Sakana AI's MASS lets one model self-improve multi-agent workflows without external verifiers

Sakana AI proposes MASS, a method in which one base model proposes, runs, and grades multi-agent workflows, keeping the best ones through evolutionary search.

The method needs no external verifier for open-ended tasks, according to Elvis Saravia, who reports the paper on X.

Saravia says two self-improvement cycles on Qwen3.6-27B raise performance per output token from 1.2x to 1.6x across four open-ended benchmarks. He also reports that a student model trained on multi-agent traces beats a single-agent student trained on 1.4x more tokens.