← Papers

Workers who pair specialised expertise with cross-domain skills — those on the 'diversity frontier' — learn faster, gain promotions and move into lower-automation roles; specialising or generalising alone is less adaptive.

Navigating the skill diversity frontier: How skill complexity explains worker resilience

Mar Carpanelli, Jedrzej Duszynski, Fabian Stephany · August 03, 2026

arxiv correlational medium evidence 8/10 relevance Full text usable extracted full text Source PDF

Structured author observations

Linked only from stored provider relations; the raw author line above is never matched by name.

Arxiv

Latest observation: August 4, 2026

Mar Carpanelli unresolved corpus identity
Jedrzej Duszynski unresolved corpus identity
Fabian Stephany unresolved corpus identity

Semantic Scholar

Latest observation: August 24, 2026

M. Carpanelli provider ID
J. Duszyński provider ID
F. Stephany provider ID

Using LinkedIn data and a network-derived 'diversity frontier', the paper shows that workers who combine specialised downstream skills with cross-domain diversity (i.e., are near the frontier) are more likely to acquire new skills, be promoted, and transition into occupations with lower automation exposure than peers with similar specialisation but narrower portfolios.

Citation observations

Cumulative provider counts captured on specific dates; providers are never combined.

0 cumulative citations

OpenAlex · Observed August 10, 2026

View corpus context

0 cumulative citations

Semantic Scholar · Observed August 24, 2026

View corpus context

As artificial intelligence transforms labor markets, understanding what makes workers adaptable has become increasingly important. Existing approaches typically characterize human capital using occupations, educational credentials, or predefined skill taxonomies, providing limited insight into how the structure of workers' skill portfolios shapes resilience to technological change. We develop an agnostic network based framework that reconstructs the hierarchy and diversity of skills directly from observed patterns of skill co occurrence. Using longitudinal data on 2.4 million United States workers and 16,753 distinct skills from LinkedIn, we introduce three complementary measures of skill complexity: specialisation, capturing productive depth; diversity, capturing adaptive breadth; and the diversity frontier, measuring the highest attainable diversity conditional on a worker's level of specialisation. We show that these dimensions predict distinct career outcomes. Specialisation is most strongly associated with sorting into higher wage occupations, whereas diversity is associated with broader skill accumulation and occupational mobility. Workers closest to the diversity frontier are significantly more likely to acquire new skills, receive promotions, transition into occupations with lower exposure to automation than workers with comparable levels of specialisation but narrower skill portfolios. These findings distinguish productive from adaptive capital and demonstrate that workers' adaptive capacity depends not simply on possessing specialised expertise or broad capabilities, but on combining both. More broadly, our framework provides a data driven approach for measuring workforce resilience and identifying reskilling pathways, offering new tools for understanding human capital in rapidly changing labor markets.

Summary

Main Finding

Workers’ resilience to technological change is best predicted not by specialisation or breadth alone but by how much cross-domain diversity they achieve given their level of specialisation — their position on a “diversity frontier.” Specialisation (deep, downstream skills) predicts sorting into higher‑wage occupations and promotions (productive capital). Diversity (breadth across skill domains) predicts broader skill accumulation, occupational mobility and AI‑skill adoption (adaptive capital). Workers near the diversity frontier are substantially more likely to acquire new skills, be promoted, and move into occupations with lower automation exposure than peers with similar specialisation but narrower portfolios.

Key Points

Conceptual decomposition:
- Productive capital = specialisation (depth in downstream/ domain‑specific skills).
- Adaptive capital = diversity (breadth across distinct skill domains).
- Diversity frontier = the maximum attainable diversity conditional on a given level of specialisation (inverse‑U relationship).
Empirical patterns:
- Specialisation → higher wages / promotion probability.
- Diversity → greater skill accumulation, occupational mobility, and adoption of AI‑related skills.
- Frontier position → strongest predictor of adaptive outcomes (learning, mobility into lower‑automation roles).
Interpretation: resilience depends on combining depth and breadth optimally, not on maximizing either alone.
Caveats: findings are predictive associations from observational LinkedIn data, not causal estimates.

