UICoder: Enhancing Large Language Models to Automatically Generate User Interface Code through Feedback-Based Fine-Tuning

Large language models (LLMs) often struggle to produce reliable UI code that compiles and generates visually appealing designs. Traditional enhancement methods require costly human feedback or a distillation of proprietary models. This paper presents an innovative approach utilizing automated feedback, specifically compilers and multi-modal models, to improve LLM-generated UI code quality. Our methodology involves starting with a base LLM and iteratively enhancing model capabilities by creating a synthetic dataset from the original model. This dataset is then refined using automated tools to filter, score, and eliminate duplicates, resulting in higher quality data for finetuning the original LLM. We tested this approach on various open-source LLMs, evaluating performance against baseline models using both automated metrics and human assessments. The results indicate our refined models significantly outperformed all other available baselines, approaching the performance level of larger proprietary models. This study represents crucial advancements in LLM technology for UI code generation.

Source link