← Back to the feed
AI & softwareResearch findings

Teach a machine-learning agent from intermediate experiments

Original published 2026.10.16 02:56 KSTAdded to Blackdeck 2026.10.18 03:10 KST

What has changed?

Intermediate experiments become explicit learning units in a long-running optimization process.

Key points

  1. MLExpert is a 35-billion-parameter mixture-of-experts agent trained for iterative machine-learning work.
  2. WIRL groups decisions between experiment results into execution windows with shared outcome rewards.
  3. The authors report an MLE-Bench Lite medal-rate increase from 37.9% to 72.7% over the base model.

Connect it to your work

Engineering teams can examine how experiment feedback is credited to preceding decisions.

What remains unproven

The benchmark result needs replication under different data and operating-cost constraints.

Originals used for this note

Links open the original publisher in a new tab.

  1. arXiv · authors' researchMLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering