Synthetic Sandbox for Training Machine Learning Engineering Agents

Verifying the behavior of machine learning engineering (MLE) agents costs far more than software engineering (SWE). While SWE can be verified with fast unit tests, MLE requires running the entire pipeline—including data preprocessing, model training, and metric evaluation—at every step. This has made existing on-policy reinforcement learning (RL) approaches difficult to apply due to speed constraints.

SandMLE is a multi-agent framework proposed to address this bottleneck. It generates diverse, verifiable synthetic MLE environments from a small number of seed tasks, adopting a micro-scale approach that limits dataset size to 50–200 samples.

Experimental results show that SandMLE reduces execution time by more than 13x, enabling large-scale on-policy RL in the MLE domain. In experiments with Qwen3 models on MLE-bench-lite, the medal acquisition rate improved from 20.3% to 66.9% compared to the SFT baseline, and it also demonstrated strong generalization performance on MLE-Dojo.