Letting a Swarm of Agents Train Themselves: Sakana AI's MASS Delivers Up to 1.6x on Research Tasks—But Coding Didn't Improve
Released by Sakana AI with UC Berkeley, MASS has a single model play executor, orchestrator, and grader, using agent collaboration as training material to self-improve. It posted up to 1.6x on research benchmarks, yet on coding tasks like SWE-bench it fell below the base model.