Hacker News new | ask | show | jobs
by somebodythere 389 days ago
My guess is that they did RLVR post-training for SWE tasks, and a smaller model can undergo more RL steps for the same amount of computation.