How accurate is resume extraction, really?
The best system on ResumeBench, LlamaParse Extract Agentic Plus, scores 87.2 F1; the median system scores about 83.5. Most entries come back right, but few resumes come back perfect.
What the score means for a recruiter
| Model | F1 | Resumes with nothing wrong | Resumes at 90 F1 or more |
|---|---|---|---|
| LlamaParse Extract Agentic Plus | 87.2 | 46 of 300 | 141 of 300 |
| LlamaParse Extract Cost-Effective | 83.5 | 26 of 300 | 99 of 300 |
| GLM 5.3 Flash | 59.2 | 23 of 300 | 74 of 300 |
The scoring is strict: every field and bullet is checked word for word, no partial credit. Misread the company and you can lose the whole job. So even a score in the 80s means someone should check the output.
One job entry, parsed by every system
Brand Storyteller / Evangelist · TechVenture, from a published resume. Each field and bullet is right or wrong; there is no partial credit.
| Field | Ground truth | LlamaParse Extract Agentic Plus | GLM 5.3 Flash |
|---|---|---|---|
| company | TechVenture | right | wrong: Velocity Labs |
| position | Brand Storyteller / Evangelist | right | wrong: Head Developer |
| bullet 1 | TechVenture is a developer-tools company (not a VC firm despite the … | right | wrong: Velocity Labs was a tech-driven dev shop (client care contractor), '… |
| bullet 2 | Produced 40+ technical blog posts, 6 conference talks, and 3 video t… | right | wrong: (empty) |
| bullet 3 | Represented TechVenture at KubeCon NA 2018 and JSConf US 2019; booth… | right | wrong: (empty) |
| bullet 4 | Ghost-wrote two bylined articles for the CEO published in TechCrunch… | right | wrong: (empty) |
9 of 18 systems got this entry entirely right. The rest missed bullet, company, position.
Where accuracy drops
Averages hide the hard cases. Handwriting, scans and designed templates cost every system points: see which parser handles hard resumes best.