MLExpert
A 35B mixture-of-experts (MoE) agent model optimized through post-training methods for autonomous long-horizon machine learning engineering, capable of autonomously orchestrating experiment workflows, modifying code and adjusting optimization strategies based on execution feedback.
Timeline (last 90 days)
MLExpert uses hierarchical agent supervised fine-tuning to transfer programming capabilities to MLE tasks, and uses windowed interactive reinforcement learning to treat multi-turn decisions as execution windows with shared rewards.
MLExpert is a 35-billion-parameter mixture-of-experts (MoE) agent model that uses hierarchical agent supervised fine-tuning (SFT) combined with windowed interactive reinforcement learning.
The optimization capabilities learned by MLExpert can generalize to NatureBench Lite and algorithm optimization tasks.
WIRL (Windowed Interaction RL) treats the multiple rounds of decisions between two experiment results as one 'execution window' and assigns this window a shared, outcome-based reward.
The paper "MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering" proposes a post-training method for autonomous long-horizon MLE optimization, and trains a 35B mixture-of-experts (MoE) agent model, MLExpert.
The paper reports improvements of MLExpert on NatureBench Lite, agent scaffold optimization and algorithm optimization, indicating that the model learns a transferable optimization capability.
MLExpert raised the base model's average medal rate on MLE-Bench Lite from 37.9% to 72.7%.
MLExpert's performance on MLE-Bench Lite approaches that of larger-scale frontier models such as Kimi-K3 and GLM-5.3.
All Facts (8)
MLExpert is a 35-billion-parameter mixture-of-experts (MoE) agent model that uses hierarchical agent supervised fine-tuning (SFT) combined with windowed interactive reinforcement learning.
60%UnverifiedMLExpert uses hierarchical agent supervised fine-tuning to transfer programming capabilities to MLE tasks, and uses windowed interactive reinforcement learning to treat multi-turn decisions as execution windows with shared rewards.
50%UnverifiedThe optimization capabilities learned by MLExpert can generalize to NatureBench Lite and algorithm optimization tasks.
50%UnverifiedWIRL (Windowed Interaction RL) treats the multiple rounds of decisions between two experiment results as one 'execution window' and assigns this window a shared, outcome-based reward.
50%UnverifiedThe paper "MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering" proposes a post-training method for autonomous long-horizon MLE optimization, and trains a 35B mixture-of-experts (MoE) agent model, MLExpert.
50%UnverifiedThe paper reports improvements of MLExpert on NatureBench Lite, agent scaffold optimization and algorithm optimization, indicating that the model learns a transferable optimization capability.
50%UnverifiedMLExpert raised the base model's average medal rate on MLE-Bench Lite from 37.9% to 72.7%.
50%UnverifiedMLExpert's performance on MLE-Bench Lite approaches that of larger-scale frontier models such as Kimi-K3 and GLM-5.3.
50%