Shanghai AI Lab's MLExpert 35B MoE Nearly Doubles MLE-Bench Lite Medal Rate to 72.7%

Shanghai AI Laboratory and East China Normal University present MLExpert, a 35B MoE post-trained for autonomous machine-learning engineering via hierarchical SFT plus a Windowed Interaction RL scheme that treats multi-turn stretches between experiment outcomes as a shared-reward window. The team reports MLE-Bench Lite medal rate rising from 37.9% to 72.7%, with the paper claiming frontier-comparable results that generalize beyond its training domain. Code and weights are promised on GitHub and Hugging Face under the InternScience org.