MLE-World model predicts ML experiment outcomes, halving compute and lifting rankings

Researchers from UW, NYU and others released MLE-World, a learned world model that predicts ML experiment outcomes from historical trajectories, letting agents skip unpromising runs. It cuts compute roughly in half and improves leaderboard percentile by 19%, complementing the team's Reward-Rate Policy Gradient work.

2026-10-23 ~ 2026-10-23 · 4 related posts

• MLE-World: a learned world model predicts ML experiment outcomes, halving compute while raising scores — sherryyangML · 2026-10-23
• MLE-World Learns From Past ML Experiment Traces, Gaining 19% on MLE-bench Lite — sherryyangML · 2026-10-23
• Reward-Rate Policy Gradient: RL that optimizes MLE agents' reward per unit time — sherryyangML · 2026-10-23
• MLE-World released: reusing experience to cut experiment feedback cost, complementing RPG — sherryyangML · 2026-10-23