Data & Methods

Data:
- LinkedIn longitudinal sample of ≈2.4 million U.S. workers who listed ≥5 skills prior to wide generative‑AI diffusion (pre‑2022).
- Skill universe: 16,753 distinct skills.
- Individual‑year panel structure. Explanatory variables measured 2020–2022; outcomes measured 2023–2025 to reduce reverse causality.
Skill network construction:
- Build directed skill network from observed co‑occurrence across profiles.
- Use a hypergeometric null model to identify statistically meaningful co‑occurrences.
- Infer edge directionality from asymmetric conditional probabilities (to recover upstream → downstream relationships); prune weak edges for sparsity and interpretability.
- Apply Louvain community detection on the undirected projection to identify skill domains.
Worker‑level measures:
- Specialisation: based on local reaching centrality (LRC) of skills in the directed network. LRC locates skills in an upstream–downstream hierarchy; worker specialisation = average of (1 − normalized LRC) across their skills (higher = more specialised/downstream).
- Diversity: Shannon/Hill diversity of skills across Louvain domains (effective number of domains represented), controlling for portfolio size.
- Diversity frontier: estimate the empirical upper envelope of attainable diversity conditional on specialisation (shows inverse‑U). Worker frontier position = how close the worker’s diversity is to that envelope for their specialisation.
Outcomes analyzed:
- Sorting into higher‑wage occupations, promotions, skill acquisition (including AI skills), transitions into occupations with lower automation exposure.
Empirical strategy:
- Regressions predicting outcomes from specialisation, diversity, and frontier position, controlling for portfolio size and other covariates (details and robustness checks in appendices). Emphasis on predictive associations, not causal claims.

Implications for AI Economics

Measurement and forecasting:
- Skill‑network derived measures (specialisation, diversity, frontier position) provide fine‑grained predictors of which workers are resilient to AI‑driven disruption beyond occupation/education categories. Useful for forecasting vulnerability and mobility potential.
Policy and reskilling design:
- Reskilling programs should aim not just to deepen or to broaden skills but to move workers toward feasible cross‑domain combinations given their specialisation (i.e., help workers reach the diversity frontier).
- Targeting workers with moderate specialisation for cross‑domain upskilling may yield large adaptive returns.
Labor market dynamics and inequality:
- Productive capital (depth) yields immediate labour‑market rewards (wages/promotions), while adaptive capital (breadth/frontier position) supports long‑run mobility and protection against automation — policies should balance short‑term earnings and long‑term resilience.
Firm strategy:
- Employers can use skill‑complexity profiles to design career ladders and internal training that foster frontier‑aligned diversification, improving workforce adaptability to AI integration.
Research directions:
- Incorporate these measures into structural and macro models of automation to better capture reallocation frictions and learning dynamics.
- Investigate causal impacts (e.g., randomized or quasi‑experimental upskilling interventions aimed at shifting frontier position).
- Study heterogeneity by sector, demographic groups, and international contexts; assess how rapid diffusion of AI shifts the skill network / frontier over time.
Limitations to bear in mind:
- LinkedIn data are self‑reported and non‑representative (platform selection); measures reflect listed skills, not necessarily realized task use or proficiency.
- Observational design: associations may reflect selection into portfolios and unobserved confounders.
- Results are conditional on pre‑2022 skill structures; the frontier may evolve as AI changes skill interdependencies.

Summary takeaway: for AI economics, this paper supplies a principled, data‑driven framework to quantify workers’ productive vs adaptive capital and shows that maximizing resilience requires combining domain depth with the greatest feasible cross‑domain breadth — operationalized as closeness to a diversity frontier.

