AlphaFold, an AI system for highly accurate protein structure prediction, has the potential to lower barriers in structural biology research. We hypothesize that its release has created greater opportunities for scientists at non-elite institutions. To test this, we employ a difference-in-differences strategy, comparing the publication output of 3434 structural biologists with that of 3014 other (non-structural) biologists. Scientists are categorized as being at elite or non-elite institutions based on whether their affiliations are among the top 50 institutions ranked by their h-index. Our findings suggest that AlphaFold-like models are associated with a larger increase in the publication output of structural biologists at non-elite institutions. However, this increase is largely concentrated in lower-tier journals. These results demonstrate how AI is reshaping competitive dynamics and altering the landscape of scientific progress.
As publication-based metrics form the foundation of most university ranking systems, the growing prevalence of hyperauthored papers has emerged as an influential factor shaping ranking outcomes. Hyperauthored papers are defined here as those within the uppermost 0.1% of the cumulative author count distribution in each discipline. This study investigates universities in clinical medicine and physics included in the 2022 NTU Ranking to evaluate how hyperauthorship affects the measurement of research productivity. Four counting methods, namely whole counting, first author counting, fractional counting, and harmonic counting, are compared to assess changes in university rankings after excluding hyperauthored papers. The results demonstrate that hyperauthorship can substantially alter university rankings, with effects varying across disciplines, ranking tiers, and countries. In clinical medicine, the first-author counting method shows more pronounced effects on lower-ranked universities, while in physics whole counting also has stronger effects on lower-ranked universities. By comparison, rankings across all tiers remain relatively stable under fractional and harmonic counting methods, indicating that these methods are less influenced by hyperauthored papers. At the national level, universities in the United States and China generally rose in ranking following the exclusion of hyperauthored papers, while those in Germany and Italy declined, and institutions in the United Kingdom were largely unaffected. Whole counting exhibited the highest sensitivity to hyperauthorship, while fractional counting and harmonic counting proved the most robust. These findings highlight the importance of accounting for hyperauthorship in research evaluation frameworks and adopting counting practices that mitigate distortions in institutional performance assessment.
Communication between authors and reviewers in peer review is widely viewed as contributing to manuscript improvement. In computer science, open peer review (OPR) introduces direct exchanges during the rebuttal stage, and upward score revisions are often interpreted as evidence of substantive improvement. Whether such revisions carry the same informational value as scores that were high from the outset, however, remains unclear. Using large-scale review records from a leading AI conference, ICLR, we examine how score improvements relate to subsequent citations and author-reviewer communication. We find that reviewers' initial scores are positively associated with subsequent citations, while the corresponding estimate for final scores is smaller and less precisely estimated. Papers with score increases receive approximately 16.39% fewer expected citations than papers without score increases at the same final-score level. Further analyses show that score increases are associated with more intensive exchanges; DeBERTa-v3 classification of the messages also links score increases to academic, evaluative, and procedural communication. These findings document conditional associations between score trajectories, communication patterns, and subsequent scholarly uptake in forum-based OPR.
This study provides a conceptual and analytical clarification of the Activity Index (AI) and the Fmeasure by identifying the structural relationships with their underlying share components and explaining their context-dependent properties. We conduct both an algebraic and empirical analysis of the relation among these indicators in two analytical contexts: Countries Specialization Comparison (comparing the specialization of multiple countries within a fixed domain) and Domains Comparison (comparing specialization across multiple domains within a fixed country). To examine the indicators' dependencies, we compute Pearson and Spearman correlation coefficients between the indicators and their underlying share components, DShareC (a domain's share of a country's output) and CShareD (a country's share of world output in that domain), using bibliometric data from InCites and OpenAlex covering >200 countries across 22 and 26 research fields over the period 1980-2024. We show that the AI and the F-measure exhibit a mirrored asymmetrical association with their underlying share components across the two analytical contexts. For Countries Comparison, the AI is rank-equivalent to DShareC for comparing countries within a fixed domain, whereas the F-measure operates as a proxy for CShareD. In the Domains Comparison, this pattern reverses: the AI becomes rank-equivalent to CShareD for comparing domains within a fixed country, whereas the F-measure is modulated by the volume of national output, correlating with DShareC in high-output countries and with CShareD in low-output countries. These findings clarify that the interpretation of specialization indicators depends on the analytical context in which they are applied. These results provide a conceptual framework for interpreting specialization indicators in scientometric evaluations and clear criteria for indicator selection, distinguishing between measures of internal research prioritization and global participation in scientific output.
It has been argued that concentrating research resources is necessary to maintain Sweden’s competitiveness as a research nation. Critics from representatives of industry and academia suggest that weak-impact research tends to be concentrated in smaller universities and colleges, and that the Swedish higher education system is too fragmented, with too many institutions attempting to cover a broad range of research areas rather than developing distinctive strengths. In this context, the aim of this article is to investigate how Swedish researchers’ achievements are distributed across different types of institutions, using bibliometric methods. The study is based on publications indexed in Clarivate’s Web of Science from 2017 to 2020, with citation data up to June 2022. Document types analyzed include Articles, Letters, Proceedings Papers, and Reviews from the Web of Science Core Collection, specifically SCI-Expanded, SSCI, and A&HCI. A percentile-based model is employed to aggregate individual researchers’ publication output and impact, accounting for citation frequency. Findings show that quality differences between specialized colleges and the newer universities and colleges are relatively small, although there are discrepancies for a few institutions. Previous funding streams by the Swedish Alliance government disproportionately benefited established and discipline-focused universities, which may have influenced these patterns. The study concludes that research performance is not inherently superior at larger universities. New universities and colleges can achieve comparable results. These findings challenge prevailing assumptions and support a more equitable allocation of research funding, contributing to a broader and more sustainable research ecosystem in Sweden.
