MLExpert: Open 35B MoE Agentic Model Lifts MLE-Bench Lite Medal Rate from 37.9% to 72.7%, Matching Kimi-K3 and GLM-5.3

Background and the Problem

Autonomous Machine Learning Engineering (MLE)—where models independently design architectures, run experiments, and refine strategies based on validation feedback—is a cornerstone of AI-for-AI (AI4AI). However, each experimental outcome reflects multiple preceding coding and configuration turns, and long MLE trajectories make end-of-trajectory RL prohibitively slow and credit-ambiguous.

Architecture and How It Works

Researchers from InternScience / Shanghai AI Lab trained MLExpert, a 35B Mixture-of-Experts agentic model:

1. Kaggle-MLE Factory & Hierarchical Agentic SFT: Constructs executable ML engineering environments and transfers general coding and tool-use skills via outcome-guided supervision;
2. Windowed Interaction RL (WIRL): Groups the multi-turn decisions between two successive experiment evaluations into a discrete execution window, assigning shared outcome-based rewards and updating the policy asynchronously before the full trajectory terminates.

Benchmarks and Measured Results

Evaluations across MLE and scientific optimization benchmarks highlight remarkable efficiency:

• 72.7% Medal Rate on MLE-Bench Lite: Lifts the base model's average medal rate from 37.9% to 72.7%, matching much larger frontier models including Kimi-K3 and GLM-5.3;
• Generalizes to Agent Scaffold Optimization: Transfers learned experimental optimization capabilities to NatureBench Lite, autonomous agent scaffold optimization, and general algorithm tuning.

Getting Started for Developers

Access the framework and harness at https://github.com/InternScience/MLExpert:

• Apply WIRL to Iterative Coding Loops: Segment long debugging or hyperparameter tuning trajectories into compile-and-eval execution windows to enable dense intermediate RL updates.