Tokyo based Sakana AI, along with researchers from UC Berkeley, has introduced Multi-Agent Self-Supervision (MASS), a framework designed to help AI models improve by learning from the collective work of teams of virtual subagents.

Researchers have referred to MASS as the process of self-supervision by distilling the team’s collective knowledge and experience. The framework explores a form of recursive self-improvement (RSI) in which a shared language model performs three roles: solving tasks, optimizing the workflow used by its subagents, and evaluating their results.

Instead of relying on an external evaluator within the improvement loop, MASS uses the model’s own judgments to select promising executions and fine-tune the model on them.

Introducing Multi-Agent Self-Supervision (MASS): Recursive self-improvement through collective intelligence

Blog: https://t.co/LkQe5k7jOW

As AI systems tackle increasingly open-ended research problems, human experts may struggle to assess results and guide further progress. The model itself may be the best available optimizer and evaluator. The challenge is obtaining useful self-supervision within that homogeneous loop, where a model risks reinforcing its own blind spots rather than correcting them.

We ask whether a model can learn from the collective intelligence of its own team.

We introduce Multi-Agent Self-Supervision (MASS). One shared model solves tasks through a team of virtual subagents, searches for better multi-agent workflows, then uses its own judgments to select the best ones.

Training on the best team’s executions provides self-supervision by distilling the team’s collective experience into the shared model.

We ran two cycles using a 27B open-weights model on synthetic open-ended research tasks. After the second cycle, score per output token reached 1.2-1.6x the base model’s level across four research benchmarks.

Three key results:

First, MASS enables role generalization within an RSI loop: training only on task-solving data from the optimizee naturally improves the capabilities of both the optimizer and the evaluator.

Second, multi-agent trajectories are more efficient training data than single-agent trajectories for RSI.

Third, optimizer capacity, rather than evaluator capacity, can be the bottleneck in the speed of improvement.

We believe the community needs to investigate homogeneous RSI carefully to support AI safety. When a model is also its own evaluator, shared blind spots could allow mistakes to be accepted as progress and reinforced through training. For AI to improve beyond human expertise safely, we need to understand these failures and preserve effective ways to detect them and intervene.

Paper: https://t.co/SzXDlAzrUO
Code: https://t.co/xc870ZHRgh

This work was a collaboration with researchers at UC Berkeley.

— Sakana AI (@SakanaAILabs) October 11, 2026

According to Sakana AI, MASS has been introduced to address a challenge: human experts may struggle to assess complex outputs, particularly when verifying a proposed discovery requires specialized knowledge, costly evidence or experiments.

The situation then leads to the frontier model itself turning into its most capable supervisor through homogenous RSI.

“The central question is how to obtain useful self-supervision within this homogeneous loop. We introduce Multi-Agent Self-Supervision (MASS). One shared model solves tasks through a team of virtual subagents, searches for better multi-agent workflows, then uses its own judgments to select the best ones.Training on the best team’s executions provides self-supervision by distilling the team’s collective experience into the shared model,” read a post by Sakana AI.

How MASS Works?

According to Sakana AI, MASS uses workflow optimization to elicit collective intelligence, then distills the team’s experience into the shared model.

MASS operates through two interconnected loops. In the inner loop, the model searches for better multi-agent workflows while its weights remain fixed. These workflows specify the roles assigned to subagents, their instructions, the outputs they must provide and the order in which they communicate.

The model evaluates the results produced by different workflows and uses those judgments to refine the process. In the outer loop, it selects high-ranking team executions and fine-tunes the shared model using the recorded conversations, tool calls and results. The updated model then resumes its roles as task solver, workflow optimizer and evaluator.

“Using a 27-billion-parameter open-weights model, the researchers completed two full cycles on synthetic research tasks in domains such as finance, robotics, and pharmacy. After the second cycle, the model achieved 1.2 to 1.6 times higher scores per output token across four public research benchmarks compared with its starting point,” read a post from Sakana AI.

Despite the reported gains, MASS leaves an important question unresolved: whether an AI system can reliably identify its own mistakes as its capabilities improve. Sakana AI cautions that shared blind spots could cause a model to accept flawed outputs as progress and reinforce those errors through subsequent training. The findings therefore represent an experimental step towards recursive self-improvement, rather than proof that AI systems can improve indefinitely without human oversight.

Also Read: “Father of Modern AI” Jürgen Schmidhuber Joins Sakana AI as Chief Scientific Advisor