1 articles with this tag
Zhang et al. train a 4B proposer to edit frozen-agent harnesses from execution feedback, beating a 35B teacher on held-out Reasoning Gym families.
