CV parsing technology converts unstructured CV text into structured data fields using one of three methods: rules-based extraction, machine-learning classification, or LLM agents. Each one reads a CV and pulls out details such as name, work history, education, and skills. They differ in how they read the document and how well they cope with messy, real-world layouts.
Rules-based parsers follow hand-written patterns. Machine learning parsers learn from labelled examples. LLM agents read context and meaning, and they can summarise and grade a CV, not just extract from it. Newer approaches tend to be more accurate on awkward CVs, but not always, and accuracy is not the only thing that matters. In a 2024 peer-reviewed study, an LLM-agent framework classified CV content with an F1 score of 87.7% while running about 11 times faster than manual screening. This guide explains what each approach is, how accurate each one is, where each one fails, and what a recruiter or agency owner should look for when choosing a tool.
This is a primer on the technology, not the workflow. For where parsing fits in an agency's day, see CV parsing for recruiting, and for a plain walk-through of the parsing steps, see CV parsing explained for recruiters. If you just want a tool that gets this right, RefineCV parses CVs and rebuilds them into clean branded templates. Try it free with 10 CVs, no card.
Key takeaways
- CV parsing technology falls into three generations: rules-based systems, machine learning, and large language models. Each reads a CV and turns it into structured data, but they handle messy layouts very differently.
- Rules-based parsers follow hand-written patterns and keywords. They are fast and predictable, but they break on any layout the rules did not anticipate.
- Machine learning and large language model parsers learn from data instead of fixed rules, so they cope better with unusual wording and formats.
- LLM agents go a step further than plain parsers. They can summarise and grade a CV, not just extract from it. In a 2024 peer-reviewed study, an LLM-agent framework classified CV content with an F1 score of 87.7% while running about 11 times faster than manual screening.
- Newer does not always mean more accurate. In one 2025 study, a plain deep-learning baseline scored well below a tuned commercial parser, while large language models scored highest. Design matters more than the label.
- Accuracy is not the only thing to weigh. Large language models can be slower and more costly per CV, and AI scores do not match human judgement, so a person should still review the output.
What every CV parser does
Every parser, whatever technology sits underneath it, does the same basic job. It reads the text of a CV, works out which part is the work history and which part is the skills section, and pulls named details into separate fields your system can search. Software reads a document left to right and top to bottom, uses standard section headings like Experience, Education, and Skills to sort the content, and cannot pull text from images or graphics unless it reads them first. Those mechanics are the same across all three generations.
Getting this right is not a small detail. In a 2021 Harvard Business School and Accenture study, 88% of employers agreed that qualified high-skills candidates are vetted out of the hiring process because they do not match the exact criteria set in the job description. When your client's system parses a CV you submitted, a parsing failure can drop a strong candidate for a formatting reason, not a skills reason. The technology behind the parser is what decides how often that happens.
The three generations of CV parsing technology
The three approaches are best understood as generations, each solving a weakness of the one before. Many real tools mix them, for example using rules to clean up the output of a model.
1. Rules-based parsers (keyword and grammar)
The oldest approach. A developer writes patterns and rules by hand that describe how a CV usually looks, for example "job title, then company, then a date range". Keyword lists and regular expressions match the text against those patterns. Rules-based parsers are fast, cheap to run, and predictable: the same CV always parses the same way. Their weakness is that they only handle what the rules anticipated. A creative heading, an unusual date format, or a two-column layout the rules did not expect can cause a field to come back empty or wrong.
2. Machine learning and NLP parsers
Instead of fixed rules, these parsers learn from thousands of labelled CVs. Models such as conditional random fields and, more recently, transformer-based classifiers, work out the probability that a piece of text is a job title, a skill, or an employer. A technique called named entity recognition pulls out specific details like company names and dates. Because they learn patterns rather than follow hand-written ones, they cope better with wording and formats the developer never saw. They need good training data to work well, and a poorly trained model can still miss things a simple rule would have caught.
