A model got better by retraining on its own agent teams' best work.
Tokyo lab Sakana AI and UC Berkeley published a method called MASS on 8 October. One model organises agent teams, judges their work and is then retrained on the best results. After two rounds its score per unit of output rose up to 1.6 times, though coding tests showed no gain.
Why it mattersLabs are racing toward models that improve themselves without new human data. This early test is small, and other AI models did the judging.
Read in the app — free, in 9 languages
Related stories
Teams of AI coding agents cost up to five times more for little gain.2026-10-12Microsoft shipped a cage for AI agents that they cannot unlock.
2026-10-11Claude can now split one job among up to 1,000 agents.
2026-10-11Agents rebuilding an old shooter tried to edit their own checker.
2026-10-11Nikon stripped a microscope video prize over generative AI.
2026-10-11