MLE-World: a learned world model predicts ML experiment outcomes to cut agent costs
Brian Ko · hf · 2026-10-28
A paper from UW, UIUC, Oxford, Columbia, and NYU introduces MLE-World, a world model finetuned from Qwen3.5-4B that predicts an ML program's terminal output (error traces, training logs, validation scores) from its code and data description, without executing it, plus a confidence score. Since a prediction costs one inference instead of a full training run, it makes test-time scaling affordable for AI research agents; the authors build five TTS policies (sequential revision, best-of-N, pairwise judging, fusion, and a UCB-style rule) that only execute candidates when confidence is high.
More from coding & agent
- 41 leftover git worktrees eat 45GB: fix it with a cleanup rule in AGENTS.md — dotey · 2026-10-28
- One prompt fixes AI-generated duplicate code: ask it to dedupe and simplify the architecture — facontidavide · 2026-10-28
- 10 open-source repos that rewrite prompts and evolve agent workflows — victor_explore · 2026-10-28
- The Trinity: recursive self-improvement for multimodal agents via capability trees — Weidong Tang · 2026-10-28
- Lead Claude kills a subagent over a bad figure, cites its 475k-token context — Sauers_ · 2026-10-28
- Having agent teams build dashboards of their own progress is a surprisingly good workflow — francesc · 2026-10-28