Assessment

Paper Typecorrelational Evidence Strengthmedium — Very large, rich longitudinal sample and carefully constructed network-derived measures give strong predictive power and allow temporal separation of predictors and outcomes, but evidence is purely observational, relies on self-reported LinkedIn skills, and lacks exogenous identification to rule out omitted variable confounding or selection bias; therefore findings are informative but not definitive causal evidence. Methods Rigormedium — Methodological approach is sophisticated (statistically principled network construction, directionality inference, LRC, Louvain domains, and well-chosen diversity metrics) and uses temporal sequencing and controls, but it remains vulnerable to measurement error, platform selection effects, omitted variables (e.g., unobserved ability, employer training, local labor demand), and no instrumental variable, natural experiment, or fixed-effect strategy is reported in the supplied text to support causal claims. SampleRandom sample of approximately 2.4 million U.S.-based LinkedIn users who listed at least five skills prior to 2022, with 16,753 distinct skills observed; data structured as an individual-year panel (skill portfolios measured 2020–2022, outcomes 2023–2025); skill co-occurrence network constructed from these profiles. Themesskills_training labor_markets IdentificationObservational longitudinal analysis using a random sample (~2.4M) of US LinkedIn users; construct a directed skill network from observed skill co-occurrences (hypergeometric null model and asymmetric conditional probabilities) and identify skill domains via Louvain community detection; derive worker-level predictors (specialisation via normalized local reaching centrality, domain diversity via Shannon/Hill diversity, and 'diversity frontier' as relative diversity given specialisation); measure predictors in 2020–2022 and outcomes in 2023–2025 to enforce temporal ordering; estimate predictive regressions controlling for portfolio size and other covariates (robustness checks and network construction thresholds reported in appendix). No experimental or quasi-experimental source of exogenous variation is used; authors frame results as predictive associations rather than causal effects. GeneralizabilityLinkedIn users are not representative of the entire workforce (over-represents white-collar, digitally-visible professionals)., Self-reported skills may reflect signaling or profile-optimisation rather than actual ability (measurement error)., Sample restricted to U.S. workers and to users with >=5 listed skills limiting applicability to other countries or less formally credentialed workers., Pre-2022 measurement window may not capture rapid post-2022 changes in skills demand following generative AI diffusion., No causal identification; results may not generalize to causal policy interventions without further validation.

Claims (7)

Claim	Direction	Outcome	Confidence & Evidence	Details
Workers with more specialised skill portfolios are more likely to sort into higher-wage occupations. Wages	positive	Sorting into higher-wage occupations	Reading fidelity high Study strength medium	n=2400000 0.3
Workers with more diverse skill portfolios are associated with broader skill accumulation. Skill Acquisition	positive	Subsequent acquisition and accumulation of skills	Reading fidelity high Study strength medium	n=2400000 0.3
Workers with more diverse skill portfolios are associated with greater occupational mobility. Employment	positive	Occupational mobility and role transitions	Reading fidelity high Study strength medium	n=2400000 0.3
Workers closest to the diversity frontier are significantly more likely to acquire new skills than workers with comparable specialisation but narrower skill portfolios. Skill Acquisition	positive	Acquisition of new skills	Reading fidelity high Study strength medium	n=2400000 significantly more likely 0.3
Workers closest to the diversity frontier are significantly more likely to receive promotions than workers with comparable specialisation but narrower skill portfolios. Employment	positive	Receipt of promotions	Reading fidelity high Study strength medium	n=2400000 significantly more likely 0.3
Workers closest to the diversity frontier are significantly more likely to transition into occupations with lower exposure to automation than comparable workers with narrower skill portfolios. Automation Exposure	positive	Transitions into occupations with lower automation exposure	Reading fidelity high Study strength medium	n=2400000 significantly more likely 0.3
Adaptive outcomes are more closely associated with workers' diversity-frontier position than with specialisation or diversity considered in isolation. Skill Acquisition	positive	Adaptive career outcomes, including learning, mobility, promotions, and AI-relevant skill adoption	Reading fidelity high Study strength medium	n=2400000 0.3