On HotpotQA multi-hop reasoning, self-evolution starting from a minimal skeleton (Mode 5) begins with a validation F1 of 77.59% (below Mode 3), subsequently surpasses Mode 3, and reaches a peak validation F1 of 83.31% at Generation 10, demonstrating that evolution from a minimal skeleton can yield more effective guidance than evolving from a hand-designed prior. · CiteArk
On HotpotQA multi-hop reasoning, self-evolution starting from a minimal skeleton (Mode 5) begins with a validation F1 of 77.59% (below Mode 3), subsequently surpasses Mode 3, and reaches a peak validation F1 of 83.31% at Generation 10, demonstrating that evolution from a minimal skeleton can yield more effective guidance than evolving from a hand-designed prior.
