AI / ML evaluation
MLE-STAR Ablation
A controlled study of which components in Gemini-generated ML pipelines actually improve model quality.
About the project
A controlled study of which components in Gemini-generated ML pipelines actually improve model quality.
Evidence & results
- Generated candidate pipelines with the Gemini API and executed them in a controlled experiment runner.
- Disabled feature engineering, scaling, tuning, ensembling, and model-selection components to isolate their contribution.
- Reported N=20 runs with 95% confidence intervals; the diabetes study found no full-pipeline improvement in 10/10 comparisons.