Automating Resume Tailoring Using Large Language Models : A Comparative Study of Model Performance

Sammanfattning

In today’s competitive job market, job applications must be tailored to align with specific job listings to maximise the applicant’s chances. However, manually tailoring applications is time-consuming, and for consulting firms, this time quickly adds up, potentially limiting the number of opportunities they can pursue. This thesis addresses this problem by implementing a system that automates the tailoring process using Large Language Models, specifically OpenAI’s GPT-4o-mini and GPT-4.1 models. The system, which relies solely on prompt engineering, takes an existing resume and a job listing as input and produces a tailored resume as well as a motivational statement. This output is then evaluated in terms of quality, time efficiency, and cost-effectiveness. Results demonstrate that the system can generate tailored resumes that match the overall quality of manually tailored versions and even outperform in less personal sections, such as the skills and job experience sections. However, the motivational statements produced by the system were rated lower than those written by humans, and generated summaries also scored lower. These findings suggest that, although Large Language Models can streamline the process, human oversight remains essential for ensuring accuracy and personal relevance. Furthermore, the GPT-4.1 model did not outperform GPT-4o-mini in this use case, despite being approximately twelve times more expensive and having higher theoretical capabilities. This suggests that lower-cost models may still meet performance requirements in many practical scenarios. The study demonstrates that LLMs effectively support resume writing, and a hybrid AI-and-human approach will likely deliver the best results. For the consulting firm Boulder, this system enables faster tailoring of resumes while maintaining quality and accuracy, ultimately increasing responsiveness to job opportunities. Future work could explore how more advanced prompt strategies or fine-tuned models might improve personalisation without compromising accuracy.