Which resume extraction model handles resumes with hidden text and prompt injections best?
14 of the 300 resumes are resumes with hidden text and prompt injections. The typical system scores 1.4 points higher on them. Best: LlamaParse Extract Agentic Plus, 93.2 F1. LlamaParse Extract Agentic is within two points for $23 per 1,000 pages.
Resumes carrying instructions to the parser in white, 1-point or invisible text, or in the PDF metadata, plus text layers that do not match the page. A good parser ignores them. Best document API: LlamaParse Extract Agentic Plus (93.2). Best vision LLM: Claude Opus 4.8 (90.8). LlamaParse Extract Cost-Effective ranks 6 places higher here than overall.
Every system on resumes with hidden text and prompt injections
F1 on the 14 resumes with hidden text and prompt injections, against each system’s F1 on all 300 resumes. Cost and time are for all resumes.
| Model | F1 here | F1 overall | Gap | Per 1,000 pages | Median time |
|---|---|---|---|---|---|
| LlamaParse Extract Agentic Plus | 93.2 | 87.2 | +6.1 | $54 | 70 s |
| LlamaParse Extract Agentic | 91.6 | 85.5 | +6.0 | $23 | 55 s |
| Claude Opus 4.8 | 90.8 | 85.5 | +5.3 | $90 | 31 s |
| LlamaParse Extract Cost-Effective | 90.5 | 83.5 | +7.0 | $8.18 | 29 s |
| Reducto Deep Extract | 90.1 | 85.6 | +4.5 | $40 | 65 s |
| Extend Max Context | 89.6 | 84.7 | +4.9 | $50 | 43 s |
| Qwen 3.8 Flash | 89.0 | 84.1 | +4.9 | $2.98 | 122 s |
| Reducto Extract | 87.2 | 81.7 | +5.5 | $20 | 16 s |
| Gemini 3.5 Flash | 86.7 | 85.5 | +1.2 | $35 | 43 s |
| Extend Extract | 82.4 | 81.0 | +1.5 | $12 | 31 s |
| Qwen 3.8 27B | 82.4 | 83.8 | −1.4 | $9.59 | 36 s |
| Qwen 3.5 35B | 79.2 | 77.9 | +1.3 | $3.91 | 78 s |
| GPT-5.6 | 77.5 | 84.7 | −7.2 | $74 | 38 s |
| Qwen 3.5 9B | 70.2 | 76.0 | −5.9 | $1.31 | 138 s |
| GPT-5.5 | 69.1 | 83.5 | −14.4 | $87 | 48 s |
| Gemma 4 26B | 48.6 | 62.8 | −14.3 | $0.56 | 29 s |
| GLM 5.3 Flash | 46.4 | 59.2 | −12.8 | $0.80 | 46 s |
| GPT-5.4 Nano | 22.0 | 61.9 | −39.9 | $2.76 | 20 s |
What to do about it
Hidden instructions are a real attack on automated screening. Never let parsed text act as instructions downstream.
Published examples
2 public resumes of this kind.