吴令飞Nature发文：大团队渐进发展，小团队颠覆式创新
2019年02月14日 10:45 分类：未分类 阅读：877 评论：0
导语
2019年2月13日，Nature 杂志官网在线发表了一篇以Large Teams Have Developed Science and Technology; Small Teams Have Disrupted It为题文章（article），介绍了对于团队创新的最新研究成果，发现小团队比大团队更能做出颠覆式的创新成果。
本文第一作者吴令飞是集智科学家，集智-凯风研读营学者，腾讯研究院×集智俱乐部AI&Society第三期讲者。第二作者王大顺是美国西北大学凯洛格商学院、复杂系统研究中心副教授。通讯作者 James A. Evans 是美国芝加哥大学社会学系教授，知识实验室主任，腾讯研究院×集智俱乐部AI&Society第十二期讲者。 http://swarma.blog.caixin.com/archives/198404
Advocates of team science have claimed that a shift to larger teams in science and technology fulfils the essential function of solving problems in modern society that are complex and which require interdisciplinary solutions6–8. Although much has been demonstrated about the professional and career benefits of team size for team members9, there is little evidence that supports the notion that larger teams are optimized for knowledge discovery and technological invention9. Experimental and observational research on groups reveals that individuals in large groups think and act differently—they generate fewer ideas10,11, recall less learned information12, reject external perspectives more often13 and tend to neutralize each other’s viewpoints14. Small and large teams may also differ in their response to the risks associated with innovation. Large teams, such as large business organizations, may focus on sure bets with large potential markets, whereas small teams that have more to gain and less to lose may undertake new, untested opportunities with the potential for high growth and failure15, leading to markedly different outcomes. These possibilities led us to explore the consequences of smaller and larger teams for scientific and technological advance, and how such teams search and assemble knowledge differently.
Previous research demonstrates that large article and patent teams receive slightly more citations2,16. However, citation counts alone cannot capture distinct types of contribution. This can be seen in the difference between two well-known articles: one about self-organized criticality17 (the BTW model, after the authors’ initials) and another about Bose–Einstein condensation18 (for which Wolfgang Ketterle was awarded the 2001 Nobel Prize in Physics) (Fig.1, Extended Data Fig.1b). The two articles have received a similar number of citations, but most research subsequent to the BTW-model article has cited only the model itself without mentioning references from the article. By con-trast, the Bose–Einstein condensation article is almost always co-cited with Bose19, Einstein20 and other antecedents. The difference between the two papers is reflected not in citation counts but in whether they suggested or solved scientific problems—whether they disrupted or developed existing scientific ideas, respectively21. The BTW model launched new streams of research, whereas the experimental realiza-tion of Bose–Einstein condensation elaborated upon possibilities that had previously been posed.
To systematically evaluate the role that small and large teams have in unfolding scientific and technological advances, we collected large-scale datasets from three related but distinct domains (seeMethods): (1) the Web of Science (WOS) database that contains more than 42 million articles published between 1954 and 2014, and 611 million cita-tions among them; (2) 5 million patents granted by the US Patent and Trademark Office from 1976 to 2014, and 65 million citations added by patent applicants; (3) 16 million software projects and 9 million forks to them on GitHub (2011–2014), a popular web platform that allows users to collaborate on the same code repository and ‘cite’ other repositories by copying and building on their code.
For each dataset, we assess the degree to which each work disrupts the field of science or technology to which it belongs by introducing something new that eclipses attention to previous work upon which it has built. We use a measure that was previously designed22 to identify destabilization and consolidation in patented inventions; this measure varies between −1 and 1, which corresponds to science and technology that develops or disrupts, respectively (Fig.1a). We validate the dis-ruption measure in several ways. First, we investigate the distribution of disruption across scientific papers (Fig.1b); the disruptive BTW-model article is located in the top 1%, whereas the developmental Bose–Einstein condensation paper is in the bottom 3% of the disrup-tion distribution. We also find that, on average, Nobel-prize-winning papers register among the 2% most disruptive articles. Review articles are developmental with a negative mean of disruption (bottom 46%), whereas the original research works that they review have a positive mean (top 23%). Articles that headline prominent prior work—such as the Bose–Einstein condensation article—lie in the bottom 25% (Supplementary Table1). We further confirmed these results with a survey in which we asked scholars from diverse fields to propose dis-ruptive and developmental articles; this symmetrically confirmed the disruption measure (Supplementary Table2). Finally, we find that in the titles of articles different words associate with disruptive (‘introduce’, ‘measure’, ‘change’ and ‘advance’) versus developing (‘endorse’, ‘confirm’, ‘demonstrate’, ‘theory’ and ‘model’) papers (Fig.1c, Supplementary Table3).
