PDF-to-LaTeX reconstruction on TEXOCR-Bench 1.0 (test)
83.4Section Accuracy (SA)GPT-5.3
Evaluation Results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT-5.3Model Type=Proprietary2026.04 | 83.4 | 86 | 65.2 | 78.2 | 72.4 | 98.8 | 82.7 | 84.6 | 69.2 | 57.9 | 91.1 | 72.7 | 78.5 | |
| Infinity-Parser-7BModel Scale=7B2026.04 | 77.5 | 3.7 | 17.6 | 32.9 | 68.9 | 96.5 | 44.5 | 70 | 76.8 | 66.1 | 80.8 | 74.6 | 59.1 | |
| TEXOCRTraining Protocol=SFT + RLVR, Base Model=Qwen3-VL-2B2026.04 | 76.7 | 85.9 | 86.8 | 83.1 | 64.7 | 95.2 | 45.2 | 68.4 | 72.8 | 58.4 | 89.2 | 73.5 | 75 | |
| TEXOCRTraining Protocol=SFT, Base Model=Qwen3-VL-2B2026.04 | 73.5 | 74.5 | 74.1 | 74 | 61.8 | 91.8 | 44.3 | 66 | 69 | 53.3 | 88 | 70.1 | 70 | |
| Qwen3-VL-8BModel Scale=8B2026.04 | 72.9 | 35.6 | 9.7 | 39.4 | 66.1 | 98.7 | 59 | 74.6 | 77.2 | 62.2 | 77 | 72.1 | 62.2 | |
| Mistral-Small-3.1-24BModel Scale=24B2026.04 | 68.9 | 43.2 | 44.6 | 52.2 | 60.4 | 98.4 | 39.1 | 66 | 62.1 | 47.4 | 79.3 | 62.9 | 60.4 | |
| Qwen2.5-VL-32BModel Scale=32B2026.04 | 68 | 15.9 | 8.6 | 30.8 | 62.1 | 96.2 | 55.8 | 71.4 | 67.6 | 62.7 | 81.9 | 70.7 | 57.6 | |
| Qwen3-VL-4BModel Scale=4B2026.04 | 64.2 | 17.3 | 8.5 | 30 | 63.1 | 98.2 | 58.4 | 73.2 | 69.1 | 58.7 | 52.6 | 60.1 | 54.5 | |
| Qwen3-VL-32BModel Scale=32B2026.04 | 61 | 62.6 | 42.9 | 55.5 | 67 | 98.3 | 58.9 | 74.7 | 75.7 | 63.8 | 88.9 | 76.1 | 68.8 | |
| InternVL3-38BModel Scale=38B2026.04 | 58.3 | 21.1 | 10.4 | 29.9 | 40.8 | 92.3 | 39.6 | 57.6 | 26.3 | 38.2 | 61.3 | 41.9 | 43.1 | |
| InternVL2.5-38BModel Scale=38B2026.04 | 54.6 | 13 | 8.1 | 25.2 | 45.1 | 88.9 | 42.7 | 58.9 | 36.9 | 35 | 43.1 | 38.3 | 40.8 | |
| Qwen2.5-VL-7BModel Scale=7B2026.04 | 53.9 | 4.8 | 9.2 | 22.6 | 52.3 | 94.6 | 58.2 | 68.4 | 56.8 | 53.2 | 51.1 | 53.7 | 48.2 | |
| Qwen2.5-VL-3BModel Scale=3B2026.04 | 52.2 | 1.3 | 0.9 | 18.1 | 43.2 | 76.9 | 39.5 | 53.2 | 43.2 | 43.8 | 45.2 | 44.1 | 38.5 | |
| Pixtral-12BModel Scale=12B2026.04 | 52.1 | 4.8 | 2.1 | 19.7 | 50.8 | 92.7 | 64.5 | 69.3 | 47.9 | 42.7 | 62.8 | 51.1 | 46.7 | |
| InternVL3-8BModel Scale=8B2026.04 | 51.9 | 17 | 3.6 | 24.2 | 49.1 | 93.1 | 35.1 | 59.1 | 46.6 | 37.6 | 50.4 | 44.9 | 42.7 | |
| Qwen3-VL-2BModel Scale=2B2026.04 | 49.6 | 21.8 | 1.5 | 24.3 | 50.3 | 97.9 | 57.4 | 68.5 | 67.3 | 55.2 | 68.8 | 63.8 | 52.2 | |
| olmOCR-2-7BModel Scale=7B2026.04 | 43.7 | 0.5 | 0.2 | 14.8 | 65.9 | 96.1 | 36.5 | 66.2 | 74.4 | 60.5 | 49.3 | 61.4 | 47.5 | |
| InternVL2-8BModel Scale=8B2026.04 | 36.9 | 5.1 | 1.6 | 14.5 | 24.4 | 61.5 | 24.4 | 36.8 | 6 | 28 | 20.8 | 18.3 | 23.2 | |
| InternVL2.5-8BModel Scale=8B2026.04 | 35 | 5.8 | 2 | 14.3 | 26.2 | 73.3 | 36.8 | 45.4 | 7.4 | 28.1 | 20.8 | 18.8 | 26.2 | |
| LLaVA-OneVision2026.04 | 26.4 | 2.2 | 0.6 | 9.7 | 22.7 | 65.4 | 46.1 | 44.7 | 6.1 | 26.8 | 12.6 | 15.2 | 23.2 | |
| Phi-3.5-vision2026.04 | 22.5 | 1.2 | 1.1 | 8.3 | 15.6 | 87.7 | 44.9 | 49.4 | 2.5 | 26.6 | 4.8 | 11.3 | 22.9 | |
| Phi-4-multimodal2026.04 | 18.6 | 7.7 | 6.2 | 10.8 | 19.2 | 98.3 | 56.8 | 58.1 | 1.5 | 26.6 | 5.4 | 11.2 | 26.7 | |
| DeepSeek-OCR2026.04 | 4.1 | 0.3 | 0 | 1.5 | 33.8 | 94.7 | 50.1 | 59.5 | 44.3 | 48.3 | 2 | 31.5 | 30.8 |