3. Large language model (LLM) parsers
The newest generation. Large language models read a CV the way they read any text, using context and meaning rather than fixed patterns. That lets them tell "Python" the programming language from "Python" in a company name, and infer structure even when a CV has no clear headings. On awkward, non-standard CVs, this tends to be the most accurate approach. The trade-offs are speed and cost: a large model can take several seconds per CV and costs more to run than a rules-based parser, and its output still needs checking.
| Approach | How it reads a CV | Strength | Weak spot |
|---|---|---|---|
| Rules-based | Hand-written patterns and keywords | Fast, cheap, predictable | Breaks on layouts the rules did not expect |
| Machine learning / NLP | Learns from labelled example CVs | Handles varied wording and formats | Only as good as its training data |
| Large language model | Reads context and meaning | Best on messy, non-standard CVs | Slower and more costly per CV |
How accurate is each approach?
How much more accurate are the newer approaches? A 2025 study by Zhu and colleagues tested several parsers on the same set of real CVs and measured field-level F1, a score from 0 to 1 where higher means more fields extracted correctly. A commercial traditional parser scored 0.82. A large language model reading the same CVs scored 0.92, and a layout-aware LLM pipeline reached 0.96. A plain deep-learning baseline in the same test scored just 0.49, well below the traditional parser.
That last number is the important one. It shows that "AI" is not automatically more accurate. A well-tuned traditional parser beat a poorly designed deep-learning model by a wide margin. The design of the system matters more than which generation it belongs to. The averages also hide something important: a parser does not fail evenly across a CV, so the next section looks at where each approach breaks.
| System tested | Approach type | Field-level F1 |
|---|---|---|
| Deep-learning baseline | Machine learning | 0.49 |
| Commercial parser | Traditional | 0.82 |
| LLM with OCR | Large language model | 0.92 |
| Layout-aware LLM pipeline | Large language model | 0.96 |
For a concrete example on real CVs, we ran our own LLM-based parser through a field-by-field faithfulness test and published every score, including which fields it missed: see how accurate CV parsing is in practice.
Where each parser fails
The averages hide where parsers actually break. Two things decide whether a field comes back clean: the layout of the CV, and the type of field being read. Here is where each generation struggles, using per-field results from the same 2025 study.
| CV feature | Why it is hard | Where it breaks |
|---|---|---|
| Multi-column layout | About 20% of real CVs use a non-linear, multi-column layout that breaks top-to-bottom, left-to-right reading order | Rules-based parsers and plain LLMs scramble the content. A layout-aware step is what fixes it. |
| Long descriptions | Job and project text runs across many sentences and is hard to pull out cleanly | The commercial parser scored just 0.50 F1 here, against 0.88 on dates. The deep-learning baseline could not read these fields at all. |
| Scanned or image CVs | The text sits inside an image, so it has to be read as an image first | No parser works without optical character recognition (OCR) to turn the picture into text. |
| Dates and short fields | Short, regular patterns like date ranges are predictable | Handled well even by older parsers. Dates were the commercial parser's strongest field group at 0.88 F1. |
The pattern is clear. A parser fails on the messy, human parts of a CV: multi-column sidebars and free-text descriptions. The same 2025 study shows this plainly. A plain large language model reading only OCR text scored 0.55 on long descriptions, but adding a layout-aware step that rebuilds the reading order raised the same model to 0.85. Stripping the layout and OCR preprocessing back out of the full system cost around 10 points of accuracy on those fields. The lesson repeats: how a system handles layout and OCR matters more than which generation it belongs to.
How LLM agents are used in recruitment
An LLM agent is more than a parser. A plain parser extracts fields and stops. An LLM agent can read the whole CV, summarise it, grade it against a role, and even recommend who to interview. This is the newest way large language models are being applied in recruitment, and it is where the research is most active.
The clearest published example comes from a 2024 study in the peer-reviewed Journal of Information Processing. Gan, Zhang, and Mori built a framework where LLM agents summarise and grade each CV in a large batch, then decide which candidates to offer or interview. After fine-tuning, their model classified CV content with an F1 score of 87.7% and ran about 11 times faster than manual screening. The fine-tuned model also beat a GPT-3.5 baseline on the summarising and grading step.
The field has kept moving. A 2025 study tested an ensemble where several fine-tuned models each read a different section of the CV, then combined their answers by weighted voting, which improved extraction over the best single model. The direction is the same: split the work, then combine, rather than asking one model to read everything in one pass. Speed and scores like these are real, but they are not the whole story, as the next section shows.
