A research framework is using language models to automatically discover and refine the data-cleaning steps that typically require the most human labor in a machine learning project.
Evolutionary Feature Engineering, or EFE, treats data preprocessing as a search problem. It generates candidate Python programs with a standard fit/transform interface, runs them against real datasets, and uses validation performance as feedback to evolve better versions — all without a human in the loop. The researchers tested two variants: EFE-Time, aimed at time-series forecasting, and EFE-Tab, which targets tabular prediction tasks. EFE-Time reduced forecasting errors by 3% or more on average across datasets and by as much as 19% on one COVID mortality dataset, working on top of existing time-series foundation models including Chronos-2.
The harder sell here is interpretability. Most automated feature engineering produces black-box transforms that accuracy-focused practitioners accept and everyone else ignores. EFE-Tab, by contrast, generates compact, readable feature programs — and the paper reports it was especially effective on classical decision trees, where a small set of evolved features matched or beat competing LLM-based methods while keeping the model explainable. That matters for any domain where a regulator or a skeptical colleague asks why the model did what it did.
Feature engineering automation has been a recurring promise in AutoML for a decade, with tools like Featuretools and various AutoML platforms making similar claims. What EFE adds is the generative flexibility of a language model as the mutation operator, replacing hand-coded transformation libraries with open-ended code synthesis — a bet that LLMs are better at creative search than at reliable reasoning.