We predict that work by small teams will be substantially more dis-ruptive than work by large teams. Our databases of papers, patents and software strongly confirm this prediction. Our sources differ in scope and domain, but we consistently observe that over the past 60 years, larger teams produce articles, patents and software with a disruption score that markedly and monotonically declines with each additional team member (Fig.2a–c, Extended Data Fig.3). Specifically, as teams grow from 1 to 50 team members, their papers, patents and products drop in percentiles of measured disruption by 70, 30 and 50, respec-tively (Extended Data Fig.3a). In every case, this highlights a transition from disruption to development. These results support the hypothesis that large teams may be better designed or incentivized to develop cur-rent science and technology, and that small teams disrupt science and technology with new problems and opportunities.
This phenomenon is amplified when we focus on the most disruptive and impactful work (Fig.2d–f). We measure the impact of each article, patent and software using the number of citations each work received. As shown in Fig.2d, solo authors are just as likely to produce high-im-pact papers (in the top 5% of citations) as teams with five members, but solo-authored papers are 72% more likely to be highly disruptive (in the top 5% of disruptive papers). By contrast, ten-person teams are 50% more likely to score a high-impact paper, yet these contributions are much more likely to develop existing ideas already prominent in the system, which is reflected in the very low likelihood they are among the most disruptive. By repeating the same analyses for patents (Fig.2e) and software development (Fig.2f), we find that disruption and impact consistently diverge as teams grow in size.
Differences in disruption between works produced by small and large teams are magnified as the impact of the work increases (Fig.3a); high-impact papers produced by small teams are the most disruptive, and high-impact papers produced by large teams are the most devel-opmental. As article impact increases, the negative slope of disruption as a function of team size steepens sharply. Even within the pool of high-impact articles and patents (Fig.3a, top 5% of citations), which are statistically more likely produced by large teams (Fig.2d), small teams have disrupted the current system with substantially more new ideas. We further split papers by time period (Extended Data Fig.3c) and scientific field (Fig.3b, Extended Data Fig.4), and found that these patterns linking disruption and team size are stable for all eras and for 90% of disciplines. The only consistent exceptions were observed for engineering and computer science, in which conference proceedings rather than journal articles are the publishing norm (the WOS database indexes only journal articles).
We considered whether observed differences between the work of small and large teams could simply be attributed to differences in disruptive potential for the different types of articles that they produce; for example, small teams may generate more theoreti-cal innovations and large teams more empirical analyses. Drawing on a previous approach23, we matched papers from www.arXiv.org with the WOS database and repeated our analyses controlling for the number of figures in each article (Extended Data Fig.5a), as empirical papers tend to have more figures than theoretical ones. Our results suggest that most of the difference in disruption between work from smaller and larger teams is not driven by differences in whether they contributed theoretical versus empirical papers (that is, had more or less figures). The association remains the same when we consider other distinctions, including review versus original research articles. Review articles with fewer authors are more disruptive than those with more, just as with original research articles (Extended Data Fig.5b).
Another possible explanation for our results is that the team effect thatwe observe occurs because the scientists, inventors and software designers involved in larger teams are qualitatively different from those comprising smaller teams. But when we predict disruptiveness as a function of team size, controlling for publication year, topic and author (Fig.3c, Extended Data Fig.3b, Supplementary Table4), we find that the decrease of disruption with the growth of team size continues to hold, and controlling for authors greatly improves the percentage of variance explained (Supplementary Table4).We further test the robustness of our results against several differ-ent definitions of the disruption measure, including the removal of self-citation links, exclusion of all but high-impact references and other variations (Extended Data Fig.5g–i). Across all variations, our conclu-sions remain the same.The considerable difference in disruption between large and small teams raises questions regarding how these teams differ in searching the past to formulate their next paper, patent or product. When we dissect search behaviour, we find that large and small teams engage in notably different practices that may be related to divergent contributions in disruption and impact. Specifically, we measure search depth as average relative age of references cited and search popularity as median citations to the references of a focal work. We examine these search strategiesLETTER RESEARCHacross fields, time periods and impact levels in science, technology and software. We also relate these search strategies to temporal delay in the impact these works receive using the ‘Sleeping Beauty index’24, which captures a delayed burst of attention traced by convexity in the citation attention that a work receives over time.
