Hello, I'm

Zirui Li 李梓睿

I am a second-year Ph.D. student in Computer Science at Harbin Institute of Technology, Shenzhen, supervised by Prof. Xuefeng Bai and Prof. Kehai Chen. My current research interests include LLM reasoning, interpretability, and Latent Chain-of-Thought. Before starting my Ph.D., I was a Master's student at the University of Manchester, working with Yizhi Li and Siwei Wu in Prof. Chenghua Lin's lab, where I focused on Multimodal retrieval. I also completed my Bachelor's degree in Computer Science at the University of Manchester, supervised by Prof. Rizos Sakellariou, during that period of time, I was interested in Social Network Analysis.

Natural Language ProcessingLLM efficiency reasoningLatent Chain-of-ThoughtOn policy distillationLLM interpretability
LocationShenzhen, China
Portrait of Zirui Li at Park Güell, Barcelona

Research

Dynamics Within Latent Chain-of-Thought

Zirui Li, Xuefeng Bai, Kehai Chen, Yizhi Li, Jian Yang, Chenghua Lin, Min Zhang
ICML 2026Seoul, Korea

DocMMIR: A Framework for Document Multi-modal Information Retrieval

Zirui Li, Siwei Wu, Yizhi Li, Xingyu Wang, Yi Zhou, Chenghua Lin
EMNLP 2025 FindingsSuzhou, China

Education

Harbin Institute of Technology, Shenzhen

Sep 2025 - present

Ph.D. in Computer Science

Natural Language ProcessingMachine translationLLM reasoningLLM interpretability

University of Manchester

Sep 2021 - Jun 2025

M.Eng Computer Science (2:1)

Natural Language UnderstandingDatabase DesignSoftware EngineeringGame TheoryAlgorithmsPythonPyTorchJavaJavaScriptSQLFlaskDjangoLinux

Experience

TeleAI

Aug 2026 - present

Research Intern

  • Exploring MoPD methods for improving preference-aligned distillation in large language models.
  • Investigating stable long-horizon distillation for OPD to improve training reliability over extended reasoning trajectories.

University of Manchester

Jun 2024 - Jul 2025

Research Intern - Prof.Chenghua Lin's research group

  • Designed DocMMIR, a document multimodal retrieval benchmark that unifies Wikipedia-scale imagery with text.
  • Built PyTorch Lightning pipelines and fine-tuned CLIP, BLIP2, SIGLIP and more MLLMs for cross-modal alignment.
  • Preprocessed 300k+ image-text samples and crafted contrastive objectives for query-to-document mapping.
  • Accepted for EMNLP2025 findings as the first author

AutoBizLine Inc.

Aug 2023 - Sep 2024

Full-stack Engineer - MySecondLine

  • Shipped 8k+ lines across a Django + JavaScript product, modernising messaging flows and dashboards.
  • Automated regulatory review pipelines with compliance tooling tailored to business rules.
  • Implemented geo-aware tax computation supporting 200+ scenarios to streamline onboarding.