Fairness is increasingly elusive in the current landscape of Artificial Intelligence and Large Language Models. These technologies can easily inject fake or inaccurate information into the data, often misrepresenting what truly exists. This problem is widely spread in many domain applications, including those dealing with user profiles. In particular, in the job market, this afects both recruiters and job seekers. Resumes are frequently optimized to fit the job call in rather than to reflect genuine qualifications, while automated screening tools may overlook authentic but non-standard profiles. This work proposes a resume analysis and enhancement system. It enables iterative improvement through the use of Large Language Models while preserving the original content. This leads to a consistent improvement in similarity and match quality with job applications. Fairness is achieved not by altering who the candidate is, but by ensuring their actual capabilities are accurately and contextually recognized, thus empowering both evaluators and applicants through authentic enhancement.
restructure their narratives just to be considered by the system [6].
AI-driven resume screening systems, currently widely This work presents a resume analysis and
enhanceadopted in company recruitment scenarios, have rede- ment system designed around the principle of
"contexifned the process of candidate evaluation[ 1]. While these tual fairness" [6]. The system avoids modifying or
arsystems possess scalability and consistency, they often tificially enhancing a candidate’s narrative. Instead, it
prioritize standardization over content [
system is designed for high-throughput resume analysis,
ofering structured outputs that assist recruiters in
canTraditional Applicant Tracking Systems (ATSs) rely on didate filtering. Similar to our work, their approach uses
keyword-based filtering [ 9], which fails to capture the instruction-tuned LLMs for interpreting and processing
contextual nuances in resumes, leading to biased or inac- resume content. However, the two systems diverge
signifcurate candidate evaluations. Recent approaches lever- icantly in purpose and design philosophy. While Gan et al.
age transformer-based models to assess semantic similar- focus on classification and summarization to streamline
ity between resumes and job descriptions. Resume2Vec hiring pipelines, our system emphasizes "contextual
fair[9] introduced a framework using models like BERT, ness"—providing non-intrusive, section-wise suggestions
RoBERTa, and LLaMA [9] to generate embeddings and that retain the candidate’s narrative integrity. Instead of
improve candidate–job alignment through cosine simi- generating summaries or altering resume tone, our
syslarity. Their system outperformed conventional ATSs in tem enhances clarity and alignment using a hybrid model
both ranking accuracy and alignment with human judg- architecture: Sentence-Transformers
multi-qa-MiniLMment across multiple domains. Unlike keyword-centric L6-cos-v1[7] for semantic similarity scoring and LLaMA
methods, Resume2Vec emphasizes context and fairness 3.2[8] for targeted feedback. However other LLMs (i.e.
by preserving the semantic richness of candidate data. LLaMantino [
Lavi et al. (2021) [10] introduced conSultantBERT, a 3. Methodology ifne-tuned Siamese Sentence-BERT model tailored for resume-job matching, addressing challenges such as data Our framework follows a pipeline with consecutive steps heterogeneity, cross-linguality, and noisy resume for- Figure 1. Such pipeline begins by taking in two primary mats. By leveraging cosine similarity between multilin- inputs: the resume uploaded by the job seeker and the gual embeddings, their model significantly outperformed job description submitted by the recruiter. both TF-IDF and pre-trained BERT baselines in predicting resume-vacancy matches. Their findings afirm the Resume upload and processing. We support a reimportance of domain-specific fine-tuning to preserve sume uploading process for documents in Word (i.e. semantic integrity in candidate profiles while improving ".docx" extension) or PDF (i.e. ".pdf" extension) format. matching accuracy. Like our system, conSultantBERT em- The system uses python-docx 1 for Word documents and phasizes contextual matching without resorting to super- pdfplumber2 for PDFs. These libraries enable accurate ifcial keyword overlap, highlighting the role of semanti- extraction of plain text and preserves section structure cally grounded embeddings in achieving fair and scalable as well as formatting semantics. Each parsed resume is recruitment solutions. While conSultantBERT focuses stored in a document database (Firestore DB and Storage) on semantic matching between resumes and job descrip- alongside unique metadata including a resume identifier , tions using fine-tuned embeddings, our approach not user email, timestamp, and a designated resume name for only evaluates similarity but also provides customized future tracking and analysis purposes. resume enhancements using LLMs
Yadav et al. (2025) [
Gan et al. (2024) [
2https://pypi.org/project/pdfplumber/ 3Currently, such aspects are not automatically extracted from the job position but we consider to do that as a future work. Improvements/suggestions Get the fields: 1.resume text 2.job text 3. Matching terms 4.keyword score 5.semantic
score 6.ATS score Generate improved suggestion
Generate ATS improvements Llama 3.2: latest model
Suggestions are sentout to user
STOP
Enter Job description(JD)
fields Store in Database
Update changes
Firebase
Matching & missing terms Upload resume
PDF/DOCX PDF Plumber
Docx based extraction Store resume text along with identifiers in Database
Resume text and JD fields Extract keywords using KeyBERT Get missing terms & matching terms using RapidFuzz ATS score, semantic score & keyword score
GET RESUME/JD
similarity
Get resume text & JD, flatten the fields of JD feed to
encoder Encode resume text & JD text to tensor & use cosine similarity to calculate
similarity Resume text
JD text Pre process: text to lower text, regular expression based text cleaning Keyword match: extract keywords from resumes &
JD text using count vectorizer Semantic similarity score calculation: encode texts using all-MiniLM-L6-v2 & calculate cosine similarity ATS score=Sesmeamnaticnsticcore score*semantic weight+keyword score*keyword weight
Keywor d score for displaying details or generating questionnaires) and unstructured access (e.g., for semantic similarity and ATS scoring).
