Has Scientific Talent Shifted from Depth to Breadth?Evidence across Papers, Knowledge Inputs, Careers, and Teams

1 Introduction

• As the stock of knowledge grows, the burden-of-knowledge account predicts longer training, narrower individual expertise, and greater reliance on collaboration.Teams have become dominant producers of scientific and technological knowledge.Source: §1 · p. 1
• AI tools that support literature search, coding, data inspection, and translation across fields motivate the possibility of broader search, rapid learning, and cross-domain coordination.The introduction presents this as a consequential hypothesis about the scientific profile, not an established trend.Source: §1 · p. 1
• Breadth can describe paper topics, cited sources, individual histories, or team expertise, and measures combining variety, balance, and disparity can yield contrasting conclusions.This motivates evaluating expertise through several linked levels of observation.Source: §1 · p. 1
• Prior work links unusual combinations to high-impact science and greater failure risk, while specialization can earn persistent rewards.Returns to interdisciplinarity vary across contexts, and specialist–generalist advantages depend on career stage and the pace of frontier change.Source: §1 · pp. 1–2
• The study frames a broadening of individual scientific expertise as a hypothesis to test across linked levels of research, rather than inferring it from any single measure.The analysis connects paper topics, cited knowledge, contributor histories, teams, and mature citation outcomes.Source: §1 · pp. 1–2
• The analysis uses public bibliographic data over 2010–2025, six prespecified primary tests with Holm familywise-error control, and complementary robustness and counterfactual analyses.Figure 1 links articles to cited knowledge and contributor histories, then links capability profiles to mature citation outcomes.Source: §1 · p. 2, Figure 1 · p. 2

2.1 Observable structure of scientific expertise

• Three patterns would support broader individual expertise: papers spanning greater intellectual distances, more diverse recent histories among established contributors, and positive returns to breadth or its complementarity with depth.The article sample contains 48,000 records before exclusions; cited knowledge expansion resolved 51,736 of 55,223 sampled reference IDs (93.7%).Source: §2.1 · pp. 2–3
• Paper topic breadth, reference breadth, individual breadth, focal depth, and team size capture distinct parts of the research process and allow specialization to coexist with integration across knowledge sources.The framework distinguishes changes in individual repertoires from organizational responses through larger teams.Source: §2.1 · pp. 2–3

2.2 Primary and triangulating tests

• The six primary hypotheses test 2010–2022 trends in team size and paper breadth, 2019–2022 changes in individual breadth and depth, breadth–depth interaction in mature FWCI, and post-2022 deviations from earlier forecasts.The 2019–2022 comparison precedes widespread generative AI access; mature citation models end in 2022.Source: §2.2 · p. 3
• Reference breadth is an independent measure of knowledge inputs, with pilot, expanded, clustered, and holdout estimates used to assess sampling precision and stability.The article sample estimates balanced trends across six fields, while the contributor sample represents a randomly selected established contributor to a sampled article.Source: §2.2 · p. 3, Table 1 · p. 3

3.1 Expanding teams and diverging measures of breadth

• Team size rose 37.3% [34.4%, 40.3%] while paper topic breadth fell 0.0144 [–0.0199, –0.0089] on the 0–1 scale, 2010–2022.Topic breadth declined by 0.00120 per year; estimates remained negative across several sensitivity checks.Source: §3.1 · p. 3, Figure 2 · p. 4
• Cited knowledge breadth increased by 0.00228 per year [0.00002, 0.00454] in the expanded sample, while coarser input measures were approximately stable.Clustering within field and year cells gave a wider interval [–0.00053, 0.00509], supporting a stable-to-modestly-broader input pattern alongside concentrated output topics.Source: §3.1 · p. 4

3.2 Individual repertoires changed more modestly

• Established contributors’ 2019–2022 prior breadth was 0.0190 higher [–0.0078, 0.0459] than in 2010–2014, within the specified ±0.05 equivalence bound.Focal depth changed by –0.0274 [–0.1451, 0.0903], leaving its direction less precisely resolved.Source: §3.2 · p. 4
• In 2023–2025, five-year breadth rose by 0.0327 [0.0044, 0.0610]; weighted median breadth shifted from 0.706 to 0.744, while depth medians were 0.578 and 0.585.The recent breadth estimate attenuated under a fixed publication count.Source: §3.2 · p. 5

3.3 Focal depth was associated with higher mature citation impact

• A one-standard-deviation increase in focal depth was associated with 8.2% higher 1 + FWCI [1.9%, 14.9%], while average breadth and interaction associations were within specified practically small bounds.The model adjusted for field and year, career age, recent publication volume, team size, and open-access status.Source: §3.3 · pp. 6–7
• Adjusted geometric FWCI was 0.45 for low-breadth, low-depth profiles, 0.48 for broad-shallow, 0.62 for narrow-deep, and 0.67 for broad-deep profiles.The depth association remained positive across several alternative expertise and outcome specifications, but weakened for raw citations through 2019 and a top-decile outcome.Source: §3.3 · p. 7

