LACE AutoML evolves Python pipelines with GPT-5.4-mini, beating auto-sklearn and H2O
An evolutionary loop with an LLM as mutation operator generates editable scikit-learn pipelines.
Traditional AutoML systems search within a hand-specified space of preprocessors, learners, and hyperparameters—they can tune known components but never generate novel pipeline structure. LACE (Language-model Assisted Code Evolution), introduced by Sofoklis Kitharidis and colleagues, breaks this limit by evolving complete executable pipeline programs. The framework maintains a population of scikit-learn-compatible Python classes, with a large language model (GPT-5.4-mini) acting as the mutation operator. Because every candidate is ordinary code, the final pipeline and the search history can be read, edited, and reused directly—no opaque framework objects required.
Evaluated under a leakage-controlled protocol that hides dataset identity from the generator, LACE covered all 68 OpenML classification benchmarks. It significantly outperformed auto-sklearn, H2O, and a fixed XGBoost baseline, and showed no detectable difference against AutoGluon, the strongest search-based system tested. Tabular foundation models achieve higher accuracy on tasks they support, but apply a fixed pretrained predictor, whereas LACE returns a task-specific, editable program. Its contribution is a code-defined search space with full coverage, practical reusability, and a component set that grows simply by editing the prompt.
- First AutoML system to evolve full executable pipeline programs using an LLM as a variation operator
- On 68 OpenML tasks, LACE with GPT-5.4-mini beats auto-sklearn, H2O, and XGBoost; matches AutoGluon
- Outputs ordinary scikit-learn Python classes—editable and reusable—and extends components via prompt changes
Why It Matters
LACE makes AutoML transparent and editable, letting practitioners reuse and customize generated pipelines instead of relying on black-box models.