Detecting Soft Skills in ML Engineering Roles CVs

This paper investigates how soft skills are articulated in the CVs of ML engineers, data scientists, and software engineers, addressing a gap in the literature where most evidence comes from the demand side (job ads, surveys). The authors state: "Soft skills shape collaboration among ML engineers, data scientists, and software engineers building ML-enabled systems, yet what we know about them comes almost entirely from the demand side... How candidates themselves articulate these competencies has not been studied, and existing CV-mining work is both keyword-based, so it cannot see skills conveyed through narrative, and descriptive, reporting frequency rankings without testing whether group differences exceed sampling variation."

To close these gaps, the authors built a balanced corpus of 300 curated CVs (100 per role) by combining a previously established multi-labeled CV corpus with a supplementary set of contemporary CVs sourced from public web indexes. They developed an LLM-based extraction pipeline (using Gemini models) validated against a human-annotated ground truth, achieving a precision of 0.69, recall of 0.74, and F1-score of 0.72. The pipeline distinguishes between explicitly listed skills (in dedicated skills sections) and implicitly narrated skills (conveyed through professional experience descriptions). The authors state: "We then convert the demand-side literature’s claims into 13 falsifiable hypotheses about role signatures, seniority progression, and disclosure style, and test them with effect sizes under family-wise error control, so that candidate-side data can corroborate or contradict the demand-side account rather than merely illustrate it."

Key findings are summarized as follows: "Eleven hypotheses are supported, one partially, and one refuted. Candidates disclose soft skills through narrative rather than keyword lists by roughly three to one, and most so for the competencies employers value most: leadership, coordination, and mentoring (88–96% narrative). Seniority nearly triples the odds of articulating leadership. That competency, assumed universal in prior work, is articulated by software engineers at half the rate of their peers. Technical candidates do articulate soft skills, in a form keyword-based screening systematically misses."

More specifically, the results show:

• RQ1 (Role signatures): Collaboration (61.3%), leadership (38.3%), and communication (34.3%) are the most common soft skills overall. Data scientists emphasize communication (54.0% vs. 24.5% pooled, OR 3.60) and problem solving (34.0% vs. 15.0%, OR 2.90) more than other roles. ML engineers emphasize mentoring more (28.0% vs. 15.5%, OR 2.11). The no-skill rate differs across roles (18% ML, 16% SE, 4% DS). The hypothesis that leadership is role-invariant (H1e) was refuted: software engineers articulate leadership at roughly half the rate (26.0%) of data scientists (47.0%) and ML engineers (42.0%).

• RQ2 (Seniority): Using a five-year experience threshold (114 junior, 186 senior), the authors found that leadership mentions roughly double from junior to senior CVs (23.7% → 47.3%, OR 2.88), and this effect survives adjustment for role (adjusted OR 3.06). Coordination (16.7% → 31.2%, OR 2.26) and strategic thinking (5.3% → 14.0%, OR 2.92) also increase. The no-skill rate drops from 21.1% to 7.5%. Collaboration accumulates rather than being displaced by leadership (50.9% → 67.7%, OR 2.02). The seniority effect is statistically homogeneous across roles.

• RQ3 (Disclosure style): Candidates are nearly three times more likely to disclose soft skills through narrative than through keyword lists (1,007 implicit vs. 347 explicit skill–CV pairs). Only 3.3% of candidates rely exclusively on explicit keywords, while 59.0% communicate soft skills entirely through narrative. Narrative dominance is concentrated in high-value competencies: coaching (100% narrative), coordination (96.1%), mentoring (94.9%), presentation (90.0%), and leadership (87.8%). In contrast, communication (54.4% narrative), problem-solving (59.5%), and analytical ability (34.4%) are as likely or more likely to be keyword-listed. Disclosure style depends on role (H3b partially supported), with data scientists more likely to use both channels (44.0%).

The authors conclude: "the much-discussed soft skills gap in ML-related professions is, at least in part, an articulation gap. The competencies are present in candidate narratives but remain invisible to keyword-based screening, and invisible in a way that is systematically biased against the most discriminating evidence."

Improvements to AI systems:

1. Narrative soft-skill extraction: Build an AI system that parses free-text professional experience sections (not just keyword lists) to detect implicit soft skills via semantic and contextual cues, using LLM-based pipelines fine-tuned on the validated corpus (F1 0.72). This system can identify leadership, coordination, and mentoring from action verbs, project outcomes, and team interactions—skills that keyword search misses.

2. Role-aware skill profiling: Implement a classifier that adjusts soft-skill detection based on job role (ML engineer, data scientist, software engineer). The system can recognize that communication and problem-solving are more salient for data scientists, mentoring for ML engineers, and leadership is less common in software engineers—preventing false negatives and role-biased screening.

3. Seniority-sensitive skill inference: Use experience duration (e.g., years) as a feature to predict expected soft-skill progression. The system can infer that senior candidates are 3x more likely to articulate leadership and coordination, and adjust extraction thresholds or weighting accordingly—improving ranking for senior roles and reducing junior-candidate false positives.

4. Dual-channel disclosure integration: Create a hybrid extraction system that separately scores explicit keyword mentions and implicit narrative mentions, then combines them with a weighting scheme (e.g., narrative-heavy for leadership, keyword-heavy for analytical ability). This system can generate a more complete skill profile and flag candidates who rely solely on narrative—useful for recruiters using ATS.

5. Bias-corrected screening: Develop an AI screener that explicitly corrects for the articulation gap—where technical candidates under-disclose soft skills in keywords but express them in narratives. The system can re-rank candidates by narrative-derived skills, reducing systematic bias against software engineers and junior applicants, and improving fairness in hiring pipelines.

6. Hypothesis-testing skill ontology: Implement a structured skill taxonomy (13 falsifiable hypotheses) that the AI can use to validate or refute role-skill associations in real-time. This allows the system to generate explainable outputs (e.g., candidate shows leadership, consistent with senior ML engineer profile) and adapt to new data without retraining from scratch.

Soft skills shape collaboration among ML engineers, data scientists, and software engineers building ML-enabled systems, yet what we know about them comes almost entirely from the demand side. Job advertisements, surveys, and hiring manager interviews capture what employers ask for. How candidates themselves articulate these competencies has not been studied, and existing CV-mining work is both keyword-based, so it cannot see skills conveyed through narrative, and descriptive, reporting frequency rankings without testing whether group differences exceed sampling variation. We close both gaps. Using a balanced corpus of 300 curated CVs spanning the three roles, we extract explicitly listed and implicitly narrated soft skills with an LLM-based pipeline validated against a human-annotated ground truth, a distinction that existing extractors were not designed to make. We then convert the demand-side literature's claims into 13 falsifiable hypotheses about role signatures, seniority progression, and disclosure style, and test them with effect sizes under family-wise error control, so that candidate-side data can corroborate or contradict the demand-side account rather than merely illustrate it. Eleven hypotheses are supported, one partially, and one refuted. Candidates disclose soft skills through narrative rather than keyword lists by roughly three to one, and most so for the competencies employers value most: leadership, coordination, and mentoring (88-96% narrative). Seniority nearly triples the odds of articulating leadership. That competency, assumed universal in prior work, is articulated by software engineers at half the rate of their peers. Technical candidates do articulate soft skills, but a keyword-based screening systematically misses them.