Knowledge Independence Breeds Disruption but Limits Recognition

TL;DR

Knowledge Independence (KI) is introduced as a paper-level metric measuring the independence of a paper's references via KI = (n_ind − n_dep)/(n_ind + n_dep), capturing how often references do not cite one another. Across 114 million publications, KI strongly predicts disruption and mediates the disruptive advantage of small, onsite, and fresh teams, while KI itself declines over time, explaining disruption’s paradoxes of rising knowledge yet declining disruption and delayed recognition. The authors validate causality with Coarsened Exact Matching, Propensity Score Matching, and two Monte Carlo simulations (random rewiring and network-genesis) and extend the analysis to SciSciNet, Web of Science, and OECD patents, demonstrating broad generality. A mechanistic picture emerges in which KI-driven “knowledge brokers” link independent ideas, yet higher KI correlates with lower and slower recognition, offering a unified explanation for disruption–impact tradeoffs. Overall, the study proposes a universal law: Knowledge independence breeds disruption but limits recognition, with implications for research strategy and science of science policy.

Abstract

Despite extensive research on scientific disruption, two questions remain: why disruption has declined amid growing knowledge, and why disruptive work receives fewer and delayed citations. One way to address these questions is to identify an intrinsic, paper-level property that reliably predicts disruption and explains both patterns. Here, we propose a novel measure, knowledge independence, capturing the extent to which a paper draws on references that do not cite one another. Analyzing 114 million publications, we find that knowledge independence strongly predicts disruption and mediates the disruptive advantage of small, onsite, and fresh teams. Its long-term decline, nonreproducible by null models, provides a mechanistic explanation for the parallel decline in disruption. Causal and simulation evidence further indicates that knowledge independence drives the persistent trade-off between disruption and impact. Taken together, these findings fill a critical gap in understanding scientific innovation, revealing a universal law: Knowledge independence breeds disruption but limits recognition.

Figures (47)

• Figure 1: $|$ Quantifying knowledge independence. (A) Three examples of how knowledge independence ($\text{KI}$) is calculated for a given focal paper (blue diamond) based on its references (gray circles). A reference is of $ind$-type (green) if it does not cite any other reference, or of $dep$-type (orange) if it does. $\text{KI}$ is then the difference in the fraction of $ind$-type and $dep$-type references. (B) The $\text{KI}$ distribution of 73,256,108 papers with at least two references published between 1950 and 2024 in the OpenAlex. (C) The pervasive downtrend of $\text{KI}$ over time across disciplines. Bootstrapped 95% confidence intervals are shown as shaded bands.
• Figure 1: $|$ Distribution of reference count. The distribution of reference count for 87,419,190 articles recorded in OpenAlex follows a stretched exponential pattern, with a large number of papers containing relatively short reference lists.
• Figure 2: $|$ Knowledge independence is associated with scientific disruption. (A) For the 55,576,192 papers published before 2023 in the OpenAlex that are cited at least once, the average disruption percentile (blue curve, left y-axis) and the disruption positive ratio (green curve, right y-axis) increase with the $\text{KI}$ percentile. (B) The association between $\text{KI}$ and disruption persists, regardless of the impact of the focal paper. Here, impact is measured as the number of citations received within the first five years after publication, denoted by $C_5$. (C) The association between $\text{KI}$ and disruption persists across decades. (D) The association between $\text{KI}$ and disruption persists, regardless of discipline. Bootstrapped 95% confidence intervals are shown as shaded bands in panels A-D. (E) The ATT (average treatment effect on the treated) matrix of $\text{KI}$ on disruption via coarsened exact matching (CEM). Each controlled group is set as a baseline, and ATTs are calculated for comparisons between the baseline and each of the treated groups. Blue cells represent negative ATTs, while red ones represent positive ATTs, with color intensity proportional to the absolute value. Within each controlled group, the ATT increases as the treated size grows, transitioning from negative to positive when the treated size approaches the controlled size. Each ATT is tested against the null hypothesis (ATT equals 0) using a two-sided t-test. (*$p < 0.05$, **$p < 0.01$, ***$p < 0.001$).
• Figure 2: $|$ The association between disruption and $\text{KI}$ is robust across different reference counts. (A-B) When controlling for the percentile of reference count, both the percentile and positive ratio of disruption continue to increase with $\text{KI}$ across all levels of reference count. (C-D) When controlling for the reference count within the bottom group ($<=10$), where each reference count induces a limited range of $\text{KI}$ values, both the percentile and positive ratio of disruption continue to increase with $\text{KI}$ across all groups of reference count.
• Figure 3: $|$ Team's preference for knowledge independence. (A-C) The $\text{KI}$ percentiles in each bin are rescaled by the average value of the respective time period to highlight the trends. Curves with larger bounds are displayed in the inset to improve visualization. Over time, the relative $\text{KI}$ percentile decreases with team size (panel A) and team distance (panel B), and increases with team freshness (panel C). Here we calculate the geographic team distance by the coordinate information of affiliations, and categorize teams into two types---onsite ($\leq100$ km) and remote ($>100$ km) teams (panel B). The relationship between $\text{KI}$ and average collaboration distance is displayed in inset of panel B. (D-F) Solid lines depict the relationships between disruption and team properties with fixed $\text{KI}$ values, alongside the uncontrolled cases (insets). Dashed lines represent linear fitted curves. Teams with higher $\text{KI}$ are consistently more disruptive. Moreover, the slopes of the fitted curves with fixed $\text{KI}$ are flatter compared to the uncontrolled curves (insets). Bootstrapped 95% confidence intervals are shown as shaded bands.