NEWS

---

GMLTM 2.0.0 (2026-08-27)

• New GMLTM_corr(): a variant of GMLTM() that replaces the independent prior on theta (Sigma = sigma^2 * I) with a multivariate Normal prior with a general correlation matrix Sigma among cognitive components, estimated via a Cholesky factor with an LKJ prior (argument lkj_eta). Implements innovation (iii) of the GMLTM-D as formulated in Ramirez et al. (2024): admitting correlations between components. Returns EAP$Sigma and quantiles$Sigma in addition to the elements already returned by GMLTM().
• New extract_correlation(): extracts the posterior correlation matrix Sigma from a GMLTM_corr fit, reports pairwise credible intervals for each pair of components (flagging whether the interval excludes 0), and draws a correlation heatmap consistent with the style used in generate_Q_with_interactions().
• compute_model_validation() now labels models by class in $Summary and $Comparison (e.g. "GMLTM-D (Sigma libre)" for GMLTM_corr fits, "GMLTM-D (Sigma=I)" for GMLTM fits, "MLTM-D", "LLTM"), so GMLTM, GMLTM_corr, MLTM, and LLTM fits can be freely mixed in the same comparison list.
• New student_report(): builds a per-component and per-rule mastery report for a single examinee (theta EAP and credible interval, mastery cutlines gamma_m/tau_km, decision-confidence indices, and rule/item difficulty), following the mastery-diagnosis procedure for the MLTM-D in Embretson (2019). Includes a scannable badge-style summary plot (report$plots$badges) alongside the existing per-component continuum plots. This function was previously named informe_estudiante(); that name is kept as a deprecated alias and will be removed in a future release.
• New student_report_batch(): pages student_report() over many examinees. Always computes the lightweight numeric summary (theta EAP, component/rule mastery and confidence) for every requested student, but only builds the ggplot2 plots for the students in the current offset/limit page, so large batches stay cheap to page through.
• priors$alpha in GMLTM()/GMLTM_corr() now accepts a family element, either "normal" (default; half-Normal, unchanged) or "lognormal" (Log-Normal), letting users match the alpha prior family actually used when fitting. New prior_predictive_check() simulates data directly from a declared priors specification (including alpha$family) under the GMLTM-D's generative structure, before any data is fitted, following the prior predictive checking workflow of Gelman et al. (2020, Sect. 2.4); plot_prior_predictive_check() visualizes the resulting proportion-correct distribution against a configurable substantively plausible range.
• Fixed a bug in GMLTM(), GMLTM_corr(), MLTM(), and LLTM(): when data was a plain matrix (as opposed to a data.frame), the response vector was built via unlist(data), which does not flatten a matrix (its dim attribute is left untouched) and made model fitting fail with a dimension-mismatch error from Stan. Fixed by flattening with as.vector(as.matrix(data)), which produces the same column-major order for both data.frame and matrix inputs.
• All package documentation and source comments are now in English.
• Internal reorganization for maintainability, with no change in behavior: reliability.R and conditional_reliability.R were split into smaller, topic-focused files (reliability-enhanced.R, reliability-diagnostics.R, conditional-reliability-compare.R); related exported functions are now grouped via @family tags in their documentation; and a package-level overview page (?GMLTM) was added.
• Removed two orphaned, unused precompiled Stan model objects that had been accidentally left in inst/.

GMLTM 0.1.0 (2026-06-30)

• GMLTM1() and GMLTM2() have been removed. Their functionality is fully covered by the priors argument of GMLTM(). See the vignette for equivalent prior specifications.
• Initial release to CRAN.
• Implements LLTM (Fischer, 1973), MLTM-D (Embretson & Yang, 2013), and GMLTM-D (Ramirez et al., 2024) via Bayesian HMC with Stan.
• All models support user-defined prior distributions via the priors argument.
• Includes reliability estimation, posterior predictive checks, and ICC visualization tools.
• Stan backend uses rstan for full CRAN compatibility.