The Decoder· Jonathan Kemper·· 3 天前AI 评分12
Sakana AI MASS 方法实现模型自我改进但存在盲区
Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots
AI 导读
Sakana AI 推出 MASS 方法，让 Qwen3.6-27B 通过多智能体协作实现递归自我改进。该方法在 MLR-Bench 等基准测试中得分提升 1.2 至 1.6 倍，两轮后胜率升至 69.9%。然而，由于共享盲区，错误可能被误判为进步并进入下一轮训练循环。
来源：The Decoder · the-decoder.com