On MultiChallenge, initializing Mode 3 with a mismatched hand-designed expert graph yields an initial validation accuracy of only 54.0%. The self-evolution loop recovers through two phases: pruning mismatched transition edges (reducing edge count from 13 to 10 by Generation 5) and restructuring control flow, raising validation accuracy to 93.9% (approaching scratch-built Mode 5 at 94.9%).

来源：paper:p35-b367