In the appendix, the paper compares DeepGate3 against DeepGate2 and MGVGA on five small circuit designs (bc0, apex1, k2, i10, mainpla) with just thousands of gates, due to DeepGate3's quadratic attention complexity limiting it to small circuits. DeepGate3 achieves an average F1-score of 0.390 and AUC of 0.834, outperforming DeepGate2 (F1=0.338, AUC=0.752) on the same small designs.
Evidence
correlational
Key metric
average F1-score = 0.390, AUC = 0.834 (5 small designs)
Caveat
Limited to small designs (thousands of gates) due to DeepGate3's quadratic time complexity in attention computation; not comparable to the 10-design evaluation in the main text