Long-generation on LongBench
96.75Sequence Length (Sl)LARFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LARFTBase Model=Qwen2.5-7B-Instruct2026.02 | 96.75 | 82.09 | |
| LARFTBase Model=Llama-3.1-8B-Instruct2026.02 | 94.75 | 62.41 | |
| RLBase Model=Qwen2.5-7B-Instruct2026.02 | 93.84 | 75.72 | |
| RLBase Model=Llama-3.1-8B-Instruct2026.02 | 90.38 | 58.88 | |
| LARFTBase Model=Llama-3.2-3B-Instruct2026.02 | 88.71 | 54.78 | |
| LARFTBase Model=Qwen2.5-3B-Instruct2026.02 | 88.34 | 73.61 | |
| RLBase Model=Llama-3.2-3B-Instruct2026.02 | 88.25 | 48.23 | |
| RLBase Model=Qwen2.5-3B-Instruct2026.02 | 87.23 | 69.49 | |
| SFT+RLBase Model=Llama-3.2-3B-Instruct2026.02 | 83.88 | 54.17 | |
| RulerBase Model=Qwen2.5-7B-Instruct2026.02 | 83.77 | 76.72 | |
| PositionIDBase Model=Qwen2.5-7B-Instruct2026.02 | 83.14 | 77.59 | |
| SFTBase Model=Qwen2.5-7B-Instruct2026.02 | 82.04 | 80.83 | |
| SFTBase Model=Llama-3.1-8B-Instruct2026.02 | 80.87 | 64.84 | |
| SFTBase Model=Qwen2.5-3B-Instruct2026.02 | 79.88 | 66.04 | |
| SFT+RLBase Model=Qwen2.5-3B-Instruct2026.02 | 79.46 | 66.49 | |
| SFT+RLBase Model=Qwen2.5-7B-Instruct2026.02 | 79.46 | 86.07 | |
| RulerBase Model=Llama-3.2-3B-Instruct2026.02 | 79.35 | 51.68 | |
| RulerBase Model=Llama-3.1-8B-Instruct2026.02 | 79.35 | 64.47 | |
| PositionIDBase Model=Llama-3.1-8B-Instruct2026.02 | 79.15 | 66.13 | |
| OriginBase Model=Qwen2.5-7B-Instruct2026.02 | 78.89 | 78.55 | |
| RulerBase Model=Qwen2.5-3B-Instruct2026.02 | 77 | 63.73 | |
| OriginBase Model=Qwen2.5-3B-Instruct2026.02 | 76.38 | 68.49 | |
| PositionIDBase Model=Qwen2.5-3B-Instruct2026.02 | 74.07 | 63.11 | |
| SFTBase Model=Llama-3.2-3B-Instruct2026.02 | 73.09 | 47.36 | |
| PositionIDBase Model=Llama-3.2-3B-Instruct2026.02 | 72.35 | 51.49 | |
| SFT+RLBase Model=Llama-3.1-8B-Instruct2026.02 | 71.26 | 63.43 | |
| OriginBase Model=Llama-3.1-8B-Instruct2026.02 | 64.75 | 62.9 | |
| OriginBase Model=Llama-3.2-3B-Instruct2026.02 | 57.73 | 48.9 |