Sakana AI introduces MASS, a recursive self improvement method where one model acts as a team, optimizes its workflow and trains on its best runs
Introducing Multi-Agent Self-Supervision (MASS): Recursive self-improvement through collective intelligence Blog: 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: t.co/SzXDlAzrUO Code: t.co/xc870ZHRgh This work was a collaboration with researchers at UC Berkeley.
Context
The paper offers a concrete recipe for a model improving itself without human graders, while its authors warn that the same setup could reinforce shared blind spots.
Sakana AI is a Tokyo based research lab, and this work was a collaboration with researchers at UC Berkeley, including Matei Zaharia. Recursive self improvement, or RSI, means a model improving the very process that improves it. MASS stands for Multi Agent Self Supervision.
Deep dive
On October 11, Sakana AI announced Multi Agent Self Supervision, or MASS. In MASS, one shared model solves tasks through a team of virtual subagents, searches for better team workflows, and then uses its own judgments to pick the best ones. Training on the executions of the best team distills that collective experience back into the model, which then fills every role in the next cycle.
According to the paper, the team ran two cycles using Qwen3.6 27B, an open weights model, on synthetic open ended research tasks. After the second cycle, score per output token reached 1.2 to 1.6 times the base model's level across four public research benchmarks: MLR Bench, DSBench, ScienceAgentBench and AstaBench.
Sakana highlighted three findings. Training only on task solving data improved the model's abilities as optimizer and evaluator as well. Multiple agent trajectories made more efficient training data than single agent ones, with the paper reporting that a student trained on them beat one trained on 1.4 times more single agent tokens. And optimizer capacity, rather than evaluator capacity, can be the bottleneck in how fast the model improves.
Sakana said the community needs to study this kind of loop carefully for AI safety, because when a model is also its own evaluator, shared blind spots could let mistakes be accepted as progress.
What to expect next
Watch whether other groups reproduce MASS on larger models and run more than two cycles, which would show whether gains keep compounding or level off.
Also watch for follow up safety work on detecting when a self evaluating model accepts its own mistakes, the failure mode Sakana itself flagged, and for labs adopting similar self supervised loops in frontier training.