Sakana AI has introduced a research framework called Multi-Agent Self-Supervision (MASS) that lets a single language model improve itself without external human feedback or hand-labeled answers. The idea is to have the model act simultaneously as the worker solving a problem, the optimizer refining how a team of its own copies collaborate, and the judge deciding what counts as a good result. It matters because it tackles a core bottleneck in AI progress: once models approach or exceed human expertise on certain tasks, humans may no longer be able to reliably check their work, so the model itself may need to become its own best supervisor.

The problem: who checks an AI that is smarter than its checkers

Most AI training today relies on some form of external supervision, whether that is human-labeled data, human feedback, or automatic rule-based scoring like unit tests or math answer-checking. That works well for tasks with clear right answers. But Sakana AI points to a harder category: non-verifiable research tasks, where there is no automatic rule to confirm correctness, and where even expert humans may lack the time, money, or specialized knowledge to properly evaluate a result. Judging a novel scientific proposal, for instance, can require expensive experiments that take far longer than generating the proposal itself.

This creates a supervision gap. If AI is meant to eventually produce discoveries beyond the existing body of human knowledge, there may not always be a reliable external judge available. Sakana AI’s answer is what they call homogeneous recursive self-improvement (RSI): the same model plays the role of the thing being improved, the optimizer driving the improvement, and the evaluator judging whether the improvement is real.

How MASS works: a two-loop system inspired by collective intelligence

MASS is built around an analogy to human societies. Progress in science and industry has rarely come from a single person working in isolation; it comes from groups dividing labor, combining ideas, and passing improvements on to others. Sakana AI cites philosopher Thomas Kuhn’s observation that scientific communities learn partly through shared examples of past achievements. MASS tries to recreate that dynamic inside a single AI system by forming a multi-agent team out of copies of the same model.

The framework runs two nested loops:

• The inner loop keeps the model’s weights fixed and searches for the best way to organize a multi-agent workflow, meaning how the agents should divide a task, share results, and check each other’s work. The model’s own evaluator role scores different workflow configurations and picks the best one.
• The outer loop then updates the model’s actual weights by training it on the executions produced by that optimized, self-judged-best workflow. This distills the team’s collective output back into a single shared model.

The cycle then repeats: the newly updated model goes back to filling all three roles, task solver, workflow optimizer, and evaluator, ideally starting the next round from a stronger baseline. Sakana AI describes this as distilling a team’s collective experience into self-supervision signal that drives recursive self-improvement, since no external labels are used to tell the model what “good” looks like.

What makes this different from standard AI training approaches

Conventional reinforcement learning from human feedback (RLHF) or supervised fine-tuning relies on an external source of truth, either human raters or verifiable answers. Multi-agent AI systems are also not new; many products already chain multiple model calls together to divide labor on a task. What sets MASS apart is that the same underlying model performs every function in the loop, including judging its own output, and that judgment is then used to directly update the model’s weights, closing the loop without a human or separate reward model in between.

The task-solving agent in MASS produces a tangible workspace, which the research describes as containing results, plots, and a written research report, rather than just a short text answer. The evaluator’s job is to compare two such workspaces, decide which one better satisfies the task, and explain its reasoning, and that comparison is what feeds back into revising the workflow.

How the results were measured and what it means in practice

Because MASS trains on non-verifiable research tasks, where there is no ground-truth answer to check against, Sakana AI needed an independent way to tell whether the loop was actually producing improvement rather than the model simply rewarding itself for mediocre work. To do this, they used two external models, GPT-5.5 and Claude Opus 4.8, to separately assess the completed work, and crucially, those judgments were withheld from the MASS training loop itself. According to Sakana AI, MASS improved its score per output token over the base model on research benchmarks, suggesting the self-supervised loop produced genuinely better outputs rather than just different ones.

For developers and researchers, MASS is currently a research framework and proof of concept rather than a ready-to-use product, and Sakana AI frames it explicitly as a step toward what it calls AI’s “Odyssean moment,” a period of continual self-learning where the thing being optimized and the thing doing the optimizing are one and the same. The practical caveat worth keeping in mind is that self-judged improvement is not the same as externally validated correctness; using outside models like GPT-5.5 and Claude Opus 4.8 as a check is a useful safeguard here, but it also underscores that fully autonomous, self-supervised AI improvement is still an open research problem rather than a solved one.

Frequently asked questions

What does “non-verifiable task” mean in this context? It refers to tasks where there is no automatic rule or ground-truth answer to check correctness, such as open-ended research work, as opposed to tasks like math problems or code tests that can be automatically graded.

Is MASS a publicly available tool? Based on the available material, MASS is presented as a research framework from Sakana AI, published as a research article rather than a released commercial product.

Why use multiple agent copies instead of one model working alone? Sakana AI’s reasoning is that a team of agents dividing work and checking each other can surface collective intelligence and workflow patterns that a single agent working in isolation may not reliably produce on its own.

Source: Recursive self-improvement through collective intelligence