Gate self-evolved skills before they enter the pool: 72% pass@1 with a skill library 5x smaller
arXiv 2608.05810·medium signal
VaG attacks skill contamination in self-evolving agents with pre-commit gating — three complementary critics vet a candidate skill before it joins the library. It reaches 72% pass@1 while keeping the skill pool roughly 5× smaller than unconditional accumulation, and the three critics intercept largely disjoint classes of harmful skills, so dropping any one leaves a real gap. The practical read: if your agent writes its own skills, the admission check matters more than the generation step, and one critic is not enough.