Sakana AI 自改进模型 MASS：解题与评判变强，却难察自身盲点
原标题：Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots
AI 摘要
Sakana AI 与 UC Berkeley 提出 Multi-Agent Self-Supervision（MASS）方法，让单一语言模型通过自我分配任务、评估结果并优化工作流实现递归自我改进，无需更强的教师模型。基于 Qwen3.6-27B 的实验显示，两轮后对基础模型的胜率从 53.9% 升至 69.9%，评估器与外部评委一致率从 73% 升至 93%。但该方法存在风险：内部评估器有时会把有缺陷或被操纵的步骤当作成功，并固化进下一轮模型，且逐任务比较时增益不再显著。
正文节选
Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots Key Points - A single language model improves itself through "Multi-Agent Self-Supervision" (MASS), without external help or a stronger teacher model. - 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. - Win rates against the base model rise to nearly 70 percent after two rounds, but the metho