Yuxiao Li
Bio
Yuxiao Li is an AI safety and mechanistic interpretability researcher currently based in Bilbao, Spain, where she is a postdoctoral researcher at the Basque Center for Applied Mathematics (BCAM). She holds a PhD in Computer and Information Sciences from Tsinghua University (2018-2023). Her research focuses on understanding the internal representations of large language models through techniques such as sparse autoencoders, variational inference, and geometric analysis of feature spaces. She was previously affiliated with MIT's Tegmark group and the Beneficial AI Foundation, where she was first author on "The Geometry of Concepts: Sparse Autoencoder Feature Structure" (arXiv 2024), a study of how concepts are geometrically organized in LLM activations. She has also participated in the ML Alignment & Theory Scholars (MATS) program and the Supervised Program for Alignment Research (SPAR), contributing multi-part research on structured priors and block-diagonal geometry in language model activations. She currently serves as a mentor in the Algoverse AI Safety Fellowship and has received independent research funding for inference-based AI interpretability work.
Grants
from Long-Term Future Fund
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.
Details
- Last Updated
- Mar 23, 2026, 2:06 AM UTC
- Created
- Mar 20, 2026, 3:00 AM UTC