We find that solo authors and small teams much more often build on older, less popular ideas (Fig.4, Extended Data Fig.6). Larger teams more often target recent, high-impact work as their primary source of inspiration, and this tendency increases monotonically with team size. It follows that large teams receive more of their citations rapidly, as their work is immediately relevant to more contemporaries whose ideas they develop and audiences primed to appreciate them. Conversely, smaller teams experience a much longer citation delay; the average Sleeping Beauty indexpercentile for solo and two-person research teams is twice that of ten-person teams (Extended Data Fig.7). As a result, even though small teams receive less recognition overall owing to the rapid decay of collective attention25–27 (as shown in Fig.2a), their successful research produces a ripple effect, which becomes an influential source of later large-team success (Extended Data Fig.8).
We also consider the relationship between these distinctive search mechanisms and recent findings28 that suggest multi- and inter- disciplinary teams more often link work from divergent fields. We examined the novelty of journal combinations within article reference lists and also keyword combinations within articles in relation to team size. These show consistent diminishing marginal increases to novelty with team size, such that with each new team member, their contribution to novel combinations decreases (Extended Data Fig.9). Moreover, using a previous measure of atypical combinations28, we find that atypical combinations increase slowly up to teams of approximately ten and then decrease sharply below the value associated with a solo investigator. Whereas larger teams facilitate broader search, small teams search deeper.
In summary, we report a universal and previously undocumented pattern that systematically differentiates the contributions of small and large teams in the creation of scientific papers, technology patents and software products. Small teams disrupt science and technology by exploring and amplifying promising ideas from older and less-popular work. Large teams develop recent successes, by solving acknowledged problems and refining common designs. Some of this difference results from the substance of science and technology that small versus large teams tackle, but the larger part appears to emerge as a consequence of team size itself. Certain types of research require the resources of large teams, but large teams demand an ongoing stream of funding and success to ‘pay the bills’29, which makes them more sensitive to the loss of reputation and support that comes from failure30. Our findings are consistent with field research on teams in other domains, which demonstrate that small groups with more to gain and less to lose are more likely to undertake new and untested opportunities that have the potential for high growth and failure15. Our findings are also in accordance with experimental and observational research on groups that demonstrates how individuals in large groups think and act differently from those in small groups10–14.
Both small and large teams are essential to a flourishing ecology of science and technology. Taken together, the increasing dominance of large teams, a flurry of scholarship on their perceived benefits2,6–9,28,31 and our findings call for new investigations into the vital role that indi-viduals and small groups have in advancing science and technology. Direct sponsorship of small-group research may not be enough to pre-serve itsbenefits. We analysed articles published from 2004 to 2014 that acknowledged financial support from several top government agencies around the world, and found that small teams with this funding are indistinguishable from large teams in their tendency to develop rather than disrupt their fields (Extended Data Fig.10). In contrast to Nobel Prize papers, which have an average disruption among the top 2% of all contemporary papers, funded papers rank near the bottom 31%. Thiscould result from a conservative review process, proposals designed to anticipate such a process or a planning effect whereby small teams lock themselves into large-team inertia by remaining accountable to a funded proposal. When we compare two major policy incentives for science (funding versus awards), we find that Nobel-prize-winning articles significantly oversample small disruptive teams, whereas those that acknowledge US National Science Foundation funding oversample large developmental teams. Regardless of the dominant driver, these results paint a unified portrait of underfunded solo investigators and small teams who disrupt science and technology by generating new directions on the basis of deeper and wider information search. These results suggest the need for government, industry and non-profit funders of science and technology to investigate the critical role that small teams appear to have in expanding the frontiers of knowledge, even as large teams rapidly develop them.
Online content Any methods, additional references, Nature Research reporting summaries, source data, statements of data availability and associated accession codes are available at https://doi.org/10.1038/s41586-019-0941-9
Received: 20 April 2018; Accepted: 7 January 2019; Published online 13 February 2019.
