Berkeley and Sakana AI propose MASS multi-agent self-supervision
Researchers at UC Berkeley and Sakana AI propose Multi-Agent Self-Supervision (MASS), in which the same model alternates between task-solving agent, workflow optimizer and evaluator: it first optimizes how multiple copies of itself collaborate with weights frozen, then distills the resulting trajectories back into shared weights. Running two MASS rounds on Qwen3.6-27B inside the qwen-code environment raised its win rate against the base model on held-out test tasks from 53.9% to 69.9%, and lifted per-output-token scores on MLR-Bench, DSBench, AstaBench E2E-Hard and ScienceAgentBench to roughly 1.2–1.6x the base model, while Terminal-Bench 2.0 and SWE-bench Verified came in at about 0.96x and 0.94x respectively. The paper also reports the evaluator's agreement with external judges rising from 0.73 to 0.93 and better training efficiency from multi-agent trajectories than single-agent ones; GPT-5.5 and Claude Opus 4.8 served as external judges in the experiments.
It reframes “AI improving AI” around multi-agent collaboration and self-supervision quality, offering a testable research path toward recursive self-improvement.