Data Cleansing. Both the resume and job descrip- and preprocessed, the system performs a multi-level tion texts are normalized by converting to lowercase and alignment assessment through semantic similarity and applying regular expression-based cleaning into alpha- keyword relevance scores. This step is central to producing numeric characters. This removes extraneous symbols, a fair and interpretable Applicant Tracking System (ATS) spacing irregularities, and control characters, ensuring score that reflects both explicit and contextual alignment input consistency before model encoding. This initial between candidate profiles and job requirements. acquisition and preparation phase ensures that both resumes and job descriptions are available in clean, com- Semantic Similarity Score. To ensure robustness parable formats for downstream tasks such as similarity and fairness in semantic evaluation, the system levercomputation, keyword relevance analysis, and improve- ages two independent transformer models from the ment suggestion generation. SentenceTransformers library 4:
• all-MiniLM-L6-v26
tracted resume keywords and job description keywords
is used to calculate a match ratio:
Each model independently encodes the cleaned resume
text and job description text into tensor embeddings. Co- match_score = |matched_keywords| (1)
sine similarity[
Applicant Tracking System (ATS) Score. The final Keyword Relevance Score. Keyword-based scoring ATS score is computed as a weighted sum of semantic complements semantic alignment by focusing on lex- similarity and keyword relevance scores: ical overlap. This scoring process follows the steps ATS_score = (sem_score · 1) + (keyword_score · 2) described below: (i) Initial Extraction. The job de- (2) scription is vectorized using CountVectorizer from Where: trsrekaclctteitoaenrrm.nR.effsrueeqmauteeuknrecyeyw_aeonxradtlsyrasaricse.te(ixiiot)rnaR.cetteseudxmutse7in,KgaelKyloewwyoBirnEdgRETdx8i--, •• 21==kseeymwaonrtdicwweeigighhtt(d(deeffaauultlt::00.5.5)) which identifies top N significant phrases based on con- This hybrid scoring formula balances surface-level textual embedding similarity. Keyword extraction is es- term relevance with deep contextual alignment. By assential and plays a vital role in ensuring fairness during signing separate weights, the system allows recruiters to evaluation. As shown in Figure 1 the extracted matching prioritize either direct skill inclusion or holistic candidatekeywords are utilized by the LLM to generate context- job compatibility. aware suggestions, providing targeted improvements The calculated ATS score serves as a crucial factor for that align more closely with the job description. This both recruiters and job seekers by helping recruiters efstep enhances both the relevance and fairness of the feed- ficiently shortlist candidates based on relevance, while back provided to users. guiding job seekers in optimizing their resumes. Unlike
Two diferent keyword extraction approaches are used traditional systems that rely solely on keyword matching, to account for the inherent diferences in data structure this score combines keyword relevance with semantic and consistency. Job descriptions are entered by users in similarity, capturing not just the presence of required a structured JSON format and are generally concise and terms but also the contextual alignment between the standardized, making them ideal for keyword extraction resume and job description. This hybrid approach enusing CountVectorizer, which captures raw term frequen- sures greater fairness, adaptability across domains, and cies. In contrast, resumes are uploaded as binary files reduced bias, making it more insightful than conven(PDF or DOCX) and converted to plain text, often in an tional ATS scores that often overlook phrasing variations unstructured and inconsistent manner - hence, KeyBERT or implied competencies. is employed to extract context-aware key phrases using To prevent artificial score inflation and preserve cansemantic embeddings, ensuring reliable keyword identi- didate authenticity, the system avoids injecting new keyifcation despite formatting noise or phrasing variability. words or altering the resume’s core content. Instead, it focuses on identifying and enhancing existing expressions—both semantically and lexically ensuring fairness to the job seeker while giving recruiters a transparent, accurate alignment signal. 