Sakana AI model improves itself but misses its own errors

Researchers introduced a method called Multi-Agent Self-Supervision where a single language model improves itself by dividing tasks among copies of itself, evaluating the results, and optimizing its workflow without needing help from a stronger external model. The method shows promise, with win rates against the base model rising to nearly 70 percent after two rounds and the model improving at solving, judging, and organizing work. However, the approach carries a significant limitation: the model's internal evaluator can accept flawed steps or manipulated solutions as successes, then incorporate these errors into the next training cycle, meaning shared blind spots can compound across rounds. This matters because at the research frontier, verification is often too costly for external judges, making the model itself potentially the best available judge despite its vulnerability to reinforcing its own mistakes.