MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering
MLExpert raised its base model’s average medal rate on MLE-Bench Lite from 37.9% to 72.7%, the authors report.
The 35-billion-parameter mixture-of-experts model learns to organize machine learning experiments and adapt using execution feedback. Its training combines outcome-guided supervision with reinforcement learning that rewards the decisions made between successive experiment results. The authors report performance comparable to larger models including Kimi-K3 and GLM-5.3 on that benchmark. Additional experiments showed gains on NatureBench Lite, agent scaffold optimization and algorithm optimization, suggesting some transfer beyond the training tasks.
HF Daily Papers' note
The 35-billion-parameter mixture-of-experts model learns to organize machine learning experiments and adapt using execution feedback. Its training combines outcome-guided supervision with reinforcement learning that rewards the decisions made between successive experiment results. The authors report performance comparable to larger models including Kimi-K3 and GLM-5.3 on that benchmark. Additional experiments showed gains on NatureBench Lite, agent scaffold optimization and algorithm optimization, suggesting some transfer beyond the training tasks.
HF Daily Papers' note
score 5