5https://huggingface.co/sentence-transformers/ multi-qa-MiniLM-L6-cos-v1 6https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2 7https://scikit-learn.org/stable/modules/feature_extraction.html 8https://pypi.org/project/keybert/
To enhance specific resume sections by rephrasing, clarifying, or restructuring them using best practices in resume writing we design an enanchement step grounded on Large Language Models (LLMs) (i.e., generate_improved_sections_with_llm). It focuses on strengthening the candidate’s input by: • Reinforcing matched keywords in previous ATS
Score Calculation step • Improving clarity and formatting of the resume • Highlighting quantifiable impacts and actiondriven phrasing actions • Focusing on improving sections like Professional
Summary, Experience, Skills, and Education A structured prompt is generated (see Table 1), including the resume text, job description, and the list of already matched keywords. The LLM is explicitly instructed not to introduce missing or hallucinated terms, ensuring that improvements remain factual and grounded in the candidate’s original input. The model returns suggestions in strict JSON format, each linked to a specific resume section for traceable integration. We provide the LLM with the flattened job description and resume text, along with the matchingTerms, similarityScore, and atsScore, to give it fuller context for generating accurate, traceable suggestions. We pass the flattened job description and flattened resume text, matchingTerms, similarityScore, atsScore are also passed so that we are giving more context to the LLM model - LLaMA 3.2 latest.
Field Resume Text JobDescription MatchedKeywords ExplicitInstructions OutputFormat
Description Extracted, cleaned resume text uploaded by the user.
Flattened string of title, skills, experience, and role.
Key terms found in both the resume and job description.
Directs model to avoid hallucination and ensure factual edits only.
JSON array with fields: sectionName, suggestion.
The second service we designed, generate_ats_score_and_improvements, operates at a global resume level rather than focusing on specific sections. It analyzes the entire resume in the context of the job description and the list of matched keywords but deliberately avoids altering content or injecting new, unverified terms. Instead, it identifies opportunities for structural and stylistic enhancements that can improve ATS performance without compromising authenticity.
Its outputs include: • Parsing the resume text and job description. • Evaluating aspects like formatting consistency (e.g., bullet points, section headers), action verb usage, and sentence clarity. • Referencing the matched keywords to ensure better usage and placement, rather than adding unrelated terms. • Returning the output in a strict JSON structure, which includes: (i) A list of factual, actionable suggestions; (ii) An estimated ATS score; (iii) Highlighted areas where improvements can be made to enhance readability and alignment.
Fairness and Transparency Considerations. Both services are governed by strict instruction constraints to: • Prevent hallucination of unverified skills • Avoid inflating match quality with artificial edits • Respect candidate identity and experience as originally stated
able content, this dual-LLM framework ensures that suggestions are ethical, transparent, and aligned with fair AI principles - providing job seekers with meaningful improvement pathways without compromising truthfulness.
