Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails Paper • 2609.09134 • Published Sep 8 • 10