Definition
Multi-Agent Self-Supervision (MASS) is an recursive-self-improvement method where one model acts as executor, workflow optimizer, and evaluator, distilling successful multi-agent team trajectories into fine-tuning signal for non-verifiable open-ended tasks.
Key Points
- 2026-10-11: sakana-ai reports two MASS cycles on Qwen3.6-27B yielding ~1.2–1.6× score per output token on four research benchmarks; coding benchmarks showed no gain (2026-10-11-sakana-mass-recursive-self-improvement)
- Authors warn homogeneous evaluator loops can reinforce shared blind spots—safety framing is explicit in paper and blog
- Code released on GitHub; arXiv 2610.12176