MASS paper: one model designs, runs and grades its own agents

A single language model proposes a multi-agent workflow, executes it, grades the output, then fine-tunes on the winning traces. That is the recursive loop sketched in a paper posted to arxiv on October 8 by Hyunin Lee, Matei Zaharia, Yujin Tang and co-authors.

The authors call the method MASS, for Multi-Agent Self-Supervision. It "alternates between evolutionary workflow optimization and supervised fine-tuning on self-generated trajectories," prompting one base model to iteratively propose, execute and self-evaluate multi-agent workflows. Over two MASS cycles with Qwen3.6-27B, the paper reports "1.2-1.6x higher performance per output tokens on four open-ended public benchmarks." Multi-agent traces also look more training-efficient: a student trained on them "outperforms a single-agent student trained on 1.4x more training tokens." Three researchers in our Who's Who tracker shared the arxiv link soon after it posted.

The abstract frames the motivation as a saturation argument: when a model's outputs exceed human expertise, the model itself becomes the best available evaluator. That logic collapses optimizer, evaluator and student into one base model. "Training only on task-solving data from the optimizee naturally improves the capabilities of both the optimizer and the evaluator," the paper argues. The four benchmarks are not named in the abstract, and no absolute scores are published alongside the per-token ratios.