Lei Xu (徐 磊)
Ph.D. in Computer Science, MIT
Applied Scientist at Amazon AWS
Email:
leix328@gmail.com
[Publications]
About Me
I’m an Applied Scientist at Amazon AWS. My research covers a wide range of topics in natural language processing and machine learning. Recently, my work has focused on the application of large language models in the medical domain. My long-term objective is to pave the way for robust and deployable LLM applications. This involves addressing various challenges such as improving data quality, optimizing task-specific model performance, and developing reliable evaluation methods.
I received my Ph.D. in Computer Science from MIT in 2023, and my B.Eng. from Tsinghua University.
News
- Oct 2024: Check out our preprint CriSPO: Multi-Aspect Critique-Suggestion-guided Automatic Prompt Optimization for Text Generation for more effective automatic prompt engineering!
- Oct 2024: Our paper Salient Information Prompting to Steer Content in Prompt-based Abstractive Summarization is accepted to EMNLP 2024 Industry Track. [code]
- May 2024: Our position paper ConSiDERS-The-Human Evaluation Framework: Rethinking Human Evaluation for Generative Large Language Models is accepted to ACL 2024 main conference.
- Jul. 2023: AWS Announces HealthScribe Service.
- Feb. 2023: I joined Amazon AWS as an applied scientist.
Selected Projects
Robustness of Text Classifiers
With deployment of large language models, the security and robustness of these models have become a real issue. In this project, we thoroughly explore how well text classifiers can handle attacks from different angles. We developed R&R and SP-Attack for traditional transformer-based text classifiers. We also proposed AToP/BToP attacks for prompt-based language models. And we’re putting effort into defending against these attacks with In Situ Augmentation.
Synthetic Tabular Generation
In this project, we aim to generate synthetic tabular data, which can help remove the barriers in data sharing. We started with an LSTM-based TGAN model, then further improved it by proposing CTGAN.