Sakana AI News 10/12/2026 AI Rating: Medium

Sakana AI Publishes MASS: a Recursive Self-Improvement Method Using Multi-Agent Self-Supervision

#Sakana AI#MASS#recursive self-improvement#multi-agent#self-supervision#research

Sakana AI, together with researchers from UC Berkeley, published a new research paper on October 12, 2026, introducing Multi-Agent Self-Supervision (MASS), a method for recursive self-improvement (RSI) in which a single shared model plays the role of an entire team of virtual subagents, then learns from its own team’s best work.

Details

  • Core idea: one shared model solves tasks by acting as a team of subagents, searches for better multi-agent workflows among its own attempts, and uses its own judgment to select the strongest ones; the best team’s combined experience is then fed back into training the same shared model
  • Central question: the authors frame the work as testing “whether a model can learn from the collective intelligence of its own team”
  • Experimental setup: two rounds of recursive improvement applied to a 27-billion-parameter open-weight model, using synthetic open-ended research tasks
  • Results: after the second improvement cycle, score-per-output-token reached 1.2x to 1.6x the base model’s level across four research benchmarks
  • Key findings: training only on task-solving data improved performance in both the “optimizer” and “evaluator” roles within the loop; multi-agent trajectories proved more data-efficient for training than single-agent trajectories; optimizer capacity, rather than evaluator capacity, may be the bottleneck limiting the pace of improvement
  • Safety caveat raised by the authors: because the same model both performs and evaluates the work, shared blind spots could let mistakes pass as progress and get reinforced through further training — the paper calls for careful study of this “homogeneous RSI” so failures can be detected and intervened on
  • Openness: the work is published as an arXiv paper (2610.12176) with accompanying open-source code on GitHub
  • Authors: Hyunin Lee, Jinglue Xu, Jeffrey Seely, and Yujin Tang of Sakana AI, with UC Berkeley collaborators Somayeh Sojoudi, Matei Zaharia, and Donghyun Lee

What happened next

MASS adds to the growing body of work on self-improving AI systems, following Sakana’s earlier research into agent and robot trajectory scaling. Because the method is published with a full paper and open code rather than as a product announcement, independent researchers can now attempt to reproduce the 1.2-1.6x gains and probe the safety concerns the authors themselves flagged — particularly whether a model marking its own homework at scale can mask systematic errors rather than catch them.