MLE-World: a learned world model predicts ML experiment outcomes, halving compute while raising scores
sherryyangML · x · 2026-10-23
Researchers from UW, NYU and others introduce MLE-World, a learned world model (fine-tuned Qwen3.5-4B) that predicts execution feedback — terminal output, exception class, validation metric — before ML research agents actually run experiments, with a confidence gate deciding when real execution is needed. On 17 MLE-bench Lite tasks it lifts mean leaderboard percentile from 45.4 to 54.1 (+19%) over baseline while cutting real execution time from 9.94 to 4.77 hours per trial. Paper, model, data and code are open.
More from coding & agent
- Developers Don't Code Anymore — Now Their Job Is Rotating Secrets Daily — Dan_Jeffries1 · 2026-10-23
- New paper: coding agents in a shared sandbox self-create rules, reciprocity debts, and coalitions — nandofioretto · 2026-10-23
- Developer runs AI project managers that manage their own AI agents to build his game — draginol · 2026-10-23
- RentAKaren: a real service that rents humans to handle complaints your AI agent can't win — mathemagic1an · 2026-10-23
- Clerk's Auth for MCP growing 13% week-over-week, fastest since 2023 — stuffyokodraws · 2026-10-23
- Shopify builds remote mobile dev environments for AI agents — rseroter · 2026-10-23