For a wider view of where AI helps and where it does not across hiring, see our balanced overview of AI in recruitment.
The limits of AI parsing
The newer the technology, the easier it is to over-trust it. Two limits are worth keeping in mind.
First, speed and cost. In the Zhu study, the plain large language model was accurate but slow, taking roughly 20 seconds per CV against about 1.5 seconds for the traditional parser. At volume, that difference in speed and running cost is real, which is why the best systems combine techniques rather than throwing a large model at every document.
Second, AI is not a substitute for human judgement. A 2025 study by Vaishampayan and colleagues compared a large language model's ratings of 736 real applicants against human recruiters' ratings. The model's scores correlated only weakly with the humans', meaning the two are not interchangeable. Careful prompting improved the match, but did not remove the gap. A parser can extract and even score a CV, but a recruiter should still read the result before it reaches a client.
How to evaluate a CV parsing tool
You do not need to know how a parser is built to judge it. Focus on how it performs on your work and how much control you keep over the result.
What to check before you commit
- Ask which technology sits underneath, and how it handles a two-column or table-heavy CV, since that is where older parsers fail
- Check that it reads scanned or image-based CVs, which need optical character recognition before any parser can work
- Confirm you can review and correct the extracted fields before anything is sent, because no parser is perfect
- Test it on your own messy CVs, not a clean sample, so you see real accuracy on the documents you actually receive
- Weigh speed and cost per CV if you format at volume, not just headline accuracy
- Check the range of input formats it accepts, so you are not stuck converting files by hand first
One more thing to keep straight, because tools blur it. Parsing pulls data out of a CV into fields. Formatting builds a clean, branded document to send to your client. They are different jobs. A parser can read a CV perfectly and still leave you with raw data you then have to lay out by hand. Your real task as a recruiter is submission, so you need both: accurate parsing to capture the details, and formatting to turn them into a document a client will read. The best tools do both in one step and let you check the result.
This is the standard RefineCV's CV formatting software is built to. It parses CVs, reads scanned files with OCR, accepts a wide range of input formats, and rebuilds each CV into a clean branded template that you review before export. See transparent pricing, or read our comparison of the best CV formatting software for agencies.
Frequently asked questions
What is CV parsing technology?
CV parsing technology is the software that reads a candidate's CV and extracts the details into structured fields such as name, contact details, work history, education, and skills. It turns a document into organised data your system can search and store. The technology behind it falls into three main types: rules-based parsers that follow hand-written patterns, machine learning parsers that learn from labelled examples, and large language models that read context and meaning.
What are the three types of CV parsing technology?
The three generations are rules-based, machine learning, and large language model parsers. Rules-based parsers use hand-written keyword patterns and are fast but brittle. Machine learning and natural language processing parsers learn from thousands of labelled CVs, so they cope better with varied formats. Large language models read a CV using context and meaning, which makes them strongest on messy, non-standard layouts but slower and more costly to run.
Is AI or LLM CV parsing more accurate than older parsers?
Usually, but not automatically. In a 2025 study by Zhu and colleagues, a commercial traditional parser scored 0.82 field-level F1 while large language models scored 0.92 to 0.96 on the same CVs. However, a poorly designed deep-learning baseline in the same test scored just 0.49, below the traditional parser. The design of the system matters more than which generation it belongs to, so test any tool on your own CVs rather than trusting the label "AI".
How are LLM agents used in recruitment?
LLM agents go beyond extracting fields. They can read a whole CV, summarise it, grade it against a role, and recommend who to interview. In a 2024 study in the peer-reviewed Journal of Information Processing, an LLM-agent framework classified CV content with an F1 score of 87.7% and ran about 11 times faster than manual screening. The results are strong, but LLM ratings still correlate only weakly with human recruiters, so a person should review the output before it reaches a client.
What is the difference between CV parsing and CV formatting?
Parsing and formatting are two different jobs. CV parsing reads a CV and extracts the details into structured fields such as name, work history, and skills. CV formatting takes those details and builds a clean, branded document to send to a client. A parser can capture the data perfectly and still leave you to lay it out by hand. As a recruiter you usually need both: parsing to capture the details, and formatting to produce the client-ready CV.
