Math Reasoning on AMC (Accuracy (%))
68.2Accuracy (%)Prefix-RFT
Evaluation Results
| Method | Links | |
|---|---|---|
| Prefix-RFTBase Model=Qwen2.5-Math-7B2025.07 | 68.2 | |
| LUFFYBase Model=Qwen2.5-Math-7B2025.07 | 65.6 | |
| ReLIFTBase Model=Qwen2.5-Math-7B2025.07 | 64.9 | |
| GRPOModel=Qwen3-4B-Base, QB (Query-preserved mini batching)=true, RB (Reward-balanced batching)=false2026.05 | 64.4 | |
| GRPOModel=Qwen3-4B-Base, QB (Query-preserved mini batching)=true, RB (Reward-balanced batching)=true2026.05 | 64 | |
| GRPOModel=Qwen2.5-Math-7B, QB (Query-preserved mini batching)=true, RB (Reward-balanced batching)=true2026.05 | 63.9 | |
| SFT+RFTBase Model=Qwen2.5-Math-7B2025.07 | 62.7 | |
| RFTBase Model=Qwen2.5-Math-7B2025.07 | 62 | |
| GRPOModel=Qwen3-4B-Base, QB (Query-preserved mini batching)=false, RB (Reward-balanced batching)=true2026.05 | 61.5 | |
| GRPOModel=Qwen2.5-Math-7B, QB (Query-preserved mini batching)=true, RB (Reward-balanced batching)=false2026.05 | 61.2 | |
| Oat-ZeroBase Model=Qwen2.5-Math-7B2025.07 | 61.2 | |
| GRPOModel=Qwen2.5-Math-7B, QB (Query-preserved mini batching)=false, RB (Reward-balanced batching)=true2026.05 | 60.1 | |
| UFTBase Model=Qwen2.5-Math-7B2025.07 | 58.8 | |
| GRPOModel=Qwen3-4B-Base, QB (Query-preserved mini batching)=false, RB (Reward-balanced batching)=false2026.05 | 58 | |
| GRPOModel=Qwen2.5-Math-7B, QB (Query-preserved mini batching)=false, RB (Reward-balanced batching)=false2026.05 | 54.9 | |
| SimpleRL-ZeroBase Model=Qwen2.5-Math-7B2025.07 | 54.9 | |
| SFTBase Model=Qwen2.5-Math-7B2025.07 | 52.8 | |
| RL w/ SFT LossBase Model=Qwen2.5-Math-7B2025.07 | 49.7 | |
| GRPOModel=Qwen3-1.7B-Base, QB (Query-preserved mini batching)=true, RB (Reward-balanced batching)=true2026.05 | 43.5 | |
| GRPOModel=Qwen3-1.7B-Base, QB (Query-preserved mini batching)=false, RB (Reward-balanced batching)=true2026.05 | 42.6 | |
| GRPOModel=Qwen3-1.7B-Base, QB (Query-preserved mini batching)=true, RB (Reward-balanced batching)=false2026.05 | 40.2 | |
| GRPOModel=Qwen3-1.7B-Base, QB (Query-preserved mini batching)=false, RB (Reward-balanced batching)=false2026.05 | 37 | |
| Qwen2.5-Math-7B2025.07 | 31.3 |