‹ back to 2026-07-30

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

arXiv cs.CL LLM Agents Junlin Yang, Che Jiang, Yu Fu, Tianwei Luo, Can Ren, Weizhi Wang, Kaikai Zhao, Hongyi Liu, Yuxin Zuo, Yuru Wang, Yuchen Fan, Kai Tian, Zhenzhao Yuan, Xiaojian Lin, Li Sheng, Rushi Qiang, Guoli Jia, Xingtai Lv, Ermo Hua, Dianqiao Lei, Youbang Sun, Ning Ding, Bowen Zhou, Kaiyan Zhang 2026-07-30

TL;DR - Frontis-MA1 is a 35B machine-learning-engineering agent trained to iteratively draft, improve, debug, and combine programs using execution feedback. Its open OpenMLE stack advances reproducible research into AI systems that improve AI-development workflows.

OpenMLE integrates verifiable task environments, execution-grounded SFT/RL, and long-horizon evolutionary search.
On MLE-Bench Lite, OpenMLE-Evo increased Frontis-MA1’s Medal Average from 39.39% to 60.61%, reaching 71.21% with experience priors and asynchronous search.
Model training and the search framework transferred independently to held-out NatureBench Lite, improving Match-SOTA from 50% to 70% and 20% to 50%, respectively.
The authors released the model weights and complete OpenMLE stack.

view merged work →