Can AI CV parsing replace human recruiters?
No. A 2025 study by Vaishampayan and colleagues compared a large language model's ratings of 736 real applicants against human recruiters and found the scores correlated only weakly, meaning the two are not interchangeable. AI parsing is good at pulling data out of a CV quickly, but it makes mistakes and its judgement does not match a person's. A recruiter should always review the extracted fields before a CV goes to a client.
What should I look for in a CV parsing tool?
Test it on your own messy CVs, not a clean sample. Check how it handles two-column and table-heavy layouts, whether it reads scanned or image-based files with optical character recognition, and whether you can review and correct the extracted fields before anything is sent. If you format at volume, weigh speed and cost per CV alongside accuracy, and check the range of input formats it accepts so you are not converting files by hand.
The bottom line
CV parsing technology comes in three generations: rules-based parsers that follow hand-written patterns, machine learning parsers that learn from labelled data, and large language models that read context and meaning. Newer approaches usually parse messy CVs more accurately, but the evidence is clear that design matters more than the label. A well-built older parser can beat a weak AI model, and even the best large language model is slower, costs more, and still needs a human to check its work. When you evaluate a tool, ignore the marketing and test it on the CVs you actually receive.
Related reading: CV parsing explained for recruiters and the difference between an ATS and a recruitment CRM.
Skip the parsing guesswork
RefineCV parses CVs with OCR support, rebuilds them into clean branded templates, and lets you review every field before export. Start free with 10 CVs, no credit card. Then $0.40 per CV, or $50 per month for 200.
Sources
- Zhu, Yu, Chen, et al., "Layout-Aware Parsing Meets Efficient LLMs: A Unified, Scalable Framework for Resume Information Extraction and Evaluation", arXiv (October 2025): Field-level F1 on real CVs for a commercial traditional parser (0.82), a large language model with OCR (0.92), a layout-aware LLM pipeline (0.96), and a deep-learning baseline (0.49); the plain LLM took around 20 seconds per CV versus about 1.5 seconds for the traditional parser. About 20% of resumes used non-linear, multi-column layouts. Per field, the commercial parser scored 0.50 F1 on long descriptions against 0.88 on dates, and the deep-learning baseline did not support long-text fields. A plain LLM on OCR text scored 0.55 on long descriptions, rising to 0.85 with layout-aware reordering; removing the layout and OCR fusion steps cost around 10 points on those fields.
- Gan, Zhang, and Mori, "Application of LLM Agents in Recruitment: A Novel Framework for Resume Screening", Journal of Information Processing, Vol. 32 (peer-reviewed) (2024): An LLM-agent framework where agents summarise and grade each CV and decide which candidates to offer or interview classified CV content with an F1 score of 87.7% after fine-tuning, ran around 11 times faster than manual screening, and beat a GPT-3.5 baseline on summarising and grading.
- Walid, Younes, Shaban, Hassan, and Hamdi, "MSLEF: Multi-Segment LLM Ensemble Finetuning in Recruitment", arXiv (September 2025): An ensemble in which several fine-tuned models each read a different CV section and are combined by weighted voting improved recruitment resume extraction over the best single model.
- Vaishampayan, et al., "Human and LLM-Based Resume Matching: An Observational Study", Findings of the Association for Computational Linguistics: NAACL 2025 (2025): Across 736 real applicants, a large language model's zero-shot ratings correlated only weakly with human recruiters' ratings, meaning the two are not interchangeable; prompt design narrowed but did not remove the gap.
- Fuller, Raman, Sage-Gavin, and Hines, "Hidden Workers: Untapped Talent", Harvard Business School Project on Managing the Future of Work and Accenture (September 2021): 88% of employers agreed that qualified high-skills candidates are vetted out of the hiring process because they do not match the exact criteria set in the job description.
- University of Minnesota Duluth Career Center, "Applicant Tracking System (ATS) Tips" (Accessed 2026): Parsers read a document left to right and top to bottom, rely on standard section headings such as Education, Experience, and Skills to sort content, and cannot read text from graphics or images without OCR.