MASS uses multi-agent self-supervision for recursive improvement
What happened
On October 12, Hacker News featured the MASS research. One model takes on three roles: task execution, workflow optimization and evaluation. Multi-agent collaboration trajectories train a shared model, producing recursive self-improvement. Starting from Qwen3.6-27B, two rounds of training brought the model to a 69.9% win rate over the base model on 3 synthetic test tasks. GPT-5.5 and Claude Opus 4.8 were used only for external evaluation, not self-supervision. The disclosed results come from those synthetic tasks.
Written by AI from the coverage · updated 1 hour ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageRecursive self-improvement through collective intelligence
MASS 让同一模型承担执行、工作流优化和评估，以多智能体协作轨迹训练共享模型，实现递归自我改进。以 Qwen3.6-27B 为起点完成两轮训练后，模型在 3 个合成测试任务上对基础模型的胜率达到 69.9%，GPT-5.5 和 Claude Opus 4.8 仅用于外部评估，不参与自我监督。
Heat over time
Not enough continuous observations to draw a trend yet.