Morning Edition · Tuesday, October 13, 2026Published at 2:13 AM EDT · New York

Sakana AI and UC Berkeley train a 27B model on its own multi-agent runs, raising its win rate to 69.9% after two rounds

In the method, called Multi-Agent Self-Supervision (MASS), a single Qwen3.6-27B model acts as executor, workflow optimizer and judge without a stronger teacher. The authors say errors that all three roles miss get reinforced as progress.

Sakana AI, the Tokyo lab, and researchers at UC Berkeley published MASS on arXiv on October 8. The method lets a model improve without human labels or an external teacher model, AI ML Big Data reported. One shared model plays three roles. A…