Simultaneous Machine Translation on ACL60/60 En-Ja (eval)
100BLEUReference
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| ReferenceTuning method=–, Tuning data=–, Inference prompt=–2026.01 | 100 | 100 | 0 | 0.9721 | 0.8447 | 3.298 | |
| Step-wise (Drop / Cut)Tuning method=–, Tuning data=–, Inference prompt=Incremental drop / cut selection2026.01 | 55.33 | 34.49 | 99.06 | 0.8913 | 0.8447 | 2.613 | |
| Step-wise (Actions)Tuning method=–, Tuning data=–, Inference prompt=Incremental action selection2026.01 | 55.31 | 34.33 | 101.42 | 0.8998 | 0.8439 | 2.524 | |
| TransLLaMA (Reference)Tuning method=Fine-tuning, Tuning data=Dev set (Reference), Inference prompt=–2026.01 | 54.88 | 30.48 | 98.58 | 0.8789 | 0.8429 | 3.085 | |
| TransLLaMA (Action)Tuning method=Fine-tuning, Tuning data=Dev set (Action-adapted refs), Inference prompt=–2026.01 | 54.08 | 30.69 | 97.72 | 0.8837 | 0.8351 | 2.927 | |
| TransLLaMA (Salami)Tuning method=Fine-tuning, Tuning data=Dev set (Salami refs), Inference prompt=–2026.01 | 53.12 | 30.11 | 98.35 | 0.8735 | 0.8313 | 2.938 | |
| Few-shot (Reference)Tuning method=Prompt tuning, Tuning data=Dev set (Reference), Inference prompt=Few-shot demonstrations2026.01 | 52.91 | 35.02 | 104.72 | 0.8934 | 0.8423 | 3.146 | |
| Few-shot (Action)Tuning method=Prompt tuning, Tuning data=Dev set (Action-adapted refs), Inference prompt=Few-shot demonstrations2026.01 | 52.78 | 35.08 | 107.31 | 0.8913 | 0.8428 | 2.915 | |
| Few-shot (Salami)Tuning method=Prompt tuning, Tuning data=Dev set (Salami refs), Inference prompt=Few-shot demonstrations2026.01 | 52.27 | 34.93 | 108.25 | 0.8926 | 0.8379 | 3.049 | |
| SalamiTuning method=Fine-tuning, Tuning data=Dev set (Salami refs), Inference prompt=Salami prompt2026.01 | 50.16 | 31.98 | 106.37 | 0.8854 | 0.8295 | 2.955 |