Sakana AI and Berkeley researchers show a model improving itself over two rounds

A paper by researchers including Yujin Tang and Matei Zaharia, published on Sakana AI’s site with code on GitHub, proposes Multi-Agent Self-Supervision (MASS) for self-improvement on open-ended tasks such as research that cannot be checked automatically. One model acts as executor, workflow optimiser and evaluator: an inner loop searches for better multi-agent workflows and an outer loop fine-tunes the model on the resulting traces. Using Qwen3.6-27B over two rounds, performance per output token reached 1.2 to 1.6 times the base model on four open-ended benchmarks, with no gain on Terminal-Bench 2.0 or SWE-bench Verified; win rate against the base model on held-out synthetic tasks rose from 53.9% to 69.9%. The authors say the test set is small and the results are their own evaluation.

When

8 Oct 2026

ID

WTA-N-2026-0286

Where

Japan, United States

Who

Sakana AI, University of California, Berkeley; Yujin Tang, Matei Zaharia

Topics

Self improvement, Research, Open-weight models

Earlier in this Storyline: What led hereSakana AI opens a recursive self-improvement lab in Tokyo with Schmidhuber as adviser24 Sep 2026

Spotted an error? Email [email protected].