Tested: Same CS Paper with LaTeX Formulas, 4 Tools
An arXiv reinforcement learning paper (8 pages): 23 LaTeX formulas, 5 data tables, pseudocode. Four tools tested on formula preservation.
Four Tools
PDFTranslate (GLM-4.5-Air), DeepL Pro, Google Translate, Mathpix Snipping Tool + DeepL (formulas extracted and processed separately).
Results
| Tool | Formula Accuracy | Table Layout | Citation Handling | Overall Readability |
|---|---|---|---|---|
| PDFTranslate | 8.0 | 8.5 | 7.0 | 8.0 |
| Mathpix + DeepL | 9.0 | 5.0 | 6.0 | 6.5 |
| DeepL Pro | 5.0 | 3.0 | 6.0 | 4.5 |
| 4.0 | 2.0 | 5.0 | 3.5 |
Surprise Finding
Mathpix + DeepL scored highest on formula accuracy (9/10) but lowest on overall readability (6.5). Why: after Mathpix extracts formulas separately, the correspondence between formula and body text is lost. Reinserting post-translation caused positional misalignment throughout.
Conclusion: if formula accuracy is the primary goal, Mathpix + DeepL wins. For full document translation (formula + body + tables + citations), PDFTranslate direct is the more practical choice.