Synthetic Sandbox for Training Machine Learning Engineering Agents

SandMLE is a multi-agent framework that creates synthetic MLE environments from a few seed tasks to enable efficient on-policy RL.

By constraining datasets to 50–200 samples per task, SandMLE reduces execution time by over 13× compared with full-scale pipelines.

Using SandMLE, experiments show larger gains than supervised fine-tuning on MLE-bench-lite across Qwen3 series models.

The trained policy also generalized to new agent scaffolds, improving HumanRank on MLE-Dojo by up to 32.4%.