MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering
35B MLExpert lifts MLE-Bench Lite medal rate from 37.9% to 72.7%, near larger frontier models.
The paper presents a post-training method for long-horizon machine-learning engineering and uses it to train MLExpert, a 35B mixture-of-experts agent. Kaggle-MLE Factory supplies executable tasks, hierarchical agentic supervised fine-tuning transfers tool use, and Windowed Interaction RL assigns shared outcome rewards to multi-turn experiment windows. On MLE-Bench Lite, MLExpert raises the base model's average medal rate from 37.9% to 72.7%, comparable to larger models such as Kimi-K3 and GLM-5.3. Further tests show gains on NatureBench Lite, agent scaffold optimization, and algorithm optimization.
52