Zichao Li

Bio

He is a PhD student working with Prof. Siva Reddy and Prof. Jackie Cheung. He has broad research interests in NLP, machine learning, and reinforcement learning. His work covers a diverse set of topics, including question answering, logical reasoning, probing, LLM editing, and evaluation. More recently, he has been working on understanding and modeling digital environments.

During his PhD, he completed a Mitacs internship at Borealis AI and is currently doing another Mitacs internship at ServiceNow, where he is working on knowledge organization and web understanding.

Contact

zichao (dot) li (at) mail (dot) mcgill (dot) ca

Research topics

The main theme of his research is understanding and enhancing the adaptation of LLM to align with the semantic structure of the external world, such as:

  • How the adaptation of LLMs/VLMs align with the transition dynamics of world (such as fictional world and digital world)? Do they capture the core abstractions to predict the future state?
  • How does knowledge updating and reasoning in LLMs align with the dependencies in (probabilistic) knowledge bases? Do they adhere to the principle of maximal knowledge consistency?
  • Do LLM representations induce a hierarachical organization of documents and knowledge? How to evaluate, probe/inspect and improve them?

Recent First-authored Publications

  • Zichao Li, Yanshuai Cao, Jackie Chi Kit Cheung. Do LLMs Build World Representations? Probing Through the Lens of State Abstraction. NeurIPS 2024. Link
  • Zichao Li, Ines Arous, Siva Reddy, Jackie Chi Kit Cheung. Evaluating Dependencies in Fact Editing for Language Models: Specificity and Implication Awareness. Findings of EMNLP 2023. Link
  • Zichao Li, Prakhar Sharma, Xing Han Lu, Jackie Chi Kit Cheung, Siva Reddy. Using Interactive Feedback to Improve the Accuracy and Explainability of Question Answering Systems Post-Deployment. Findings of ACL 2022. Link

Other Recent Publications

  • Shubham Gupta, Zichao Li, Tianyi Chen, Cem Subakan, Siva Reddy, Perouz Taslakian, Valentina Zantedesch. ReTreever: Tree-Based Coarse-To-Fine Representations for Retrieval. ArXiv preprint (2025). Link
  • Juan A. Rodriguez et al. BigDocs: An Open and Permissively-Licensed Dataset for Training Multimodal Models on Document and Code Tasks. ICLR 2025. Link