Sakana AI's MASS method trains a model on the best runs of its own agent teams

Sakana AI, working with UC Berkeley, has published Multi-Agent Self-Supervision (MASS), a method for recursive self-improvement that does not depend on human evaluators. In MASS, one shared model solves tasks through a team of virtual subagents. The same model searches for better multi-agent workflows and uses its own judgments to pick the best ones. It is then trained on the executions of the best-performing team, which feeds the team's collective experience back into the model. The researchers ran two cycles with a 27-billion-parameter open-weight model on synthetic, open-ended research tasks. After the second cycle, its score per output token reached 1.2 to 1.6 times the base model's level across four research benchmarks. Sakana reports that training only on task-solving data also improved the model in its roles as workflow optimizer and evaluator. The work addresses a growing problem: as AI systems take on open-ended research, human experts may find it harder to judge their results. RuntimeWire noted that the gains carried over less well to other benchmarks. It also noted that with only two cycles, long-run reliability and real-world costs remain open questions.