MLExpert points toward a generalist agentic model for autonomous ML engineering.
MLExpert, a 35-billion Mixture-of-Experts agentic model, uses execution feedback to lift MLE-Bench Lite medal rate to 72.7%.
Key Takeaways
- MLExpert is a 35B-parameter Mixture-of-Experts agentic model designed to autonomously plan, execute, and revise machine-learning experiments.
- On MLE-Bench Lite, it raises the base model’s medal rate from 37.9% to 72.7%, showing performance comparable to larger frontier models.
- Instead of fixed workflows, it learns from execution feedback through Kaggle-MLE Factory tasks, hierarchical agentic supervised fine-tuning, and Windowed Interaction Reinforcement Learning.
- Its gains across scientific tasks, agent scaffold optimization, and algorithm optimization suggest the learned feedback-driven optimization ability may extend beyond ML engineering.
Meta: MLExpert is a 35 billion parameter Mixture-of-Experts agentic model that raises MLE-Bench Lite medal rate from 37.9% to 72.7% using execution feedback.
Tags: MLExpert, 35 billion, Mixture-of-Experts, MLE-Bench Lite, agentic models
Featured image: Abstract, friendly illustration of a Mixture-of-Experts model planning machine-learning experiments, with no real people, no logos, and no text.
MLExpert, a Mixture-of-Experts agentic model with 35 billion parameters, is designed to run machine-learning experiments on its own. It can plan work, use tools, carry out steps, and revise them based on results. On MLE-Bench Lite, MLExpert improves the base model's average medal rate from 37.9% to 72.7%, showing performance comparable to larger frontier models.
Learning from runs, not fixed workflows
Instead of depending on manually designed workflows, MLExpert uses execution feedback to guide optimization. Its training draws on Kaggle-MLE Factory to build executable tasks, uses hierarchical agentic supervised fine-tuning, and adds Windowed Interaction Reinforcement Learning. That method lets intermediate experiment results teach the model before a full run ends.
What the result suggests
The MLE-Bench Lite score shows a large jump over the base model. MLExpert also gained on scientific tasks, agent scaffold optimization, and algorithm optimization. Those findings suggest its learned optimization abilities may reach beyond machine-learning engineering.
Why readers should notice
For readers tracking AI experiment automation, the signal is the training method: the model learns to improve by reacting to its own intermediate outputs rather than following a preset sequence. The reported gains are tied to the named benchmark and tasks, but they show how feedback-driven training can narrow the gap with larger frontier models.

Comments (0)