Sakana AI’s self-improving MASS method boosts task scores but risks reinforcing hidden errors
A new method from Sakana AI and UC Berkeley lets a single language model improve itself across solving, judging, and refining tasks. While win rates…
Read more