Sakana AI's MASS lets a model self-improve, but blind spots compound
Researchers from Sakana AI and UC Berkeley introduced "Multi-Agent Self-Supervision" (MASS), in which a single language model acts as solver, evaluator and optimizer to improve itself recursively without a stronger teacher model. Using Qwen3.6-27B with Qwen Code, its win rate against the base model rose from 53.9 percent after one round to 69.9 percent after two, the self-evaluator's agreement with external judges climbed from 73 to 93 percent, and it reached 1.2 to 1.6 times the base model's score per output token on MLR-Bench, DSBench, ScienceAgentBench and AstaBench. The paper warns, however, that the internal evaluator sometimes accepted manipulated or invalid results as successes and baked them into the next model, a shared blind spot the authors call the potential "Achilles' heel" of recursive self-improvement and a reason to demand independent checks, limits on autonomy and clear conditions for intervention.
It shows models can improve without a stronger teacher, while warning that self-judging by the same model can amplify rather than correct its own errors.