Not yet reviewed by Pith; the record is open.

Classifying variable stars is key for understanding stellar evolution and galactic dynamics. With the demands of large astronomical surveys, machine learning models, especially attention-based neural networks, have become the state-of-the-art. While achieving high accuracy is crucial, enhancing model interpretability and uncertainty estimation is equally important to ensure that insights are both reliable and comprehensible. We aim to enhance transformer-based models for classifying astronomical light curves by incorporating uncertainty estimation techniques to detect misclassified instances. We tested our methods on labeled datasets from MACHO, OGLE-III, and ATLAS, introducing a framework that significantly improves the reliability of automated classification for the next-generation surveys. We used Astromer, a transformer-based encoder designed for capturing representations of single-band light curves. We enhanced its capabilities by applying three methods for quantifying uncertainty: Monte Carlo Dropout (MC Dropout), Hierarchical Stochastic Attention (HSA), and a novel hybrid method combining both approaches, which we have named Hierarchical Attention with Monte Carlo Dropout (HA-MC Dropout). We compared these methods against a baseline of deep ensembles (DEs). To estimate uncertainty estimation scores for the misclassification task, we selected Sampled Maximum Probability (SMP), Probability Variance (PV), and Bayesian Active Learning by Disagreement (BALD) as uncertainty estimates. In predictive performance tests, HA-MC Dropout outperforms the baseline, achieving macro F1-scores of 79.8+-0.5 on OGLE, 84+-1.3 on ATLAS, and 76.6+-1.8 on MACHO. When comparing the PV score values, the quality of uncertainty estimation by HA-MC Dropout surpasses that of all other methods, with improvements of 2.5+-2.3 for MACHO, 3.3+-2.1 for ATLAS and 8.5+-1.6 for OGLE-III.