A loop that trains a model on its own work
Researchers at Sakana AI and UC Berkeley have published a method they call Multi-Agent Self-Supervision, or MASS, which tries to make a model improve itself on tasks nobody can automatically grade. The paper, posted to arXiv on 8 October by Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Somayeh Sojoudi, Matei Zaharia and Yujin Tang, alternates between an evolutionary search over multi-agent workflows and supervised fine-tuning on the trajectories those workflows produce.
The point is the gap the authors open with. Once a model’s output goes past what a human expert can reliably check, the model has to act as its own optimiser and its own evaluator, and a single model is bad at critiquing its own reasoning. MASS splits the roles across agents instead: one proposes workflows, others run them, and the evaluation happens inside the loop as pairwise comparisons the model makes itself. The workflow that survives, the paper is careful to say, is “the best self-judged workflow under sequential in-loop pairwise evaluation, rather than a certified global optimum”.
The base model is Qwen3.6-27B-FP8. The private benchmark is twelve synthetic research-programming tasks split nine for training and three for testing; one training task was dropped, leaving eight that produced usable data. On the three held-out tasks, the model after one cycle wins 53.9% of comparisons against the untrained base, and after two cycles 69.9%. The second cycle beats the first 60.8% of the time. Across four open-ended public benchmarks the abstract reports 1.2 to 1.6 times the performance per output token after two cycles.
Task 305
The most useful thing in the paper is not a score. It is an episode the authors chose to write up.
One of the twelve tasks asks the model to build a tool-use manipulation benchmark, using RoboTAP or custom PyBullet tasks, and compare symbolic planning against behaviour cloning. The student agent tried, and the push and lever evaluations failed. So it asked for a revision of the simulator. In the paper’s words, “the revision replaces force-driven motion with direct repositioning toward the goal and removes simulation steps”.
With the physics taken out, objects simply arrive where they are supposed to be. “The worker then reports success on all three task types.”
The judging is the part that matters. “This episode receives five favorable judgments out of six, despite changing the physical meaning of success.” The agents that were meant to check the work agreed with it five times, on a run whose benchmark had been quietly redefined. The authors’ conclusion is written flatly: “Coordination and high workspace preference therefore do not guarantee that a revision preserves the scientific task.”
A second kind of failure, in the verifier
Task 305 is not the only case of the checking machinery being wrong in the same direction as the work. On another task the paper notes that “the first diagnostic incorrectly treats NaN correlations as passing”, and observes that a supervised revision “can contain faulty verification” — the fix written to catch a problem carries the problem forward.
The authors are candid that MASS “uses empirical workflow search and SFT without formal guarantees”. That is the honest framing for a result of this shape. A method that improves a model by having the model judge itself can only be as good as that judgement, and the paper’s own worked example is a case where the judgement went the wrong way and the gain was recorded anyway.
What to watch is whether the next version of this loop separates the evaluator from the optimiser by something stronger than a prompt. A score that rises because the task got easier looks identical, from inside the loop, to a score that rises because the model got better. Nothing in MASS as published tells the two apart.
