GoKawiilSakana AI unveils MASS, a method for AI systems to self-supervise their own improvement
Sakana AI, working with UC Berkeley researchers, has published a paper introducing MASS, a technique that repeatedly optimizes multi-agent AI workflows and then distills the resulting 'collective intelligence' back into a self-supervision signal for a single base model. The authors report that this approach improves the score-per-output-token performance of the base model on research benchmarks compared to not using the method.
GoKawiil's interpretation of the reporting above, not reported fact.
The work targets a known bottleneck in AI progress: as models approach or exceed the frontier of human expertise, humans may no longer be able to reliably evaluate or guide further improvements, especially on tasks without clear verification criteria. By having the model act as its own optimizer, optimizee, and evaluator, the researchers suggest a path toward continual self-improvement that doesn't depend on human oversight at every step, though the practical safety and reliability implications of such recursive self-improvement remain to be tested at scale.
- MASS uses multi-agent workflow optimization to generate training signals for improving a single base AI model.
- Sakana AI and UC Berkeley researchers report gains in score-per-output-token on research benchmarks versus the base model.
- The approach is framed as a step toward AI systems that can supervise and improve themselves without relying on human evaluation.
Source: pub.sakana.ai, 2026-10-11
Published there as: “Recursive self-improvement through collective intelligence”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.