研究团队提出MASS方法实现AI递归自我改进

Sakana AI研究团队提出了一种名为MASS（多智能体自监督）的新方法，通过让单一AI模型同时扮演执行者、优化器和评估者三个角色，实现了递归自我改进 1。该方法将多智能体系统的集体智能蒸馏为自监督信号，使模型能够在无需外部监督的情况下不断优化自身性能 1。

在Qwen 3.6-27B模型上进行的两轮训练中，该方法在三个合成测试任务上的赢率从基础模型的53.9%提升至69.9% 1。在MLR-Bench、DSBench、ScienceAgentBench和AstaBench等多个研究基准上，每个输出token的得分达到基础模型的1.2-1.6倍 1。与单智能体轨迹相比，多智能体轨迹在递归自我改进中表现更加有效，使用31%更少的监督训练token就能达到相同性能 1。

研究还发现，优化器可能是递归自我改进的瓶颈，强化优化器的能力比强化评估器更为关键 1。研究团队观察到最优多智能体工作流倾向于在契约和信息跳转中编码更多任务信息，而在角色和指令中编码较少 1。

Researchers have introduced MASS, a multi-agent self-supervised approach that enables a single AI model to simultaneously assume the roles of executor, optimizer, and evaluator to achieve recursive self-improvement 1. By distilling the collective intelligence of multi-agent systems into self-supervised signals, the method demonstrates substantial performance gains when applied to the Qwen 3.6-27B model 1.

Over two training rounds on three synthetic test tasks, the win rate improved from 53.9% in the initial round to 69.9% in the second round 1. On established research benchmarks including MLR-Bench, DSBench, ScienceAgentBench, and AstaBench, the model achieved 1.2 to 1.6 times the performance per output token compared to the baseline model 1. Multi-agent trajectories proved more effective than single-agent trajectories for recursive self-improvement, requiring 31% fewer supervised training tokens to reach equivalent performance levels 1.

Analysis of the optimization process revealed that the optimizer rather than the evaluator may represent the critical bottleneck in recursive self-improvement, making strengthening the optimizer more essential than enhancing the evaluator 1. Optimal multi-agent workflows tend to encode more task information within contracts and information transfers while encoding less in roles and instructions 1. The work, authored by Lee Hyunin, Xu Jinglue, and others, has been published on arXiv as preprint 2610.12176 1.