3.4 Recent patterns in the context of longer trajectories

• The mean 2023–2025 paper-breadth residual from earlier forecasts was –0.01084 [–0.026, 0.004] (p = 0.116), with four fields below and two above projected paths.Placebo forecast origins produced mean residuals near zero.Source: §3.4 · p. 7
• Across 83 subfields, the recent paper-breadth coefficient by baseline AI intensity was 0.00000 [–0.00467, 0.00467] (p = 1.000), consistent with similar changes across that gradient.Differential-change estimates for log(1 + team size) and cited knowledge breadth retained wider plausible ranges.Source: §3.4 · p. 7

4 Discussion

• Established contributors’ repertoires changed gradually, while the discussion frames focused accumulation as compatible with knowledge and collaborators reaching beyond an individual specialty.The paper argues that different measures of breadth reveal distinct levels of the organization of science.Source: §4 · pp. 8–9
• Paper topics became more concentrated while cited knowledge remained stable to modestly broader, separating output focus from the reach of knowledge inputs.A focused article may draw on methods, evidence, or concepts from several domains.Source: §4 · pp. 7–8
• Team size grew by 37.3%, and topic concentration persisted after standardizing team composition, consistent with more contributors coordinating around focused problems.The discussion connects this organizational pattern to collaboration combining specialized contributions as the knowledge frontier becomes harder to master individually.Source: §4 · p. 9
• Focal depth had the clearest average association with mature citation impact, while breadth and its interaction with depth fell within specified bounds for practically small average associations.The broad-and-deep profile had the highest fitted value, but the narrow-and-deep profile captured much of the fitted advantage.Source: §4 · p. 9
• Focused output and expanding collaboration predated widespread generative AI access, and recent changes did not clearly vary with baseline AI intensity across 83 subfields.The AI-intensity comparison found a tightly centered paper-topic estimate, while estimates for team size and cited knowledge breadth retained wider plausible ranges.Source: Figure 7 · p. 8, §4 · p. 9
• For training and evaluation, the discussion supports deep expertise as an empirical anchor, complemented by the capacity to locate, evaluate, and connect knowledge beyond a focal specialty.Public contribution statements, method histories, and research-software records are proposed for testing these capabilities in future computational work.Source: §4 · p. 9

5.1 Data source and sampling by field and year

• 47,959 retained articles came from 500 sampled records per field and year across six fields, yielding balanced trend estimates for 2010–2025.Records were retrieved from OpenAlex; retracted and paratext records were removed.Source: §5.1 · p. 9

5.2 Topic hierarchy and paper breadth

• Paper breadth uses pairwise distances in a topic hierarchy, integrating variety, balance, and disparity; the primary trend sample fixes each paper at three topics.Robustness checks use equal topic weights, maximum pair distance, cross-field indicators, and topic-count controls.Source: §5.2 · pp. 9–10

5.3 Independent cited knowledge validation

• Reference validation expanded sampling to 35 papers per field-year cell, resolving 51,736 cited-work IDs; eligible papers needed five classified references covering at least 70% of sampled IDs.The study also calculated effective cited fields, the share outside the focal field, and a coarser field/subfield distance.Source: §5.3 · p. 10

5.4 Contributor sampling and prior histories

• A focal article and then a contributor were randomly selected, producing 1,754 unique author IDs; weighted estimates target established contributors within the specified history frame.The frame required 3–300 lifetime works and selected contributors from teams of 1–20; histories were screened for stability and sufficient prior work.Source: §5.4 · p. 10

5.5 Individual breadth and depth

• Individual breadth and focal depth use recency-weighted prior topic histories, allowing an author to be broad and deep simultaneously.Robustness definitions alter the history window or use only primary-subfield matches; displays compare repeated cross-sections.Source: §5.5 · p. 10

5.6 Trend and return models

• Trend models use year and field effects, while return models relate standardized breadth, depth, and their interaction to mature FWCI with contributor weights and HC3 errors.Robustness checks vary outcomes and histories; multiplicity was adjusted with Holm’s method, and equivalence bounds were treated as sensitivity thresholds.Source: §5.6 · pp. 10–11, §5.8 · p. 11

5.7 Post-2022 counterfactuals and AI intensity

• Post-2022 departures were tested against field-specific forecasts, while AI-intensity contrasts used baseline topic shares in 83 subfields.The counterfactual interpretation requires parallel trends and assumptions linking exposure contrasts to treatment responses.Source: §5.7 · p. 11

