PaperlayerView exact revision on arXiv

Source-linked AI summary

The conservative turn in science: The changing character of knowledge recombination

Cheng Zhang, Xinkun Liang, Yi Bu, Hongkan Chen

arXiv:2609.08468v1cs.SIcs.DL

TL;DR

Existing evidence of declining disruptiveness in individual works does not establish whether cross-disciplinary knowledge-flow mechanisms have shifted. Using OpenAlex publication metadata and semantic-distance measures based on SPECTER2 embeddings, the study finds a shift from completely external bridges toward partially shared combinations, corroborated by independent semantic-distance measures.

  • Problem

    Evidence of declining disruptiveness in individual works does not reveal whether cross-disciplinary knowledge-flow mechanisms themselves have shifted.

    Source: §1 · p. 2

  • Method

    The study draws on OpenAlex publication metadata and uses a crossover-distance measure and SPECTER2 embeddings of publication titles and abstracts.

    Source: §1 · p. 3, §2.1 · p. 4, §4.4 · p. 18

  • Results

    Over six decades, interdisciplinary citations shifted from completely external bridges toward partially shared, cross-external combinations, corroborated by independent semantic-distance measures.

    Source: §5 · p. 18

  • Takeaways & Limitations

    The conclusion points toward frameworks that reward integration depth over mere connectivity.

    Source: §5 · p. 18

    The semantic-distance measure uses SPECTER2 embeddings of titles and abstracts; JCR matching retains only a fraction.

    Source: §4.4 · p. 18

Abstract

from arXiv · show

This study examines how the dominant mode of cross-disciplinary knowledge combination has changed over time. It addresses a limitation of existing studies that document declining disruptiveness at the level of individual works but do not reveal whether the mechanisms of cross-disciplinary knowledge flow have themselves shifted. Drawing on 56 million publications and 816 million citations from OpenAlex (1960-2025), the study classifies cross-disciplinary citations into five types based on the set relationship between the subfield portfolios of citing and cited publications and introduces a semantic-distance measure derived from SPECTER-v2 embeddings of publication abstracts. It subsequently demonstrates that cross-disciplinary citations have shifted markedly away from configurations with no shared disciplinary ground toward types built on partially overlapping knowledge bases, a reorientation further reflected in a steady decline in average semantic distance between citing and cited fields. This shift co-occurs with a diffusion of brokerage positions across the network and a greater reliance on older literature, consistent with a system that has become broader and more decentralized in its interdisciplinary reach. Finally, this study discusses the implications of these findings for science policy and for the design of cross-disciplinary funding mechanisms.

1. Introduction

The study asks whether cross-disciplinary knowledge flow has shifted from distant recombination toward integration on partially shared foundations, a question not answered by evidence of declining disruptiveness in individual works. It reports a system-wide reorientation toward shared knowledge, alongside changes in brokerage and reliance on older literature, and raises questions about funding and evaluation.

1. Introduction

  • The study addresses whether cross-disciplinary knowledge flow moved from direct, distant recombination with no shared ground toward integration on partially overlapping knowledge bases.Earlier evidence of declining disruptiveness in individual works does not show whether the mechanisms of cross-disciplinary flow themselves changed.Source: §1 · p. 2
  • The authors investigate shifts across time and domains, changes in recombination across the proximity spectrum, and the evolution of interdisciplinary brokerage.They draw on 56 million publications and 816 million OpenAlex citations.Source: §1 · pp. 2–3
  • Cross-disciplinary citations shifted toward types built on partially shared knowledge, alongside a steady decline in average semantic distance between citing and cited fields.The study frames this as movement from relatively separate toward increasingly overlapping knowledge foundations.Source: §1 · p. 3
  • The shift co-occurred with diffused brokerage positions and greater reliance on older literature, consistent with broader, more decentralized interdisciplinary reach.The passage presents these patterns as co-occurring, not as causes of the shift.Source: §1 · p. 3
  • The crossover-distance measure marks where shared-foundation and fully external combinations are equally prevalent, connecting aggregate trends to optimal cognitive distance.The framework describes a trade-off: combinations that are too close add little novelty, while overly distant ones exceed absorptive capacity.Source: §1 · pp. 2–3
  • The observed shift toward shared-basis combination raises whether funding and evaluation structures adequately incentivize deeper cross-disciplinary integration.The paper presents this as a policy question, not a demonstrated effect of existing funding structures.Source: §1 · p. 3

2. Data and Method

The study assembles an OpenAlex publication corpus and operationalizes cross-disciplinary knowledge flows through subfield-overlap citation types, semantic measures, and annual network analysis.

