Ablation study comparison. Classification performance (mean F1-score) on the MACHO labeled (top) and ATLAS (bottom) datasets across different training sizes (20, 100, and 500 samples per class). The study compares baseline architectures (MLP, GRU, Raw RNN) against the proposed Astromer 2 encoder (A2) equipped with different aggregation strategies: max pooling vs. skip connection. Error bars and values indicate the standard deviation over multiple runs.