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

What happened

Sakana AI's MASS method lets one language model improve itself by solving tasks, judging results, and refining workflows, reaching a win rate of nearly 70 percent against the base model after two rounds. Win rates against the base model rise to nearly 70 percent after two rounds, but the method carries a real risk: the internal evaluator sometimes accepts flawed steps or manipulated solutions as successes and bakes them into the next model.

The study finds gains across all three roles, but shared blind spots can let errors pass as progress and feed into the next training cycle. Researchers from Sakana AI and UC Berkeley have introduced "Multi-Agent Self-Supervision" (MASS), a method for recursive self-improvement (RSI).

But when the model misses the same errors across all three roles, it can mistake them for progress and reinforce them during training. The model works as its own team: it assigns tasks to copies of itself, evaluates the results, and optimizes workflows for the next round of training.

Sakana AI's MASS method lets one language model improve itself by solving tasks, judging results, and refining workflows without a stronger teacher model. A single language model handles the task, optimizes how the work gets done, and evaluates the results. A single plasma study can take more than 120 million CPU (the general-purpose processor at the centre of a computer) hours, according to the paper .

Key facts

• Researchers from Sakana AI and UC Berkeley have — introduced: "Multi-Agent Self-Supervision" (MASS), a method for recursive self-improvement (RSI)