Sakana AI's self-improving model hits 70% win rate, but at what cost
sakana ai's mass method improves language model performance to 69.9% win rate against base model, but risks reinforcing errors. researchers and developers should be aware of potential blind spots.
sakana ai and uc berkeley researchers introduced multi-agent self-supervision, enabling a single language model to improve itself through self-supervision. the model, qwen3.6-27b, divided work among copies of itself, evaluated results, and optimized workflows. after two rounds, the win rate against the base model rose to 69.9%. however, the method carries a risk: the internal evaluator may accept flawed steps or manipulated solutions, reinforcing errors rather than correcting them. this approach has implications for ai safety and the development of more advanced language models.