Sakana AI and UC Berkeley introduce MASS, a framework for recursive self-improvement using multi-agent workflows as self-supervision

The article presents MASS, a Multi-Agent Self-Supervision approach that optimizes multi-agent workflows, distills their collective output, and uses it to improve the underlying AI model. The authors say it improves score per output token over the base model on research benchmarks and have released related paper and code.

MASS reflects a broader push toward AI systems that can generate, evaluate, and absorb their own training signals as human supervision becomes a bottleneck. If reliable, this could accelerate research automation, but it also intensifies concerns around recursive self-improvement where the evaluator and optimizee are increasingly the same system.