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

Summary

OpenMLE is a full stack: OpenMLE-Gym (5,758 quality-gated executable tasks), OpenMLE-ERL (execution-grounded SFT+RL on the four operators), and OpenMLE-Evo (experience-guided search). Frontis-MA1-35B post-trained on this stack raises MLE-Bench Lite Medal Average from 39.39% (Qwen3.6-35B-A3B) to 60.61% under OpenMLE-Evo and 71.21% under OpenMLE-Evo-Max (12h / one RTX 4090 12GB). NatureBench Lite Match-SOTA moves 50%→70% by swapping the trained model and 20%→50% by swapping the harness. Weights, gym, and traces are released.

Keywords

frontis · openmle · rsi · ai4ai · mle-bench · naturebench · evolutionary-search · moe · qwen3.6

Topics

AI4AI, machine learning engineering, evolutionary search, RSI

Research notes

• Primary: arxiv abs (CC BY 4.0, cs.CL). Code https://github.com/FrontisAI/OpenRSI and project https://frontisai.github.io/OpenRSI/. Model https://huggingface.co/FrontisAI/Frontis-MA1-35B (CC BY-NC 4.0; upstream Qwen Apache-2.0 notice preserved). Datasets FrontisAI/OpenMLE-Tasks and FrontisAI/OpenMLE-SFT-Traces. Correspondence zhangkaiyan@frontis.cn. Discord posted AlphaXiv 2607.28568v1. Paper-plus-model; OpenMLE-Tasks is the associated gym release (not a separate papers_local row).