Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

The team released OpenMLE, an open system for studying recursive self-improvement in machine learning engineering.

Frontis-MA1, a 35B model, was trained as a meta-evolution agent using execution-grounded supervised fine-tuning and reinforcement learning.

The stack combines verifiable task environments, operator learning, and long-horizon search across Draft, Improve, Debug, and Crossover actions.

It improved results on MLE-Bench Lite and NatureBench Lite, with gains also transferring to held-out tasks.

The authors also released the model weights and full stack to support reproducible research.