A single Qwen3.6-27B designs, executes and evaluates multi-agent flows, then trains on its best runs: achieving up to 1.6 times the score per token across four research benchmarks.
How do you improve a model on open-ended tasks, such as scientific research, when no one is able to reliably score their answers anymore? Sakana AI and the University of California, Berkeley propose MASS (Multi-Agent Self-Supervision), a recursive self-improvement method in which the model becomes the object to be optimized, the optimizer, and the evaluator at the same time. Among the authors is Matei Zaharia, co-founder of Databricks.
Two phases that alternate
The model first proposes multi-agent organizations (roles, information flow), executes them and compares the results, in an evolutionary search framed by structural safeguards. It is then refined on the trajectories it itself has judged to be the best, making it a better optimizer and a better judge for the next cycle. After two cycles on Qwen3.6-27B, the researchers measure a score per token produced multiplied by 1.2 to 1.6 on MLR-Bench, DSBench, ScienceAgentBench and AstaBench, and a win rate against the initial model increased from 53.9% to 69.9%.
Two lessons emerge: multi-agent traces are a better training material (a student trained on them beats a single-agent student fed 1.4 times more tokens), and the limiting link is the flow optimizer rather than the evaluator. The transfer to software engineering, on the other hand, remains disappointing, with slightly lower scores than the basic model on SWE-bench and Terminal-Bench. The code is published in open source.
Read the full article on the original source
Go to the source →