Math Reasoning on (Math-500, AMC23, GSM8k, Minerva, Olympiad)
71.8Math-500GXPO
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GXPOModel=Qwen2.5-7B, k=5, BP=32026.05 | 71.8 | 50 | 88.63 | 23.58 | 20.79 | 50.96 | |
| SFPOModel=Qwen2.5-7B, k=5, BP=62026.05 | 71.75 | 49.87 | 88.49 | 23.43 | 20.57 | 50.82 | |
| GXPOModel=Qwen2.5-7B, k=3, BP=32026.05 | 71.7 | 47.6 | 88.6 | 23.66 | 20.76 | 50.46 | |
| SFPOModel=Qwen2.5-7B, k=3, BP=42026.05 | 66.65 | 49.45 | 88.52 | 20.64 | 19.11 | 48.87 | |
| GRPOModel=Qwen2.5-7B, k=–, BP=12026.05 | 66.56 | 48.78 | 88.43 | 20.91 | 19.48 | 48.83 | |
| GXPOModel=Qwen2.5-3B, k=10, BP=32026.05 | 59.36 | 32.73 | 80.98 | 17.39 | 15.64 | 41.22 | |
| GXPOModel=Qwen2.5-3B, k=5, BP=32026.05 | 58.34 | 30.89 | 80.21 | 17.27 | 15.39 | 40.42 | |
| SFPOModel=Qwen2.5-3B, k=10, BP=112026.05 | 57.7 | 30.89 | 79.46 | 16.95 | 15.29 | 40.06 | |
| GXPOModel=Qwen2.5-3B, k=3, BP=32026.05 | 57.59 | 30.65 | 79.47 | 17.35 | 15.42 | 40.1 | |
| SFPOModel=Qwen2.5-3B, k=5, BP=62026.05 | 57.02 | 30.94 | 79.02 | 17.23 | 15.42 | 39.93 | |
| GRPOModel=Qwen2.5-3B, k=–, BP=12026.05 | 56.74 | 30.13 | 78.73 | 16.91 | 14.91 | 39.48 | |
| SFPOModel=Qwen2.5-3B, k=3, BP=42026.05 | 56.44 | 31.07 | 78.94 | 17.52 | 15.25 | 39.84 | |
| GXPOModel=Llama3.2-3B, k=5, BP=32026.05 | 36.09 | 12.92 | 68.68 | 12.05 | 6.31 | 27.21 | |
| GXPOModel=Llama3.2-3B, k=3, BP=32026.05 | 34.85 | 11.09 | 68.14 | 11.17 | 6.19 | 26.29 | |
| SFPOModel=Llama3.2-3B, k=5, BP=62026.05 | 34.15 | 11.74 | 67.44 | 11.44 | 5.91 | 26.14 | |
| SFPOModel=Llama3.2-3B, k=3, BP=42026.05 | 33.51 | 11.28 | 66.99 | 11.12 | 5.98 | 25.77 | |
| GRPOModel=Llama3.2-3B, k=–, BP=12026.05 | 33.46 | 10.57 | 67.04 | 10.98 | 5.76 | 25.56 | |
| GXPOModel=Qwen2.5-1.5B, k=10, BP=32026.05 | 32.31 | 5.08 | 54.45 | 7.29 | 8.19 | 21.46 | |
| GXPOModel=Qwen2.5-1.5B, k=5, BP=32026.05 | 31.8 | 3.26 | 52.46 | 7.5 | 7.6 | 20.52 | |
| GXPOModel=Qwen2.5-1.5B, k=3, BP=32026.05 | 30.75 | 6.38 | 51.49 | 6.53 | 7.64 | 20.56 | |
| SFPOModel=Qwen2.5-1.5B, k=10, BP=112026.05 | 30 | 5.08 | 52.51 | 6.17 | 7.15 | 20.18 | |
| SFPOModel=Qwen2.5-1.5B, k=5, BP=62026.05 | 29.98 | 4.43 | 50.43 | 6.59 | 7.45 | 19.78 | |
| SFPOModel=Qwen2.5-1.5B, k=3, BP=42026.05 | 25.71 | 4.43 | 45.29 | 5.91 | 6.78 | 17.62 | |
| GRPOModel=Qwen2.5-1.5B, k=–, BP=12026.05 | 24.18 | 2.99 | 43.53 | 5.49 | 6.1 | 16.46 | |
| FlowTracerBackbone=Llama-3.1-8B2026.06 | 9 | 5.9 | 21.1 | 6 | 3.7 | 9.1 | |
| AttentionBackbone=Llama-3.1-8B2026.06 | 8.2 | 4.9 | 19.4 | 5.9 | 3.4 | 8.4 | |
| GRPOBackbone=Llama-3.1-8B2026.06 | 8.1 | 3.2 | 18.7 | 5.3 | 3.2 | 7.7 | |
| RandomBackbone=Llama-3.1-8B2026.06 | 7.8 | 4 | 17.6 | 5.2 | 3.1 | 7.5 | |
| High-entropyBackbone=Llama-3.1-8B2026.06 | 7.8 | 4 | 17.6 | 5.2 | 3.1 | 7.5 | |
| FlowTracerBackbone=Llama-3.2-3B2026.06 | 6.7 | 5.1 | 11.4 | 3.4 | 2.8 | 5.9 | |
| AttentionBackbone=Llama-3.2-3B2026.06 | 5.9 | 4 | 10.4 | 2.9 | 2.4 | 5.1 | |
| RandomBackbone=Llama-3.2-3B2026.06 | 5.8 | 3.5 | 9.1 | 3.2 | 2.1 | 4.7 | |
| GRPOBackbone=Llama-3.2-3B2026.06 | 5.4 | 3.8 | 9.5 | 3.2 | 2.2 | 4.8 | |
| High-entropyBackbone=Llama-3.2-3B2026.06 | 5.2 | 3.8 | 10 | 3.1 | 2.3 | 4.9 |