2.2 Subfield-level citation classification

  • Five citation types classify each publication-level citation by the set relationship between the citing and cited publications’ subfield memberships.A publication-citation record may contribute to multiple subfield-pair records.Source: §2.2 · p. 5, Table 1 · p. 5

2.3 Semantic distance and citation intensity among subfields

  • Semantic distance between subfields uses 768-dimensional SPECTER-v2 publication-abstract embeddings and decade-specific subfield centroids.Centroids are recomputed by decade so the semantic structure can evolve over time.Source: §2.3 · p. 6
  • A degree-normalized index measures citation intensity between subfields using citation frequency and their total citations made and received.The index is defined for a directed subfield pair.Source: §2.3 · p. 6

2.4 Network analysis

  • Annual weighted directed subfield networks use boundary-crossing citation shares as edge weights, drop edges with fewer than five such citations, retain isolated active nodes, and calculate exact weighted betweenness.Edges represent cross-external, partially internal, and superset-external citations.Source: §2.4 · pp. 6–7

3. Results

Cross-disciplinary citation patterns vary across domains and have shifted over time toward more cross-external citations and less completely external citation. Meanwhile, brokerage has become less concentrated, even as network connections have broadened without comparable growth in tie strength.

3.1 Overall trends and descriptive statistics

  • Cross-external citations account for 56.03% of citations, compared with 18.91% for completely external citations; three other types together account for 25.06%.The three other types are completely internal, superset external, and partially internal citations.Source: §3.1 · pp. 7–8, Table 2

3.2 Citation-mode differences across domains

  • Physical Sciences favor citation modes based on shared knowledge more than Social Sciences, where completely external citations are highest and cross-external citations lowest.Physical Sciences have the lowest completely external share and the highest cross-external, partially internal, and completely internal shares.Source: §3.2 · p. 8

3.3 Temporal evolution of citation modes

  • Completely external citations fell from 0.5–0.6 to 0.2–0.3 as cross-external citations rose from 0.2–0.4 to 0.5–0.6 over time.Completely internal citations also increased slightly, from nearly 0 to 0.1; superset-external and partially internal citations showed no comparable long-run trend.Source: §3.3 · p. 10, Figure 3 · p. 11

3.4 Semantic distance and citation intensity

  • As semantic distance increases, completely external citations’ relative contribution rises while cross-external citations’ falls; their crossover shifts toward larger distances over time.The reported shift means completely external citations require greater semantic distance to match cross-external citations’ contribution.Source: §3.4 · p. 11, Figure 4 · p. 11

3.5 Network structure

  • Betweenness-centrality Gini coefficients fell from 0.60–0.75 to 0.40–0.50 across all domains and the pooled network, indicating less concentrated brokerage.Social Sciences had the highest within-domain Gini coefficient, indicating a comparatively more uneven betweenness distribution.Source: §3.5 · p. 12, Figure 5 · p. 13
  • Connected subfield pairs increased substantially, but connection strength changed little, especially for mixed-domain pairs.The findings characterize this as broader connection without enough corresponding depth.Source: §3.5 · p. 13

4. Discussion

The discussion frames the shift toward shorter, foundation-sharing recombinations as deradicalization and considers implications for evaluating and supporting interdisciplinary research. It also notes limitations of citation-based evidence and identifies extensions to patents, topic-level patterns, and recombination outcomes.

