Span-level error detection on WMT25 English-Ukrainian (official human annotations)
4.58PrecisionGemini-3-Flash
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Gemini-3-FlashError Filtering=Filtered2026.05 | 4.58 | 24.27 | 7.71 | |
| Gemini-3-FlashError Filtering=Unfiltered2026.05 | 4.57 | 24.72 | 7.72 | |
| Gemma-3-27b-itError Filtering=Unfiltered2026.05 | 3.77 | 11.37 | 5.67 | |
| Gemma-3-27b-itError Filtering=Filtered2026.05 | 3.76 | 11.11 | 5.61 | |
| EuroLLM-9B-InstructError Filtering=Filtered2026.05 | 2.3 | 8.41 | 3.61 | |
| EuroLLM-9B-InstructError Filtering=Unfiltered2026.05 | 2.28 | 8.89 | 3.63 | |
| Qwen3-VL-30B-A3B-InstructError Filtering=Filtered2026.05 | 1.84 | 20.65 | 3.39 | |
| Qwen3-VL-30B-A3B-InstructError Filtering=Unfiltered2026.05 | 1.51 | 23.43 | 2.84 |