Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots

Sakana AI has developed the MASS method, which allows a language model to iteratively improve its performance by solving tasks, evaluating its own results, and refining its workflows. This self-improvement process increased the model's win rate by nearly 70 percent against its original version after two cycles. However, the system remains prone to systematic errors, as the model often fails to identify its own mistakes when it repeats the same flawed reasoning across the solving and judging phases.