Sakana AI's MASS Method Enhances Self-Improvement in Language Models

Sakana AI has developed the MASS method, allowing a language model to enhance its performance by solving tasks and evaluating results. After two training rounds, the model achieved a nearly 70% success rate compared to the base model. This method improves the model's efficiency through self-learning, but there is a risk that the model may fail to recognize recurring errors. When the model misses the same mistakes across all three roles, it can mistakenly perceive progress and reinforce those errors during training.