Mathematical Reasoning on AIME (acc, len)
35.69AIME AccuracySA-GRPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SA-GRPOBase Model=Qwen3-4B, alpha=0.22025.09 | 35.69 | 1,788 | |
| GRPOBase Model=Qwen3-4B2025.09 | 30.11 | 2,113 | |
| GRPO-ERBase Model=Qwen3-4B, alpha=0.052025.09 | 28.94 | 1,022 | |
| GRPO-O1Base Model=Qwen3-4B, alpha=0.052025.09 | 28.62 | 1,016 | |
| SA-PPOBase Model=Qwen3-4B2025.09 | 28.19 | 1,397 | |
| PPOBase Model=Qwen3-4B2025.09 | 26.9 | 2,680 | |
| Qwen3-4BBase Model=Qwen3-4B2025.09 | 20.15 | 1,281 | |
| Gemma3-4BBackbone=Gemma3-4B, Strategy=Base2025.09 | 18.76 | 1,843 | |
| GRPO-O1Backbone=Gemma3-4B2025.09 | 18.54 | 1,814 | |
| SA-GRPOBackbone=Gemma3-4B2025.09 | 18.11 | 2,379 | |
| GRPOBackbone=Gemma3-4B2025.09 | 18.01 | 2,357 | |
| GRPO-ERBackbone=Gemma3-4B2025.09 | 17.36 | 2,268 | |
| GRPOBase Model=Qwen3-1.7B2025.09 | 13.93 | 1,278 | |
| GRPO-O1Base Model=Qwen3-1.7B, alpha=0.052025.09 | 13.61 | 910 | |
| SA-GRPOBase Model=Qwen3-1.7B, alpha=0.22025.09 | 13.5 | 992 | |
| GRPO-ERBase Model=Qwen3-1.7B, alpha=0.052025.09 | 12.22 | 1,232 | |
| SA-PPOBase Model=Qwen3-1.7B2025.09 | 10.61 | 1,015 | |
| Qwen3-1.7BBase Model=Qwen3-1.7B2025.09 | 9.97 | 1,539 | |
| PPOBase Model=Qwen3-1.7B2025.09 | 9.11 | 1,718 | |
| PPOBase Model=Phi-3.5-mini2025.09 | 3.85 | 926 | |
| SA-GRPOBase Model=Phi-3.5-mini, alpha=0.22025.09 | 3.22 | 535 | |
| GRPOBase Model=Phi-3.5-mini2025.09 | 3.21 | 672 | |
| SA-PPOBase Model=Phi-3.5-mini2025.09 | 2.79 | 501 | |
| GRPO-O1Base Model=Phi-3.5-mini, alpha=0.052025.09 | 2.57 | 583 | |
| GRPO-ERBase Model=Phi-3.5-mini, alpha=0.052025.09 | 2.47 | 606 | |
| Phi-3.5-miniBase Model=Phi-3.5-mini2025.09 | 2.35 | 943 | |
| GRPO-O1Base Model=Gemma3-1B, alpha=0.052025.09 | 2.35 | 2,449 | |
| Gemma3-1BBase Model=Gemma3-1B2025.09 | 2.25 | 2,279 | |
| GRPOBase Model=Gemma3-1B2025.09 | 2.25 | 2,613 | |
| SA-GRPOBase Model=Gemma3-1B, alpha=0.22025.09 | 2.14 | 1,509 | |
| GRPO-ERBase Model=Gemma3-1B, alpha=0.052025.09 | 2.04 | 1,696 | |
| PPOBase Model=Gemma3-1B2025.09 | 1.6 | 2,422 | |
| SA-PPOBase Model=Gemma3-1B2025.09 | 1.5 | 1,683 | |
| SA-GRPOBackbone=Qwen2-2B-VL2025.09 | 1.29 | 57 | |
| GRPOBackbone=Qwen2-2B-VL2025.09 | 0.86 | 561 | |
| Qwen2-2B-VLBackbone=Qwen2-2B-VL, Strategy=Base2025.09 | 0.54 | 2,404 | |
| GRPO-ERBackbone=Qwen2-2B-VL2025.09 | 0.43 | 383 | |
| GRPO-O1Backbone=Qwen2-2B-VL2025.09 | 0.32 | 17 |