This experiment evaluates whether resumes can be ethically enhanced to better align with job descriptions, without introducing fabricated content or misleading embellishments. The objective is to test whether a candidate’s original experience and qualifications can be made more contextually relevant while preserving the integrity and authenticity of the resume. A representative set of 10 manually crafted synthetic resumes (refer Tables 3, 4 in appendix and for column name descriptions refer Table 5) were selected and evaluated against a curated synthetic job description for the role ReactJS Frontend how our approach upholds fairness by restricting edits Developer (API Integration & UI Frameworks) using four to what the candidate’s own language can support. The key metrics: similarity score, ATS score, matching terms, experiment confirms that our system provides significant and missing terms. These metrics were computed against improvements while maintaining fairness, i.e., enhanctarget job descriptions which were manually crafted by ing the resume without misrepresenting the candidate’s analyzing real listings for similar job roles. While indi- skills or experience. vidual scores may vary, the relative diferences (score deltas) remain consistent across resumes. The resumes 4.2. Experiment 2: Efectiveness were then enhanced using our LLM-powered suggestion Comparison engine, which provides section-wise recommendations based solely on the candidate’s original content and job This experiment compares the efectiveness of two rerelevance. Enhanced resumes were re-evaluated with sume enhancement strategies, both operating under strict the same approach previously used, for observing: (i) non-hallucination constraints. The first method uses Increases in similarity and ATS scores; (ii) Growth in our domain-specific LLM-powered suggestion engine contextually valid matching terms; (iii) Retention of se- to improve resume-job alignment while preserving the mantic integrity (i.e., no direct insertion of previously candidate’s original intent and language. The second missing terms unless already implied). method uses a general-purpose GPT-4o model instructed to rewrite resumes without adding any content not origiFairness Criteria. To ensure ethical enhancement, the nally present. Each resume was evaluated in three forms: system followed three key constraints: the original version, a system-enhanced version using our custom enhancement engine, and a GPT-enhanced (GPT-4o)10 version rewritten by a large language model under strict non-hallucination instructions. All three versions were analyzed using the same backend evaluation pipeline (refer Figure 1) to compute similarity score, final ATS score, semantic similarity score, and keyword match score. • All newly introduced terms had to be contextually
consistent with the original resume. • Terms from the initial missing terms list were disallowed unless semantically implied or rephrased from existing content. • No artificial keyword stufing or hallucinated ex
periences were permitted.
The improvements were evaluated using changes in matching terms and missing terms metrics computed by comparing keyphrases from the job description with the resume text before and after enhancement (refer Table 2). These metrics served as our primary quantitative evidence, ensuring that enhancements improved alignment without introducing unrelated or fabricated content, as the suggestion engine operated strictly within the resume’s original context.
Experimental results. The system-enhanced resumes consistently outperformed the original versions in all key metrics. The summary of ATS and Similarity Scores (in %) Across Resume Enhancement Systems can be seen in Table 6. On average, similarity scores improved by 18.7% and ATS scores rose by 22.3% following enhancement. When comparing system-enhanced resumes to GPT-enhanced counterparts, our method achieved higher average similarity scores (43.52% vs. 34.13%) and comparable semantic similarity scores (46.77% vs. 47.21%), Experimental results. Following enhancement using despite the GPT-enhanced versions showing a higher fiour system, all resumes demonstrated meaningful im- nal ATS score (74.25%). However, a deeper inspection of provements while preserving fairness and integrity. New the results reveals that the elevated ATS scores in GPTmatching terms were successfully added in every case, enhanced resumes may be attributed to broader keyword and all additions were contextually aligned with the coverage rather than meaningful contextual alignment. original resume content. Crucially, none of the origi- In Figure 1, once the suggestions from our system are nal missing terms were directly reused, and no hallu- updated, a parallel process generates and applies sugcinated or unrelated information was introduced (re- gestions using Chat GPT-4o as well. Both updated verfer Table 3 and 4 in appendix). The outcomes of Ex- sions, the one based on our system’s suggestions and the periment 1, which involved evaluating ten candidate one generated from GPT-4o’s recommendations, are then resumes for the ReactJS Frontend Developer position, re-evaluated. A comparison spreadsheet is generated are summarized in Tables 3 and 4 in appendix. While containing the results of both evaluations, highlighting both LLaMA 3.2 and GPT-4o raise the overall match diferences in ATS scores, similarity scores, and overcounts, the New_Terms_Added_by_LLaMA3.2 column all improvements. The ATS scores and similarity scores grows only with fair, semantically grounded additions. In comparison across enhancement Systems can be seen contrast, New_Terms_Added_by_GPT-4o reflects GPT4o’s blind injections of extra keywords—demonstrating 10https://openai.com/index/hello-gpt-4o/ in Table 6. The system-enhanced resumes maintained though GPT-4o’s outputs show notable improvements a more focused and candidate-authentic tone while still over the unmodified baseline, they still fall short of the improving discoverability (refer Table 3 and Table 4 in Ap- results achieved by our system. This supports the efecpendix). In multiple cases, the system-enhanced versions tiveness of a fairness-oriented framework that prioritizes outperformed GPT in similarity score by margins exceed- refining existing content rather than introducing extraing 16 percentage points, with the highest observed gain neous terms. reaching 29.07% (refer Table 6). In Figure 3 the bar chart compares ATS scores for ten
Figure 2 shows that the augmented similarity scores resumes across three conditions: the original unmodi(system_updated_similarityScore) markedly exceed both ifed documents (blue bars), the LLaMA 3.2-based update the baseline (original_similarityScore) and the GPT-4o methodology (red bars), and GPT-driven enhancements derived scores (chatgpt4o_updated_similarityScore), in- (green bars). LLaMA 3.2 updates yield the highest imdicating that our LLaMA 3.2–based methodology predi- provements boosting scores from approximately 18–25 at cated on conservative, in situ enhancement of existing baseline to 35–55, whereas GPT enhancements produce text yields the most substantial improvements in seman- moderate gains, raising baseline values to roughly 26–48. tic alignment between resumes and job descriptions. Al- In every case, the LLaMA 3.2–adjusted resumes outperform both the original and GPT-enhanced versions, with strict adherence to defined fairness criteria ensures the the latter still delivering a substantial uplift relative to tool’s suitability for real-world applications where ethical unmodified resumes. standards are paramount.
