1 paper
Xiaobo Wang, Zimu Wang, Siyu Ye +13
Agent harness evolution enables recursive self-improvement (RSI) without updating model weights. We examine whether higher benchmark scores yield reliable, efficient, and transfera…