Table-based Reasoning and Grounding on HighlightBench Real-world
81.3T1: Markup Grounding AccuracyGemini2.5-Flash
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Gemini2.5-FlashSubset=Real-world2026.03 | 81.3 | 75.3 | 62.7 | 72.2 | 78.5 | 75.3 | |
| OursSubset=Real-world2026.03 | 72.9 | 74.2 | 70.8 | 82.2 | 79.8 | 77.1 | |
| Claude Sonnet 4Subset=Real-world2026.03 | 56.7 | 59.6 | 42.2 | 60.7 | 68.2 | 58.7 | |
| Qwen3-VL-8BSubset=Real-world2026.03 | 56.3 | 57.1 | 24.2 | 36.3 | 78 | 48.4 | |
| MiniCPM-V 4.5Subset=Real-world2026.03 | 56.2 | 50 | 15.5 | 35.9 | 35.7 | 42.8 | |
| InternVL3.5-8BSubset=Real-world2026.03 | 54.2 | 53.1 | 23 | 36.6 | 67.7 | 46.3 | |
| Qwen2.5-VL-3BSubset=Real-world2026.03 | 39.4 | 35.6 | 1.2 | 31.3 | 58.9 | 34.7 | |
| MiniCPM-V 2.6Subset=Real-world2026.03 | 17.2 | 29 | 9.9 | 17.5 | 3.4 | 17.5 | |
| LLaVA1.5-7BSubset=Real-world2026.03 | 3.4 | 1.5 | 1.2 | 0.2 | 0 | 1.4 |