MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering

Traditional machine learning engineering (MLE) workflows require manual, long-horizon experimentation, where feedback from successive trials is costly and difficult to attribute to specific preceding decisions. To address this, the paper introduces MLExpert, a 35B Mixture-of-Experts (MoE) agentic model designed for autonomous MLE. The framework utilizes the Kaggle-MLE Factory to generate executable tasks and an interaction harness for model-directed experimentation. It employs hierarchical agentic Supervised Fine-Tuning (SFT) to transfer tool-use and coding capabilities, alongside Windowed Interaction Reinforcement Learning (WIRL) to assign shared outcome-based rewards across multi-turn decision windows. While the paper reports autonomous optimization capabilities, a key limitation for student implementation is the high computational demand of a 35B parameter model.

To bridge this gap, Fried Engineers offers project development support and component guidance. This includes structuring smaller-scale agentic workflows, reviewing reinforcement learning reward designs, and simulating simplified MLE feedback loops using accessible open-source LLMs. This approach allows students to explore recursive self-improvement concepts within standard hardware constraints.

Message: Hi FE, I need help with “MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering” in “Artificial Intelligence & Machine Learning”