Sakana AI's MASS teaches multi-agent workflows to self-improve

Sakana AI and UC Berkeley researchers have posted MASS, a framework in which a single language model plays three roles in a multi-agent loop, task executor, workflow optimizer and evaluator, before fine-tuning itself on its own successful runs. In two rounds of training, the Round 2 model beat the base model on test tasks 69.9% of the time and beat its own Round 1 version 60.8% of the time, according to the paper posted at pub.sakana.ai.

The authors state the goal plainly: "MASS repeatedly optimizes multi-agent workflows, then distills collective intelligence from multi-agent systems." Training ran on 12 synthetic finance, robotics and pharmacy tasks, with workflows built from four named components: roles, instructions, contracts and hops.

The headline numbers extend to four public benchmarks. The paper reports a 1.60× score-per-token gain on MLR-Bench, 1.35× on DSBench, 1.20× on ScienceAgentBench and 1.28× on AstaBench. On token efficiency, multi-agent trajectories hit a 68.3% win rate while using 28.8M supervised tokens against 42M for the single-agent comparison, which the authors describe as 31% fewer.

One of the three observations the paper highlights comes from an ablation that upgraded one role at a time. "The bottleneck in RSI can be the optimizer, not the evaluator," the authors conclude, noting that upgrading only the evaluator showed minimal gains compared to upgrading the optimizer. Three researchers we track in our Who's Who directory posted the link to their feeds.

The authors flag their own blind spot. Because the same model fills every role, flawed outputs could propagate through successive updates if humans also cannot detect them, a risk the paper raises explicitly in its safety section.