-ZeroBackbone=Qwen2.5-Math-7B, Training Protocol=On-Policy RLVR2026.06 | 73.3 | 18.2 | 32.7 | 41.4 | |
| SFT+RLBackbone=Qwen2.5-Math-7B, Training Protocol=Hybrid & Off-Policy Methods2026.06 | 72.4 | 24.2 | 37.7 | 44.8 | |
| Qwen-InstructBackbone=Qwen2.5-Math-7B, Training Protocol=Instruct2026.06 | 70.3 | 24.7 | 34.1 | 43 | |
| Oat-ZeroBackbone=Qwen2.5-Math-7B, Training Protocol=On-Policy RLVR2026.06 | 70.1 | 23.7 | 41.7 | 45.2 | |
| OpenReasoner-ZeroBackbone=Qwen2.5-Math-7B, Training Protocol=On-Policy RLVR2026.06 | 66.2 | 29.8 | 58.7 | 51.6 | |
| SimpleRL-ZeroBackbone=Qwen2.5-Math-7B, Training Protocol=On-Policy RLVR2026.06 | 30.2 | 23.2 | 34.5 | 29.3 | |
| Qwen-BaseBackbone=Qwen2.5-Math-7B, Training Protocol=Base2026.06 | 18.2 | 11.1 | 16.9 | 15.4 |