Yan Li(李岩)

Previously, I obtained master degree in Machine Learning at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), advised by Prof. Kun Zhang. I work closely with Guangyi Chen and Yunlong Deng in the MBZUAI Causality Group.

From July 2025 to December 2025, I was a visiting student in the Department of Philosophy at Carnegie Mellon University (CMU), working with Prof. Peter Spirtes and Prof. Kun Zhang, and collaborating with the CMU CLeaR Group. I obtained a B.Eng. in Computer Science and Technology from Tsinghua University in June 2024, where I worked at the Research Institute of Human–Computer Interaction and Media Integration with Chun Yu, mainly on HCI and context-aware systems.

My research primarily focuses on representation learning, particularly representation learning for foundation models. My broader goal is to improve the reliability and robustness of large language models, for example by reducing hallucinations and enabling stronger generalization.

I have published extensively on causal representation learning, with much of this work examining conventional adaptation problems through a causal perspective. My collaborators and I have also constructed relevant datasets and authored position papers that aim to clarify important research directions in this area.

Since 2026, my research has increasingly shifted toward improving the capabilities of LLMs through the lens of representation learning. I developed MPM, a lossless approach for extending multimodal large language models with additional modalities. I am also collaborating with Prof. Sean Du on noise detection in LLMs, particularly in the reasoning trajectories produced by agents.

Another major direction of my current work is improving the long-horizon capabilities of agents. Instead of treating context as a single flat sequence of tokens, we organize it into structured blocks. During generation, the model can emit dedicated editing tokens from its vocabulary to perform atomic operations on specific blocks, including deletion, rewriting, and exact restoration by block identifier.