Sakana AI Publishes MASS, a 27B Recursive Self-Improvement Framework Hitting 1.2-1.6x Base Score per Output Token
Summary
Sakana AI released MASS (Multi-Agent Self-Supervision), a recursive-self-improvement framework in which one 27B open-weights model fills executor, workflow-optimizer, and evaluator roles. After the second cycle, score per output token reached 1.2-1.6x the base across MLR-Bench, DSBench, ScienceAgentBench, and AstaBench, with a 68.3% win rate using 31% fewer supervised tokens than single-agent baselines. The paper flags a central safety concern: shared blind spots when a model grades its own outputs could get reinforced through training.
Shared on Bluesky by 3 AI experts
-
hardmaru @hardmaru.bsky.social
New Sakana AI paper: Recursive Self-Improvement through Multi-Agent Self-Supervision (MASS). Blog: pub.sakana.ai/mass/ Paper: arxiv.org/abs/2610.12176 How do we keep improving AI when tasks become too complex for human…
View on Bluesky →
Originally reported by sakana.ai
Read the original article →Original headline: Sakana AI Publishes MASS, a 27B Recursive Self-Improvement Framework Hitting 1.2-1.6x Base Score per Output Token