Sakana's MASS Recycles One 27B Model as Solver, Judge, Optimizer

A single open-weights 27B model, Qwen3.6-27B, plays all three roles in Sakana AI's MASS paper: task solver, workflow optimizer, and evaluator judging its own output, with no external supervisor in the loop.

Score per output token, measured against the untouched base, hit 1.60x on MLR-Bench, 1.35x on DSBench, 1.28x on AstaBench, and 1.20x on ScienceAgentBench. On SWE-bench Verified and Terminal-Bench 2.0, both coding-heavy, the same system fell to 0.94x and 0.96x. Head-to-head win rate against the base rose from 53.9% after one round to 69.9% after two.

"MASS repeatedly optimizes multi-agent workflows, then distills collective intelligence from multi-agent systems and uses it as a self-supervision signal for RSI," the paper writes. On the token-efficiency claim, the authors report their multi-agent setup reached a 68.3% win rate at 28.8M supervised tokens, against a single-agent variant at 42M: "The MASS model thus achieves the higher score against this reference with 31% fewer supervised tokens."

The paper then names its own biggest problem. "In a homogeneous loop, a shared blind spot could allow a flawed result to be produced, accepted as an improvement, and incorporated into the next model," the authors write, calling the risk "an Achilles' heel of recursive self-improvement" and pointing at chain-of-thought monitorability and intelligence-explosion concerns as open control questions.

The 2026 preprint is credited to Hyunin Lee of UC Berkeley with Jinglue Xu, Jeffrey Seely, and Yujin Tang of Sakana AI. It arrives into a week of AI safety coverage on our tracker that includes an OpenAI grader model that wrecked its own virtual machine to force a reset.