Machine Translation on WMT 23 (English-to-Chinese)
89.79GRFDeepSeek-R1(w/ refl)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| DeepSeek-R1(w/ refl)Token=1029.68, reflection=true2026.04 | 89.79 | 1.93 | 81.04 | |
| ReflectMT-Stage2(Our)Token=30.672026.04 | 89.48 | 1.7 | 81.18 | |
| Qwen3-8B(w/ refl)Token=738.92, reflection=true2026.04 | 89 | 2 | 81.24 | |
| Qwen2.5-14B(w/ refl)Token=193.38, reflection=true2026.04 | 88.48 | 2.07 | 80.85 | |
| DeepSeek-R1Token=593.122026.04 | 88.37 | 2.06 | 79.39 | |
| Qwen3-8BToken=588.592026.04 | 88.2 | 2 | 81.18 | |
| ReflectMT-Stage1(Our)Token=28.912026.04 | 87.89 | 1.82 | 80.78 | |
| QwQ-32B(w/ refl)Token=1136.59, reflection=true2026.04 | 87.17 | 2.6 | 78.33 | |
| ReflectMT-SFT(Cold Start)Token=25.942026.04 | 86.01 | 2.21 | 80.35 | |
| QwQ-32BToken=632.562026.04 | 85.51 | 2.79 | 78.59 | |
| Marco-o1-7B(w/ refl)Token=725.51, reflection=true2026.04 | 85.34 | 2.26 | 80.03 | |
| Qwen-7B-Adapt-RL†Token=192.73, thought_mode=adaptive2026.04 | 83.7 | 1.73 | 80.77 | |
| Qwen2.5-14BToken=31.142026.04 | 83.69 | 2.17 | 80.67 | |
| Qwen-7B-Full-RL‡Token=201.31, thought_mode=full2026.04 | 82.89 | 1.74 | 80.71 | |
| Qwen-7B-Adapt-LoRAToken=150.152026.04 | 82.15 | 2.13 | 80.55 | |
| ExTrans-7BToken=1547.552026.04 | 81.83 | 2.17 | 78.96 | |
| DRT-7BToken=479.642026.04 | 79.7 | 2.54 | 78.87 | |
| Marco-o1-7BToken=548.342026.04 | 79.58 | 2.38 | 79.42 | |
| DeepTrans-7BToken=2018.962026.04 | 78.23 | 2.61 | 77.12 | |
| Qwen-7B-InstructToken=28.332026.04 | 75.54 | 2.65 | 78.83 | |
| Llama3.1-8BToken=31.712026.04 | 75.35 | 2.58 | 79.29 |