AI researcher and practitioner. I aim to: 1) diagnose LLM limitations by carefully crafting evaluation datasets and procedures. 2) advance LLMs by developing new architectures, optimizing training and inference, and designing new training datasets.