5.9 Scope of measurement and inference

• The estimates cover a balanced six-field sample, established contributors on teams of 1–20, and a three-year recent period; topic classifications use one retrieval snapshot.Book-based scholarship has more variable coverage, and mature citation returns have complete windows only through 2022.Source: §5.9 · pp. 11–12

• The study uses public OpenAlex records.Source: Data and code availability · p. 12
• A replication package with the data and code required to reproduce the reported results is organized for public deposit before journal submission.Source: Data and code availability · p. 12

• Generative AI assisted literature discovery, code implementation, diagnostic design, figure preparation, and language editing under human direction.Source: AI assistance disclosure · p. 12
• Human authors conceived the questions and framework, directed and reviewed the work, interpreted results, verified sources, and take full responsibility for the research and manuscript.Source: AI assistance disclosure · p. 12

A.1 Field variation in measurement divergence

• Paper topic breadth slopes are negative in engineering, medicine, physics and astronomy, and social sciences, approximately flat in computer science, and slightly positive in arts and humanities.Source: §A.1 · p. 12
• Cited-knowledge paths show more annual variation, and their input trajectory remains sensitive to disciplinary context.The passage attributes this variation to the smaller reference sample in each cell.Source: §A.1 · p. 12
• Table A1 reinforces the output result across alternative topic distances, weighting rules, and treatments of topic count.Source: §A.1 · p. 12

A.2 Reference validation and coverage

• 0.00228 [0.00002, 0.00454] is the expanded annual reference breadth slope under HC3 errors, while clustering gives [–0.00053, 0.00509].After excluding 187 articles overlapping the pilot, 2,469 eligible pre-2023 articles yielded a slope of 0.00255 [0.00024, 0.00486] under HC3 errors; the holdout reproduces the positive expanded estimate using separate focal articles.Source: §A.2 · p. 12
• The aggregate resolution rate is 93.7% of unique cited identifiers, distinct from the figure’s per-article share of sampled references with usable primary topics.Source: §A.2 · p. 12
• Eligibility requires at least five classified references and 70% coverage by classified references, preserving a minimum information base for pairwise distance estimates.Source: §A.2 · p. 12

A.3 Stability of the capability and citation association

• Depth remained positively associated with impact across alternative history lengths, publication counts, identity-quality thresholds, sampling weights, and depth definitions.The robustness comparison is reported in Table A2.Source: §A.3 · p. 12, Table A2 · p. 13
• 0.0575–0.1284 was the range of positive depth coefficients when each field was omitted in turn.The association was distributed across the sampled research settings.Source: §A.3 · p. 12
• Bootstrap 95% intervals spanned zero for breadth [–0.0500, 0.0820] and interaction [–0.0421, 0.0574], but not depth [0.0206, 0.1344].The interaction interval is on the log scale.Source: §A.3 · p. 12, Table A2 · p. 13
• The interaction stayed near zero across field-omission models; a nominal negative 2019–2022 dynamic estimate did not alter the overall pattern of limited average complementarity.The positive focal-depth association remained the broader pattern.Source: §A.3 · p. 12

A.4 Recent heterogeneity by prior AI intensity

• Prior breadth's estimate was 0.0235 per standard deviation of baseline AI intensity [–0.0005, 0.0476], while focal depth's was –0.0467 [–0.1719, 0.0785].The breadth estimate motivates closer examination, but its interval spans little differential change and more appreciable broadening.Source: §A.4 · p. 13
• These contrasts are unadjusted exploratory summaries; analyses retain 83 subfields for paper, team, and cited-knowledge outcomes and 77 for contributor outcomes.Source: §A.4 · p. 14

A.5 Team selection and aggregated trends

• 0.41% to 0.97% was the rise in papers with more than 20 recorded authors; physics and astronomy reached 3.80% and medicine 1.80%.The exactly three-topic analysis retained 29,329 pre-2023 papers from teams of 1–20 authors.Source: §A.5 · p. 14, Table A4 · p. 14
• –0.00122 [–0.00169, –0.00076] was the annual breadth slope among 29,329 pre-2023 papers from teams of 1–20 authors.The negative trend therefore also characterizes articles within the contributor sampling frame.Source: §A.5 · p. 14, Table A4 · p. 14
• –0.00111 [–0.00184, –0.00038] was the annual slope across 78 field-year cell means; the six-field average was also –0.00111 but had a wider interval [–0.00266, 0.00044].The field-year regression had p = 0.003; the six-field summary had p = 0.125, reflecting between-field variation.Source: §A.5 · pp. 14–15

A.6 AI vocabulary sensitivity

• The broad vocabulary added four terms to six legacy strings; the resulting exposure correlation was 0.9996 across 83 subfields, and recent heterogeneity estimates remained closely matched.Both vocabularies use topic assignments from the same current OpenAlex snapshot.Source: Table A5 · p. 14, §A.6 · p. 15