Furthermore, the system’s enhancements did not
introduce any hallucinated content and preserved the re- Comparative Efectiveness and Contextual
Alignsume’s original structure and voice (refer Table 3 and Ta- ment. Experiment 2’s comparative analysis between
ble 4 in Appendix). In contrast, GPT-enhanced rewrites, our system and GPT-based enhancements further
rewhile constrained, occasionally drifted toward general- inforces the strengths of our approach. While
GPTized language or tone inconsistencies. These observa- enhanced resumes sometimes achieved higher ATS
tions reinforce the value of targeted, context-aware en- scores—likely due to broader keyword coverage—the
hancement over generalized rewriting approaches. system-enhanced resumes consistently showed superior
or comparable semantic similarity scores, indicating a
5. Considerations and Limitations closer contextual match to the original resumes.
This distinction is important: higher ATS scores alone
The results from our experiments highlight the eficacy do not guarantee a better quality or more truthful resume.
and robustness of the proposed AI-powered resume en- The tendency of GPT-based rewrites to introduce
generhancement system, especially in terms of fairness, con- alized language or tone inconsistencies could dilute the
textual integrity, and practical relevance for applicant candidate’s unique profile, potentially reducing perceived
tracking systems (ATS). authenticity. In contrast, our system’s targeted,
contextaware enhancements retain the original voice and
structure, ofering improvements that are both meaningful
Fairness and Authenticity Preservation. Experi- and aligned with the candidate’s actual background. The
ment 1 demonstrated that our system can meaningfully observed margin of improvement in similarity scores (up
enhance resumes by adding relevant matching terms to 29.07 percentage points over GPT in some cases)
sugwithout compromising fairness or authenticity. The fact gests that our method excels at fine-grained semantic
that none of the original missing terms were reused and enhancement rather than broad-stroke rewriting. This
no hallucinated or unrelated information was introduced focused approach is likely to yield better candidate-job
is particularly encouraging. This shows that the sys- matching outcomes in ATS environments that value
pretem respects the candidate’s true skills and experiences, cise and relevant keyword and phrase usage.
avoiding unethical exaggeration or fabrication—a critical Additional limitations include the need for improved
requirement in AI-assisted recruitment tools. The aver- performance in domain-specific contexts, sensitivity to
age improvements of 18.7% in semantic similarity and input formats, and the lack of multilingual support.
Ethi22.3% in ATS scores indicate that the enhancements not cal concerns around bias, transparency, and resume
overonly preserve but also amplify the relevance of candidate optimization also warrant future exploration. Ensuring
profiles to job descriptions, improving their discoverabil- fairness, explainability, and data privacy in deployment
ity without sacrificing honesty. environments will be crucial to responsible adoption [
This balance between enhancement and fairness is a While the system shows promising results, some areas key diferentiator compared to many automated systems merit further attention. Current performance is strongest that risk introducing biases or misrepresentations. The on English-language resumes with consistent formatting; improving support for varied layouts and multilingual inputs is a valuable direction. Our evaluation, centered on synthetic resumes for a specific domain (Frontend ReactJS), provides a solid foundation but would benefit from broader validation across job types and real-world data. Additionally, while basic bias detection is included, more comprehensive fairness auditing remains an important avenue for future development. As with all LLMenhanced systems, results may vary slightly based on the quality of job description inputs. Addressing these aspects can help increase the system’s robustness, fairness, and generalizability.
