Lexin Zhou
I am a research intern at Microsoft, advised by Dr. Xing Xie and Prof. Jose Hernandez-Orallo. I did my master’s in NLP & HCI at the University of Cambridge, supervised by Prof. Andreas Vlachos. Prior to that, I did my BSc in Data Science at the Universitat Politècnica de València, where I got into research by working with Prof. Jose Hernandez-Orallo.
I am interested in research about AI Evaluation, social computing, human-AI interactions and AI safety, regularly taking inspiration from psychometrics and cognitive science. At present, I mostly spend my day thinking about (i) designing robust evaluation methods that offer explanatory and predictive power of AI’s capabilities and limitations, and (ii) assessing and anticipating societal risks associated with the deployment of AI in the quest of offering actionable insights that translate into policy and design changes to minimise the harms of AI while amplifying their benefits. I am especially intrigued by general-purpose systems like LLMs.
I’ve spent time in research/consultancy roles on AI Evaluation at Microsoft Research, Meta AI, OpenAI, Krueger AI Safety Lab, VRAIN, and European Commission JRC. My work has been featured in Nature, Financial Times, MIT Tech Review, Forbes, IEEE Spectrum, El País, New Scientists, QbitAI, IBM, among others.
If you are drawn to everything relevant to AI Evaluation and wanna stay informed, please subscribe our monthly AI Evaluation Digest newsletter! If you wanna talk about something I do, feel free to reach out via email or on Twitter.
news
| Oct 30, 2024 | 💡Invited talk on Larger and More Instructable Language Models Become Less Reliable at Microsoft Research! |
|---|---|
| Sep 25, 2024 | 📜 Larger and More Instructable Language Models Become Less Reliable is finally out in Nature! Takeaways on X and a fairly well-written article in Chinese by QbitAI. This reminds me of Goodhart’s law. |
| Sep 20, 2024 | 📜 An LLM Feature-based Framework for Dialogue Constructiveness Assessment is accepted by EMNLP 2024, receiving high review scores that placed it in the top 0.5% of all submissions! |
| Mar 01, 2024 | 👨💻 Participated in the Red Team at Meta AI for their new foundation models, focusing on adversarial testing. |
| Oct 09, 2023 | 📜 Predictable Artificial Intelligence preprint at Arxiv! - It synthesizes one of my key perspectives on the future of AI, which I personally feel really excited about. |
| Sep 09, 2022 | 👨💻 Participated in the Red Team of GPT-4 at OpenAI, focusing on capability assessment, reliability evaluation, and adversarial testing. |
selected publications
- Larger and More Instructable Language Models Become Less ReliableNature, 2024
- An LLM Feature-based Framework for Dialogue Constructiveness AssessmentEMNLP, 2024