LLMs Automate Neural Architecture Search for Cross-Lingual OCR

This paper introduces an innovative AutoML framework that leverages large language models (LLMs) like GPT-5, GPT-4o, and Claude Sonnet 4 to automate the entire process of neural architecture search. The system focuses on developing models for cross-lingual handwritten optical character recognition (OCR). Each LLM acts as an independent designer, iteratively generating, training, evaluating, and refining neural network architectures based on performance feedback. The framework was rigorously tested across Arabic, Persian, and English handwriting datasets through numerous experiments. It consistently produced highly accurate and computationally efficient models, achieving over 93% mean test accuracy and low inference latency, all without requiring manual architecture design, domain-specific preprocessing, or hyperparameter tuning. This demonstrates the potential of LLMs as powerful AutoML agents for complex machine learning tasks.