ML-Agent: A Reinforcement Learning Framework for LLM-Based Autonomous ML Engineering

A recent study presents ML-Agent, a framework designed to utilize online reinforcement learning for training large language model (LLM) agents in autonomous machine learning engineering. The researchers highlight the shortcomings of existing prompt-based methods: smaller models are unable to learn from execution paths, and larger proprietary models incur high computational costs. The framework consists of three key elements: exploration-enriched fine-tuning for generating varied actions, step-wise reinforcement learning for training on individual action steps, and a mechanism to speed up experience gathering. This research marks the inaugural investigation into learning-based agentic ML, where an LLM agent acquires knowledge through interactive experimentation with ML tasks.