Title: iML: A Code-Driven AutoML Framework That Is Executable, Problem-Specific, and Capable of Wide-Ranging Exploration

Abstract

While Automated Machine Learning (AutoML) has significantly democratized access to machine learning, prevailing methods often struggle with limited flexibility, a lack of transparency, and inconsistent execution reliability. Code-driven AutoML presents a compelling alternative by generating executable code for tasks such as preprocessing, model training, and evaluation. Nevertheless, current approaches based on Large Language Models (LLMs) frequently produce code that appears plausible linguistically but fails during execution, lacks sufficient grounding in the specific dataset, or is confined to narrow solution spaces.

To address these challenges, this paper introduces iML, a multi-agent, code-driven AutoML framework built upon three core principles: executability, problem grounding, and the broad exploration of viable solutions. The iML process begins by analyzing the task and profiling the data to create a structured blueprint. This blueprint directs modular code generation across various implementation tracks, encompassing traditional machine learning, pretrained model adaptation, and custom neural architectures. To ensure robustness, iML incorporates interface verification, dynamic execution, and iterative debugging during the integration phase.

We evaluated iML using MLE-BENCH and a newly proposed benchmark, iML-BENCH, which features diverse Kaggle-style tasks. On MLE-BENCH, iML achieved a 90% valid submission rate and a 45% medal rate, securing an Average Standardized Performance Score (APS) of 0.82. This represents a 52% to 273% improvement in APS over LLM-based baselines. Furthermore, on iML-BENCH, the framework attained the highest APS and maintained robust performance even when task descriptions were significantly simplified. These findings position iML as a reliable and highly competitive solution for code-driven AutoML.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC