Sakana AI's MASS uses multi-agent teams as their own supervisor for recursive self-improvement

What Sakana AI proposed

Sakana AI has published a research write-up on Multi-Agent Self-Supervision (MASS), a method for recursive self-improvement (RSI) that surfaced on the Hacker News front page. The core argument: once AI models are pushed past the boundary of humanity's accumulated knowledge, human experts may no longer be able to assess what the model produces — especially on tasks with no automatic way to check correctness. At that frontier, the model itself may be the most capable supervisor available, so the training loop has to be built from its own judgments.

MASS addresses this with what the authors call homogeneous RSI: one shared language model fills every role in the loop. According to the write-up, the method repeatedly optimizes how a multi-agent team works together, then distills the team's collective output into the model's weights as a self-supervision signal. The researchers report that this raises scores per output token over the base model on research benchmarks.

One model, three roles

The setup assigns three jobs to the same model, as the publication describes:

• Agent: the task solver, which produces a workspace containing results, plots, and a research report.
• Workflow optimizer: revises how agents split up the task, exchange results, and check each other's work, using evaluator feedback.
• Evaluator: compares two workspaces, picks the one that better satisfies the task, and explains the reasoning.

Training relies on non-verifiable research tasks, meaning no rule-based signal exists and only self-evaluation — whose correctness is unknown — steers the learning process.

Workflows written as prompts

A key design choice is that a workflow is just a written plan, injected into the task prompt so the model itself can inspect and rewrite its own team structure. Each plan has four parts: a role defining a subagent's responsibility, an instruction for the work to do, a contract specifying what must be returned and checked, and a hop ordering the calls. An orchestrator agent assigns subagents and integrates their results; one evolved example in the write-up arranges data preparation and later steps for a finance task.

Two nested loops

The process runs as an inner and outer loop. In the inner loop, with model weights frozen, the system searches for the best team workflow: candidate workflows run, the self-evaluator judges the resulting workspaces, and the optimizer iterates until a champion emerges. The outer loop then trains on that champion — per the accompanying diagram, the champion workflow runs 18 times, the model self-ranks the resulting workspaces, the best 16 are kept, and the model is fine-tuned on those executions with the workflow stripped from the prompt. The updated model then reassumes all three roles and the cycle repeats.

Checking progress without ground truth

Because supervision inside the loop is entirely internal, the authors measure progress using outside judges: GPT-5.5 and Claude Opus 4.8 assess the completed work, and their verdicts are withheld from MASS. This provides an external yardstick for a process driven purely by self-judgment. The write-up also notes the research draws on an analogy to human societies, citing Thomas Kuhn's account of how scientific communities learn from shared examples, and that the work was done during an internship at Sakana AI.

Why it matters

Recursive self-improvement is one of the central open questions in AI, and its usual sticking point is evaluation: someone or something has to say whether an output is better. MASS proposes an answer — treat a team's collective performance as the teacher — and claims measurable gains on exactly the tasks where verifiers don't exist. If the approach scales, it points toward continual self-learning that no longer waits on human grading, which could noticeably accelerate AI progress on open-ended research work. The risk is symmetrical: a loop supervised by self-judgment of unknown quality could just as easily reinforce a model's blind spots, and the withheld-judge evaluation reports progress without correcting it. Expect both the efficiency claims and the failure modes to draw close scrutiny as the idea circulates.