arXiv study: LLM pipeline finds soft skills hidden in CV narratives

A new arXiv paper (2608.10046) from researchers Aidan Azamnouri, Nouran Ayad, Justus Bogner, and Stefan Wagner tackles a blind spot in AI hiring: how ML engineering candidates actually communicate soft skills on their CVs. Prior work relied on job ads and surveys, and existing CV-mining tools were keyword-based, missing skills expressed through narrative. The team built an LLM-based extraction pipeline that distinguishes explicitly listed skills from implicitly narrated ones, validating it against human-annotated ground truth on a balanced corpus of 300 curated CVs spanning machine learning engineers, data scientists, and software engineers.

The results are striking. Candidates disclose soft skills through narrative rather than keyword lists by roughly 3 to 1, and the ratio is even higher for the competencies employers value most: leadership, coordination, and mentoring appear in narrative form 88–96% of the time. The researchers converted demand-side claims into 13 falsifiable hypotheses and tested them with effect sizes under family-wise error control: 11 were supported, one partially, and one refuted. Notably, seniority nearly triples the odds of articulating leadership, while software engineers articulate leadership at half the rate of their peers. The takeaway: technical candidates do express soft skills, but traditional keyword-based screening systematically misses them, suggesting hiring pipelines that rely on simple matching are filtering out qualified applicants.