How Large Language Models "Read" a Résumé

You upload a carefully designed PDF. An AI returns a score, a list of strengths, and suggested rewrites. But what actually happened between upload and output? If you want to understand how AI reads a resume, you need to look past the interface at tokenisation, context windows, and inference — the machinery behind LLM resume understanding.

It is not reading — it is predicting

A large language model does not "look" at your CV the way a recruiter does. It receives a sequence of tokens — fragments of words and punctuation — and predicts what comes next based on patterns learned from vast amounts of text. When analysing a résumé, the model is asked targeted questions: What skills are demonstrated? Does experience match this job description? What is weak about this bullet?

The answers feel intelligent because language models are good at language. They are not inherently good at facts. Without guardrails, they can confabulate skills you never claimed. That is why production AI CV analysis systems ground responses in extracted text rather than trusting open-ended generation.

Step 1: Tokenisation

Before any model sees your CV, the text is split into tokens. "Project management" might become two tokens; "C++" might become one or three depending on the tokenizer. This matters because:

• Unusual spellings and symbols can fragment awkwardly
• Acronyms only match job descriptions if they appear in the token stream
• Typos become different tokens, which can reduce match scores

Tokenization resume quirks explain why spelling out "Search Engine Optimisation (SEO)" once helps both humans and models: you cover multiple token patterns in one phrase.

Step 2: Text extraction (the hidden bottleneck)

Tokenisation assumes you have clean text. PDF extraction runs first — and it is often the weakest link. Multi-column layouts read in the wrong order; headers disappear; scanned pages need OCR. The LLM only analyses what extraction delivers. For more on this pipeline, see the state of AI in CV parsing.

Step 3: Context windows

Models have a finite context window — the amount of text they can consider at once. A CV plus a long job description plus instructions plus output schema consumes tokens fast. Very long CVs (four pages of dense academia, for example) may require chunking: the model analyses sections separately and synthesises, which can miss cross-section connections.

Practical implication: concise, focused CVs often analyse more accurately than sprawling ones — not because length is bad, but because more fits in context intact.

Step 4: Inference and prompting

The model receives a prompt — instructions plus your CV text plus optionally the job description. The prompt shapes behaviour: score against rubric dimensions, suggest rewrites, build a requirements matrix. Temperature and sampling settings affect consistency. Production systems use low randomness for scoring and may run extraction and evaluation as separate steps to reduce hallucination.

Why structure still matters

LLMs infer structure from text, but clear structure reduces errors. Compare:

Weak structure: A wall of paragraphs mixing three jobs without dates or headings.

Strong structure: Employer name, title, dates on their own lines; bullets underneath each role.

The second format helps the model attribute achievements to the correct employer — critical when screening for "5+ years in fintech" or similar requirements.

Where LLMs hallucinate

Hallucination in CV analysis usually means inventing skills, inflating titles, or citing bullets that do not exist. It happens when:

• The model is asked open-ended questions without source grounding
• Extraction dropped text and the model fills gaps plausibly
• Job-description pressure causes over-attribution of partial matches

Cvaluate mitigates this with a two-step pipeline — extract first, evaluate second — so scoring and rewrites reference verified content. Read more in our post on why Cvaluate uses a two-step AI pipeline.

A bullet the model can work with

Before: Helped with marketing.

After: Owned email campaigns for 42k subscribers; lifted open rates from 19% to 27% in two quarters.

The model can match the "after" bullet to digital marketing requirements, extract a metric, and suggest further improvements. The "before" bullet gives it almost nothing reliable to reason about.

Where to go next

For the candidate-facing overview, start with our guide on AI in hiring and your job search. Compare legacy systems in old-school ATS vs modern AI screening. When you want to see inference on your own document, run your CV through Cvaluate's free analysis.