MLE-World Predicts ML Experiment Outcomes to Gate Test-Time Scaling

MLE-World is presented as a learned world model that predicts outcomes of machine learning experiments. The supplied text is a figure caption describing a schematic in which each candidate program receives a predicted report before selection or combination. Predictions carry a confidence score that routes candidates either to skipping execution (high confidence) or to running real code (low confidence). Newly combined code receives a fresh prediction. The text notes that the full method additionally verifies new best predictions and uses outcome-specific thresholds.

Key takeaways

• Each candidate program receives a predicted experiment report before selection or combination.
• High-confidence predictions skip execution, while low-confidence candidates are run as real code.
• The returned program carries its MLE-World confidence into a gating step.
• Scores and the 0.80 threshold shown are described as illustrative, not as the full method's values.

Why it matters

If predicted experiment outcomes are reliable, skipping costly execution could make automated ML search cheaper. The approach ties prediction confidence directly to whether real runs happen.

Read with care

The supplied text is a figure caption and schematic description; reported performance, accuracy, and savings are not included. Source claims not independently checked.