Sakana AI's MASS: one model forms a team and writes its own training data
What is it?
MASS (Multi-Agent Self-Supervision) has a single model play every role: a team of subagents that solve the task, an optimizer that rewrites the team's workflow, and an evaluator that grades the results. Once it finds a better workflow, the model is fine-tuned on the best team's runs, which folds the team's experience back into its weights. Experiments used the open-weight Qwen3.6-27B, and the code is public.
Why does it matter?
After two rounds, the win rate against the base model on held-out research tasks rose to 69.9%Vendor claim, and score per output token was 1.2-1.6x higher across four benchmarksVendor claim. Multi-agent traces also beat single-agent traces with less data. That's useful for any team thinking about recycling its own agent logs as training data.
Maturity
ResearchThis is research. Scores slipped slightly on Terminal-Bench 2.0 (0.96x) and SWE-bench Verified (0.94x)Vendor claim, and the authors warn that a model grading itself can reinforce its own blind spots. The repo doesn't declare a license yet.
How to track it
- Licensing: check whether the
SakanaAI/massrepo adds a license file. - Follow-up versions and reproductions: watch arXiv revisions and repo updates for fixes to the Terminal-Bench and SWE-bench drops.
- Other base models: watch the Sakana AI project page for experiments beyond Qwen3.6-27B.