4.1 From radical recombination to conservative recombination

  • Over six decades, completely external citations bridging fields without shared foundations have given way to cross-external citations recombining partially overlapping knowledge bases.Source: §4.1 · p. 14

4.2 A moving boundary in semantic space

  • The crossover between completely external and cross-external citations moved outward in semantic space, while external citations declined and cross-external citations gained ground across distance bins.Source: §4.2 · pp. 15–16

4.3 The diffusion of brokerage

  • Betweenness brokerage became less concentrated as connected subfield pairs expanded, while connection strength showed little corresponding increase and sometimes declined.Source: §4.3 · pp. 16–17

4.4 The deradicalization trend of scientific recombination

  • The interdisciplinary frontier has not vanished but shifted from radical, foundation-breaking combinations toward moderate, foundation-sharing forms that consolidate knowledge.Source: §4.4 · p. 17
  • Because cross-disciplinary activity volume need not track integration depth, evaluation and incentives should consider the cognitive distance at which interdisciplinary links operate.Source: §4.4 · p. 17
  • The conclusions are limited because citations capture only one knowledge-transfer channel and may reflect referencing conventions, while semantic distance uses publication titles and abstracts alone.Source: §4.4 · p. 18
  • JCR matching covers only a fraction of OpenAlex, coverage varies across domains and time, 2025 data are incomplete, and findings may differ at finer or coarser granularity.The analysis operates at the subfield level.Source: §4.4 · p. 18
  • Future work could compare patents with science and test whether topic-level deradicalization is uniform or concentrated in particular areas, including in relation to AI-assisted research.Source: §4.4 · p. 18
  • Linking system-level citation structure to the scientific and technological impact of specific recombination events could clarify deradicalization’s consequences for science’s long-run health.Source: §4.4 · p. 18

5. Conclusion

The balance of interdisciplinary citations has shifted from completely external bridges toward partially shared, cross-external combinations. Semantic distance, reference ages, and network brokerage corroborate a system that increasingly consolidates rather than displaces knowledge.

5. Conclusion

  • Over six decades, interdisciplinary citations shifted from completely external bridges toward partially shared, cross-external combinations.The paper describes this as a systematic deradicalization of cross-disciplinary knowledge combination.Source: §5 · p. 18
  • Independent semantic-distance measures, older reference ages, and more diffuse network brokerage corroborate the structural reorientation.Source: §5 · p. 18
  • Together with declining disruptiveness (Park et al., 2023), the findings point to science that increasingly consolidates rather than displaces knowledge.The system recombines within shorter, more absorbable cognitive distances.Source: §5 · p. 18
  • Rising cross-disciplinary activity should not be equated with deeper integration.The conclusion calls for frameworks that reward integration depth over mere connectivity.Source: §5 · p. 18
  • The structural pull toward conservative recombination is likely to persist.Source: §5 · p. 18

Supplementary

The supplementary figures track citation types, semantic distance, network structure, brokerage, and team size across years or subfields. Their reported distributions show slight outward expansion in semantic distance and no systematic shift in citation intensity.

Supplementary

  • Figure 7 tracks annual publication volumes by domain and annual counts of the five citation types.Source: Figure 7 · p. 21, Supplementary · pp. 21–22
  • Across 1960–2019, in decade-long intervals, semantic-distance distributions between subfields exhibit a slight outward expansion over time, whereas citation-intensity distributions show no systematic shift.Figure 8 presents both distributions; for visibility, the citation-intensity plot is truncated, with an inset displaying the full untruncated distribution.Source: Supplementary · p. 22, Figure 8 · p. 22
  • Figure 9 tracks annual crossover points between cross-external and completely external contributions by semantic-distance percentile and relative contribution.The figure extracts the intersection points annually; the horizontal axis is year.Source: Figure 9 · p. 22, Supplementary · p. 23
  • Figure 10 compares the ten subfields with highest betweenness centrality in 1965 and 2015, marking entrants, exits, and persistent members.Bar colors denote domain; Artificial Intelligence is individually labeled.Source: Figure 10 · p. 23
  • Figure 11 compares subfield networks in 1965 and 2015, with node size indicating publication volume and color indicating domain.Node positions use a joint PCA projection of subfield centroid embeddings; displayed edges are limited to the top 60 nodes by betweenness centrality.Source: Figure 11 · p. 24
  • Figure 12 compares mean authors per article and 95% confidence intervals for seven selected subfields against their parent domains, plus a panel comparing four domains.The series cover 1960–2025.Source: Figure 12 · p. 24
Loading 2609.08468v1…