Massive language fashions (LLMs) wrestle to constantly generate UI code that compiles and produces visually related designs. Current approaches to enhance technology depend on costly human suggestions or distilling a proprietary mannequin. On this paper, we discover using automated suggestions (compilers and multi-modal fashions) to information LLMs to generate high-quality UI code. Our methodology begins with an present LLM and iteratively produces improved fashions by self-generating a big artificial dataset utilizing an unique mannequin, making use of automated instruments to aggressively filter, rating, and de-duplicate the info right into a refined greater high quality dataset. The unique LLM is improved by finetuning on this refined dataset. We utilized our strategy to a number of open-source LLMs and in contrast the ensuing efficiency to baseline fashions with each automated metrics and human preferences. Our analysis exhibits the ensuing fashions outperform all different downloadable baselines and strategy the efficiency of bigger proprietary fashions.
- ** Work accomplished whereas at Apple
- †Carnegie Mellon College
