MLExpert: A 35B Generalist Agentic Model for Autonomous Machine Learning Engineering

Forum topic · 小凯 · 2026-10-17

Summary

MLExpert is a 35B mixture-of-experts agentic model designed for autonomous machine learning engineering (MLE), presented in an arXiv paper (2610.22344) by researchers advancing AI for AI (AI4AI) toward recursive self-improvement. The authors propose a post-training pipeline to handle long-horizon, model-directed experimentation where feedback from each costly experiment reflects multiple earlier decisions. Key components include Kaggle-MLE Factory, a framework for building executable tasks and model-directed experiment interactions; hierarchical agentic SFT that transfers general tool-use and coding abilities into MLE competence via outcome-guided supervision; and Windowed Interaction RL (WIRL), which treats multi-turn decisions between consecutive experiment results as execution windows with shared outcome-based rewards. On MLE-Bench Lite, MLExpert raises the average medal rate of its base model from 37.9% to 72.7%, matching the performance of much larger frontier models such as Kimi-K3 and GLM-5.3. The work demonstrates that careful supervision selection and reward assignment over long interaction trajectories can enable mid-sized models to compete in autonomous AI development.

论文概要

Research Area: NLP / AI4AI Authors: Shangheng Du, Jinxin Shi, Shuaiyu Zhang, Yiming Wu, Junpeng Zhao, Shiyang Feng, Xin Li, Liang He, Xiangchao Yan, Bo Zhang Published: 2026-10-15 arXiv: 2610.22344

English Abstract (original)

Large language models (LLMs) are expanding from assisting scientific research and engineering to autonomously developing and improving AI systems, advancing AI for AI (AI4AI) as a pathway toward recursive self-improvement (RSI). Workflow-driven machine learning engineering (MLE) systems provide a practical path toward this pursuit through complex orchestration that manages long-horizon experimentation, while recent MLE systems adopt model-directed interaction, allowing models to autonomously organize experiments and adapt optimization strategies based on execution feedback. However, feedback from each experiment reflects several preceding decisions, and a long MLE run contains many such costly experiments. Training must therefore select useful supervision from long interaction trajectories.

Key Contributions

MLExpert: a 35B mixture-of-experts (MoE) agentic model trained for autonomous machine learning engineering via a dedicated post-training method.
Kaggle-MLE Factory: an interactive framework for constructing executable tasks and model-directed experiment environments.
Hierarchical agentic SFT: transfers general tool-use and coding capabilities into MLE competence through outcome-guided supervision.
Windowed Interaction RL (WIRL): treats multi-turn decision-making between consecutive experiment results as execution windows and assigns shared outcome-based rewards across each window.

Results

On MLE-Bench Lite, MLExpert improves the base model's average medal rate from 37.9% to 72.7%, achieving performance comparable to much larger frontier models such as Kimi-K3 and GLM-5.3.

---

*Auto-collected on 2026-10-17.*