Which resume extraction model handles hard resumes best?
Handwriting, scans, designed templates, hidden text and long careers each cost parsers accuracy, and not the same parsers. Each page ranks every system on one kind of resume.
| Kind of resume | # of resumes | Median system loses | Best here | F1 |
|---|---|---|---|---|
| Handwritten resumes | 9 | 31.7 points | LlamaParse Extract Agentic Plus | 67.8 |
| Scanned and image-only resumes | 28 | 10.5 points | LlamaParse Extract Agentic Plus | 84.2 |
| Resumes with custom sections | 105 | 4.4 points | LlamaParse Extract Agentic Plus | 85.1 |
| Resumes with many entries | 83 | 2.9 points | Extend Max Context | 85.6 |
| Resumes with long bullet lists | 101 | 2.8 points | LlamaParse Extract Agentic Plus | 84.9 |
| Designed resumes (Canva and LaTeX) | 32 | 0.7 points | LlamaParse Extract Agentic Plus | 86.4 |
| Rotated and warped scans | 21 | 0.4 points | LlamaParse Extract Agentic Plus | 92.5 |
| Resumes with hidden text and prompt injections | 14 | 1.4 points | LlamaParse Extract Agentic Plus | 93.2 |