Sakana Trains One Small Model to Run Its Own Research Team

Sakana AI has a new answer to a stubborn problem: how to get an AI model to keep improving itself when the tasks have no neat right answer. Its method, developed with researchers at the University of California, Berkeley, takes a single open-weights model and has it play every part in its own training loop.

The usual recipe for self-improvement leans on tasks with clear grading, like code that either passes its tests or does not. Open-ended research is different. There is no unit test for a good experiment design or a well-framed hypothesis, and models trying to improve on that kind of work tend to stall. Sakana's approach, which it calls MASS, splits the job into roles and assigns all of them to the same 27-billion-parameter model, a version of Qwen. One instance solves the task, another revises the workflow, a third evaluates the output, and the model is then fine-tuned on the trajectories it judged best.

The reported results are modest but pointed. Across four research benchmarks, Sakana says two training cycles raised the score the model achieved per unit of output by between 1.2 and 1.6 times. The team found that letting several coordinated instances tackle a problem produced a denser, more useful training signal than a single agent working alone, and that practice at solving tasks quietly sharpened the model's abilities as an optimiser and a judge as well. One finding stood out: the part of the loop that limits how fast the whole thing improves was the workflow optimiser, not the evaluator many would have suspected.

The caveats are real. Benchmark gains did not always carry over cleanly to other tasks, and the team tested only two cycles, so how far this can run before it plateaus, and at what cost, is still an open question.

The work fits a wider shift from single models toward orchestrated systems, where the interesting engineering is in how the parts are coordinated. That is the same instinct behind Anthropic's experiments with one lead agent directing a thousand subagents and Microsoft's Decision-1, which scores choices rather than writing them. Sakana's twist is to fold all those roles back into one model, and let it teach itself.