Sakana AI and Berkeley Propose MASS for Recursive Self-Improvement

Sakana AI and UC Berkeley proposed MASS, using multiple instances of a model to supervise each other for recursive self-improvement, boosting per-token performance 1.2–1.6x in two iterations, though Dietterich questioned whether the supervision bottleneck is truly solved.

2026-10-13 ~ 2026-10-13 · 4 related posts

• Sakana AI and Berkeley propose MASS: multi-agent self-supervision for recursive self-improvement — TheTuringPost · 2026-10-13
• MASS paper: two cycles lift Qwen3.6-27B per-token performance 1.2–1.6x via multi-agent self-supervision — TheTuringPost · 2026-10-13
• Sakana AI's MASS self-improvement paper draws skepticism from Thomas Dietterich — tdietterich · 2026-10-13

1 near-duplicate retellings: hardmaru