Sakana AI and UC Berkeley release MASS: a 27B model acts as its own team, optimizer and judge for two rounds of recursive self-improvement

Why it matters

【Key points】

Sakana AI and UC Berkeley released Multi-Agent Self-Supervision (MASS) on October 11. One model plays three roles: a multi-agent team doing the work, an optimizer rewriting the team's workflow, and an evaluator comparing results. An inner loop revises workflows and keeps the self-judged best; an outer loop runs that workflow 18 times, self-ranks the outputs, keeps up to 16 trajectories for LoRA fine-tuning, and the updated model takes over all three roles again.

Starting from Qwen3.6-27B over two rounds, external judges (GPT-5.5, Claude Opus 4.8) preferred the round-2 model over the base 69.9% of the time on held-out synthetic research tasks in finance, robotics and pharma. On MLR-Bench, DSBench, ScienceAgentBench and AstaBench, score per output token rose to 1.2–1.6x the base. Training only on task-solving data also improved the optimizer and evaluator; evaluator agreement with external judges rose from 73% to 93%.

【Verifiable】

Blog at pub.sakana.ai/mass, paper arXiv:2610.12176 (submitted October 8), code at GitHub SakanaAI/mass. The paper notes score per token was 0.96x on Terminal-Bench 2.0 and 0.94x on SWE-bench Verified; on MLR-Bench delivered papers rose from 31/60 to 50/60 while their mean score fell from 3.07 to 2.88.

【Our read】

1. This is a small empirical test of a model acting as its own teacher. The gain is in efficiency, not a capability jump, and two rounds cannot show whether the loop keeps compounding.

2. The notable finding is that training only on task work also strengthened the optimizer and judge. If it holds at scale, the bottleneck shifts from outside supervision to the optimizer itself.

3. The authors flag the risk themselves: when a model grades itself, shared blind spots can be trained in as progress. That they relied on outside models to measure gains shows fully closed-loop improvement still needs independent evaluation.