Sakana AI trains a model on its agents' collective research work
Sakana AI's MASS method uses one model to organize research agents, judge their results and learn from selected team runs, with gains on research benchmarks and weaker transfer elsewhere.
By RuntimeWire Staff · Published
Primary source: Sakana AI
Why it matters
MASS tests whether agent coordination can become training data for a model's next generation. Its early research gains make the idea concrete, while uneven benchmark transfer and just two cycles leave long-run reliability and real-world costs unresolved.
Sakana AI co-founder and CEO David Ha has made self-evolving AI one of the Tokyo lab's research interests. A new project from Sakana AI and UC Berkeley, described on Sakana AI's MASS project page, tests that idea: use a model to organize research agents, let it judge their work, then train the model on the strongest team results.…