Math Reasoning on AMC 2022-2024 (avg@32 accuracy, generation length)
51.24Accuracy @32GRPO + RSI-S
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| GRPO + RSI-SBackbone=Qwen2.5-7B2026.06 | 51.24 | 832.76 | 3.83 | -186.67 | |
| GRPO + EBBackbone=Qwen2.5-7B2026.06 | 49.79 | 1,036.5 | — | — | |
| GRPOBackbone=Qwen2.5-7B2026.06 | 47.41 | 1,019.43 | — | — | |
| GRPO + PBBackbone=Qwen2.5-7B2026.06 | 47.19 | 908.22 | — | — |