This project demonstrates that our AI-powered resume enhancement system efectively improves resume quality while upholding fairness and authenticity. By preserving resume integrity—without adding fabricated keywords or skills—the system consistently adds contextually relevant terms, resulting in substantial improvements in semantic similarity (18.7%) and ATS scores (22.3%). Compared to GPT-based rewrites, our approach achieves higher or comparable semantic alignment while maintaining the candidate’s original voice and structure, avoiding generalized or inconsistent language. These findings highlight the advantage of targeted, context-aware enhancement methods that responsibly boost candidate discoverability and preserve authenticity. Consequently, our LLM-based enhancement system ofers a practical, ethical, and superior solution for real-world recruitment pipelines. Future improvements could include support for multilingual resumes and enhanced robustness for unstructured or poorly formatted inputs.
react, developer, exper- reactjs, frontend, back- reactjs, frontend, react, frontend, component, reactjs, frontend, react, functional, accessibility, tise, apis, redux, ui, end, skilled, freelance, developer, expertise, development, reactjs, backend, developer, performant, flexbox, dejest, scss, api, experi- axios, frameworks, apis, redux, compo- integrations, scalable, axios, frameworks, type- sign, chrome, devtools, ence, candidate, man- typescript, es6, com- nents, component, ui, building, quality, ap- script, expertise, apis, integrations, handling, aging, university, com- ponents, component, jest, development, ux, plications, integrating, redux, es6, components, github, axios, interfaces, puter, form, grid, sci- agile, development, ux, integrations, applica- components, like, de- component, ui, jest, git, vanilla, integration, ence, user, high, rest, github, flexbox, inte- tions, api, building, liver, ux, integration agile, development, ux, frontend, component, state, years, best, con- grations, web, design, experience, integration, github, flexbox, inte- environment, styling, tract, hooks, time, hook applications, javascript, integrating, scalable, grations, scss, design, reactjs, scalable, quality, responsive, bachelor, managing, deliver, applications, javascript, agile, teams, hands, interfaces, formik, boot- university, computer, api, responsive, inter- ensuring, practices, ux, strap, fetch, building, form, grid, science, user, faces, bootstrap, fetch, tailwind, validation, devtools, integration, high, rest, state, years, building, experience, integrating, modern, integrating, css, scal- best, contract, quality, devtools, integration, in- token, typescript, buildable, layouts, initiative, time, like tegrating, css, scalable, ing, layouts, backend, responsibilities, collabo- layouts, collaborate, authentication, perforrate, degree, applying, ensuring, functional, mance, frameworks, ensuring, functional, chrome, managing, like, bootstrap, testing, chrome, router, looking, performance, hands, javascript, context, performance, hands, authentication, vanilla, gitlab, responsive, css, authentication, vanilla, gitlab, deliver, envi- development, collabogitlab, deliver, envi- ronment, university, rate, query, es6, fetch, ronment, tailwind, tailwind, computer, applications, based, styling, token, startup, styling, form, grid, components, deliver, accessibility, git, work- token, accessibility, git, library ing, handling, testing, science, user, handling, practices, teams, mod- high, testing, practices, ern, based, library, teams, rest, state, negotiable, performant, modern, years, based, quality, implementa- best, library, hooks, tion, validation, query, performant, quality, include, like, ideal, time, hook, validation, equivalent, context query, like, context Original_Matching_Terms Original_Missing_Terms LLaMA3.2_Matching_Terms New_Terms_Added_by_LLaMA3.2 GPT-4o_Matching_Terms New_Terms_Added_by_GPT-4o
Description The set of job-description keywords that already appeared in the candidate’s resume before any edits.
Keywords required by the job but absent from the unmodified resume.
After our LLaMA 3.2 “in-place” enhancement, this column lists all keywords in the resume that now match the job description—combining the original matches with those preserved by conservative rewriting.
Of the matches in the previous column, these are the new terms introduced by LLaMA 3.2. Crucially, each is semantically equivalent to language already used by the candidate.
The total set of matched keywords after GPT-4o editing—again including both originally present terms and those retained or reordered by GPT.
The new keywords injected by GPT-4o. Unlike our method, these often include terms that were not semantically aligned with the candidate’s original phrasing. resume_2_7 resume_2_1 resume_2_2 resume_2_9 resume_2_4 Declaration on Generative AI During the preparation of this work, the author(s) used ChatGPT (OpenAI), Gemini (Google), and Grammarly in order to: Paraphrase and reword, Improve writing style, and Grammar and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.