Scientific papers typically cite prior work to acknowledge intellectual debts (substantive citations). However, many citations are rhetorical, serving persuasive purposes regardless of actual influence. While often viewed negatively, potentially corrupting incentives and benefiting prominent researchers, this study explores whether rhetorical citations might offer some positive effects for science, potentially disproportionately benefiting less prominent researchers and papers. The central research question is whether the inclusion of rhetorical citations, often considered undesirable, might lead to a healthier scientific community in terms of attention distribution and the dynamism of ideas. This is investigated through the development of a comprehensive agent-based model of citation behavior.

Existing literature largely portrays rhetorical citations negatively, viewing them as undermining the integrity of scientific evaluation. Studies highlight issues such as misrepresentation of cited works and the potential for biased attention towards high-status papers (persuasion by naming-dropping). However, the paper also notes that some rhetorical citations provide context or differentiate contributions. The current consensus implicitly compares the real-world (with rhetorical citations) to a counterfactual world without them. This research aims to rigorously model and compare these two worlds, acknowledging that a substantive-only citation world might lead to a concentration of attention on a few elite papers, hindering scientific dynamism. Previous research indicates that highly cited works attract disproportionately more substantive reading and citing, creating a feedback loop where high status leads to further attention and citations. This research suggests that rhetorical citations might weaken this feedback loop by introducing factors beyond quality into the citation process. Empirical evidence shows that the increasing length of reference lists is allowing for more rhetorical citations and, surprisingly, that this seems to be associated with a decline in uncited papers, suggesting the positive influence of these citations on a wider distribution of attention.

The authors developed a family of agent-based models to simulate the citation process, incorporating both substantive and rhetorical motivations. The model integrates aspects of cognitive realism, reflecting how researchers strategically select papers to read based on perceived quality (often proxied by status and citation count) and subsequently cite them substantively if their quality surpasses a threshold. Remaining citation slots are filled rhetorically, selecting papers based on rhetorical value (which incorporates aspects of quality and perceived impact). The models include both homogeneous (identical perceptions among agents) and heterogeneous (varying perceptions) agents, thereby representing a more realistic scenario. The authors compare a ‘full model’ (with both substantive and rhetorical citations) with two null models (only substantive citations, differing in how they handle citation budget constraints). Three key community health metrics were defined: citation-quality correlation (measuring the degree to which highly cited papers are indeed high-quality), citation churn (representing the turnover of cited papers over time, reflecting dynamism), and citation inequality (measuring the concentration of citations among a small number of papers). The authors analyze how these metrics vary across models and when varying factors such as literature size, reading budget (number of papers read), and citing budget (number of references allowed). The model uses a beta distribution for quality and initial rhetorical value, reflecting distributions found in peer review data from the ICML conference. Specific parameters are determined based on literature on reading and citation practices. The models are run for 1000 time steps, with each step representing a research paper published and cited. The Pearson correlation coefficient, Gini coefficient, and the number of new citations from one timestep to the next are used to measure the community health metrics.

The simulation results show that enabling rhetorical citing leads to a stronger correlation between paper quality and citations, increased citation churn, and reduced citation inequality. This is because rhetorical citing redistributes some citations away from top-quality papers towards high-to-moderate quality papers that possess high rhetorical value. In the full model (with rhetorical citations), the correlation between citations and quality was higher (+2.5%), citation churn was significantly greater (2.36 times higher than in the null models), and citation inequality was substantially lower (30-31% lower). Analyzing citations to high-quality and mid-quality papers reveals that rhetorical citing significantly increased the relative number of citations to mid-quality papers, particularly through rhetorical citations. The analysis of the moderating effects of community characteristics shows that increasing the citing budget (i.e., allowing more references per paper) substantially improves all three health metrics, whereas the reading budget has little effect. Increasing literature size, on the other hand, shows a mixed effect, reducing citation-quality correlation and increasing both citation inequality and churn. The results indicate that allowing for a larger citation budget is key to improving the health metrics, particularly when rhetorical citation is enabled.

The findings challenge the conventional negative view of rhetorical citations. The results suggest that, counterintuitively, these citations can contribute positively to the health of a scientific community by fostering dynamism and reducing citation inequality. The redistribution of attention facilitated by rhetorical citing leads to a more proportional distribution of citations across a wider range of quality papers. The study emphasizes that citations are not an isolated phenomenon, but an outcome of the complete process of research, reading, and citation choices. The study's simulation shows that larger citation budgets (i.e., longer reference lists) can positively impact scientific communities when coupled with rhetorical citation practices, an outcome often counter to typical assumptions. This suggests that focusing on improving citation practices alone without addressing the wider incentive structure and cognitive constraints of researchers may be ineffective. The results suggest the importance of considering the complex interplay between reading practices, citation motivations, and the resulting effects on the overall health of the scientific community.

This paper presents a novel agent-based model demonstrating the potentially positive effects of rhetorical citations on the health of a scientific community. The findings suggest that rhetorical citing can enhance citation-quality correlation, citation churn, and reduce citation inequality, particularly when coupled with larger citation budgets. Further research should explore additional dimensions of community health beyond those considered in this study and investigate the long-term dynamics of rhetorical citation behavior. The results highlight the need for a more nuanced understanding of the citation process and its implications for policy decisions related to scientific evaluation.

The study relies on several simplifying assumptions, including the use of simplified representations of paper quality and rhetorical value and the assumption that agents and papers remain fixed across different models. The model focuses on short-to-medium term dynamics, neglecting long-term factors like scientific obsolescence and evolving reading practices. The choice of parameters and their distributions could influence the results, although robustness checks were conducted. The study does not exhaustively consider all potential effects of rhetorical citations, such as misinformation. Therefore, the conclusions regarding the overall benefit of rhetorical citations should be viewed as tentative.