Sakana AI MASS: what it is and what it found

MASS (Multi-Agent Self-Supervision) is a research method from Sakana AI and UC Berkeley in which a single AI model plays three parts: it does the work, redesigns how its team of copies does the work, and grades the results. The model then trains on its own best team runs, and after two rounds a 27B open-weights model got 1.2 to 1.6 times more score per output token on four research benchmarks, with no human or stronger "teacher" model in the loop.

Sakana announced it on X on Sunday. @SakanaAILabs framed it as "recursive self-improvement through collective intelligence," and co-founder @hardmaru followed with a thread on the paper. The paper is on arXiv as 2610.12176, the write-up is on Sakana's site, and the code is on GitHub.

How MASS works

The same model, Qwen3.6-27B, fills every role:

• As the agents, it solves a research task and produces code, results, plots and a report.
• As the workflow optimizer, it rewrites how those agents split up the work, pass results along and check each other.
• As the evaluator, it compares two finished workspaces, picks the better one, and says why.

There are two loops. In the inner one the weights stay frozen: the team tries a workflow, the evaluator checks it against the current best, and the optimizer suggests the next attempt. In the outer loop, the winning workflow runs 18 times per task, the model ranks those runs, and up to 16 of the best are used to fine-tune the model with LoRA. Then the updated model takes over all three roles again and the cycle repeats.

The interesting step is that Sakana strips the workflow text out of the training prompts. The model learns from what the team produced without seeing the instructions that organized the team, so the hope is that some of the team's habits end up living inside one model.

What Sakana says it found

Judged by outside models (GPT-5.5 and Claude Opus 4.8), the trained model beat the original on held-out tasks 53.9% of the time after round one and 69.9% after round two. Per Sakana's page, it also found winning workflows faster in later rounds, and the evaluator's agreement with the outside judges went from 73% to 93%.

@hardmaru's thread singles out one result: "Multi-agent traces were also much better training data than single-agent ones." The team says training only on task-solving data also made the model a better optimizer and judge, and that the optimizer, not the evaluator, seems to be what limits how quickly things improve.

The honest caveats

The paper is careful about its limits, and they're worth knowing before anyone writes "AI teaches itself" in a headline.

The gains showed up on research-style benchmarks (MLR-Bench, DSBench, ScienceAgentBench, AstaBench). On the two coding benchmarks, Terminal-Bench 2.0 and SWE-bench Verified, the model came in slightly below where it began, at 0.96 and 0.94 times. The training set is small: 12 synthetic tasks in finance, robotics and pharmacy, nine for training and three held out. Sakana also notes that the workflow search is noisy, with good iterations sometimes followed by sharp drops.

The biggest caveat comes from Sakana itself. When one model produces the work, designs the process and grades the outcome, a blind spot shared across all three roles can let a mistake count as progress and get trained back in. The authors say loops like this need careful study as a safety matter. That's the right instinct, and for anyone outside a lab it's the most useful part of the paper.

Can you use it?

You can run it, but it isn't a product. The GitHub repo has the pipeline, the 12 tasks and configs for both cycles. Its offline preview and tests run without a GPU, while the real experiments need a vLLM model server, GPUs and the model downloads. The README states no license and doesn't list hardware specs, and some benchmark adapters aren't included. This is a researcher's tool.

What it means for a business using AI

A system that checks its own work with the same eyes that produced it will miss the same things twice. That's true of an AI agent that writes a customer email and then rates it, and it's true of a person proofreading their own quote. If an automation in your business drafts, approves and sends with nothing checking it in between, add that check before anything else. It can be a different model, a simple rule, or a person.

If you'd like a second set of eyes on where AI could take work off your plate, and where it should never grade itself, New Face Design's free process audit will map both.