1. Large Language Model (LLM) research and application. Proposed LLM-ensemble, which is the first model to ensemble multiple LLMs and achieves state-of-the-art results on both offline evaluation and A/B tests. Tech report accepted by SIGIR'24. 2. Diffusion models for personalized product image generation. Train a language model to generate prompts for stable diffusion to better understand abstract concepts. Fine-tune stable diffusion to achieve the best generated images that align the prompts. Tech report accepted by WWW'24.3. Train a multi-relational graph neural network for large-scale marketplace transaction data.