Sakana AI提出MASS:多智能体自监督递归自我改进
Recommended paper from Sakana AI on recursive self-improvement.
递归自我改进是前沿方向,MASS去除了对外部验证器的依赖,且给出了具体的性能提升数据和对比实验,对探索Agent自进化有参考价值。
Recommended paper from Sakana AI on recursive self-improvement.
Sakana AI 关于递归自我改进的推荐论文。
They propose an interesting way to scale recursive self-improvement through multi-agent self-supervision.
他们提出了一种通过多代理自监督来扩展递归自我改进的有趣方法。
In this line of research, self-improvement loops usually need an external verifier, so open-ended tasks without a checker are left out.
在这类研究中,自我改进循环通常需要外部验证器,因此没有检查器的开放式任务被排除在外。
MASS removes that requirement.
MASS 消除了这一要求。
One base model proposes multi-agent workflows, runs them and grades them, and an evolutionary search keeps the workflows that score best. The model is then fine-tuned on its own traces, and the improved model starts the next cycle as a better optimizer and grader.
一个基础模型提议多代理工作流,运行并评估它们,进化搜索保留得分最高的工作流。随后,该模型在其自身的轨迹上进行微调,改进后的模型作为更好的优化器和评分者开启下一个循环。
Two cycles on Qwen3.6-27B raise performance per output token from 1.2 to 1.6x on four open-ended benchmarks.
在 Qwen3.6-27B 上经过两个循环后,四个开放式基准测试中每个输出 token 的性能从 1.2 倍提升至 1.6 倍。
A student trained on multi-agent traces also beats a single-agent student trained on 1.4x more tokens.
在多代理轨迹上训练的学生模型也优于在 1.4 倍更多 token 上训练的单一代理学生模型。
Paper: https://arxiv.org/abs/2610.12176
论文:https://arxiv.org/abs/2610.12176
Chat with Paper: https://academy.dair.ai/papers/recursive-self-improvement-through-multi-agent-self-supervision-2610.12176
与论文对话:https://academy.dair.ai/papers/recursive-self-improvement-through-multi-agent-self-supervision-2610.12176
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力