Span-level error detection on WMT25 English-Italian (official human annotations)
10.2PrecisionGemini-3-Flash
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Gemini-3-FlashError Filtering=Filtered2026.05 | 10.2 | 21.96 | 13.93 | |
| Gemini-3-FlashError Filtering=Unfiltered2026.05 | 10.13 | 22.1 | 13.89 | |
| Gemma-3-27b-itError Filtering=Filtered2026.05 | 9.82 | 11.47 | 10.58 | |
| Gemma-3-27b-itError Filtering=Unfiltered2026.05 | 9.78 | 11.69 | 10.65 | |
| EuroLLM-9B-InstructError Filtering=Filtered2026.05 | 4.44 | 7.16 | 5.48 | |
| EuroLLM-9B-InstructError Filtering=Unfiltered2026.05 | 4.35 | 7.43 | 5.49 | |
| Qwen3-VL-30B-A3B-InstructError Filtering=Filtered2026.05 | 4.32 | 23.42 | 7.3 | |
| Qwen3-VL-30B-A3B-InstructError Filtering=Unfiltered2026.05 | 3.03 | 25.75 | 5.42 |