mirtverse: Accessible, AI-ready Item Response Theory in R with mirt

Abstract

Abstract

Item response theory (IRT) is the statistical foundation of most modern tests: It links the responses people give to test items with the latent trait the test is meant to measure, and it underpins item banking, test assembly, computerized adaptive testing, score linking and fairness analyses (de Ayala, 2022; Embretson & Reise, 2000; Lord, 1980). The R environment (R Core Team, 2026) offers many IRT packages, and among them mirt (Chalmers, 2012) stands out for the breadth of models and estimation methods it supports. Its power comes with a cost, however: Choosing, checking, interpreting and reporting a model in mirt requires fluency in both R and psychometrics. mirtverse (Bulut, 2026b) is an R package that sits on top of mirt and makes it usable by a wider audience. It turns fitted mirt models into tidy tables, publication-ready ggplot2 graphics (Wickham, 2016), plain-language interpretations, item health cards, one-file HTML reports, and shareable dashboards. A point-and-click Shiny workbench runs the same functions and writes the R code for every step. Beyond the usual IRT toolkit, mirtverse adds a generative AI layer that writes analysis scripts from a plain-language request, drafts and verifies test items, trains and audits automated scorers for open-ended responses, and predicts item difficulty from item text. Every AI feature sits behind a psychometric check, so a language model can speed up the work but cannot quietly lower its standards. The package is accompanied by an open textbook (Bulut, 2026a), ten interactive teaching labs, and an instructor toolkit. This paper reviews the theory that mirtverse builds on, the current landscape of IRT software in R, the reasons for building on mirt, and the design of mirtverse, including its AI capabilities.