CV

PhD in Model Compression with Algebraic Approaches

Imperial College London

• January 2022 – May 2026; supervised by Prof. Danilo P. Mandić.
• LLM compression with algebraic approaches, including tensor decomposition and structural pruning.
• Thesis: Generative Language Model Compression with Algebraic Approaches.

MEng in Computer Science and Technology

University of Science and Technology of China

• September 2018 – July 2021; supervised by Prof. Xiang-Yang Li.
• AI Privacy and Federated Learning
• IoT Sensing
• Thesis: Set-level Membership Inference in Deep Learning Systems.

BEng in Computer Science and Technology

China University of Petroleum - Beijing

• September 2014 – July 2018; ranked first in the department.

Founder

Facthink LTD

• Since May 2026, building Graph-Hypoth, a multi-agent hypothesis generation system that develops research claims into candidate hypotheses and literature-grounded experiment plans.
• Seven specialised agents work over a shared claim graph to review evidence, propose and critique hypotheses, and design experiments, with fixed aggregation rules and versioned records for audit.

PhD Researcher

Imperial College London, Department of Electrical and Electronic Engineering

• Used representation dimension to estimate language model performance after structural pruning.
• Developed a unified taxonomy of matrix and tensor factorisation for generative language model compression.
• Compressed GPT-series models with tensor decomposition and deployed them on resource-constrained devices, including Raspberry Pi.
• Applied membership inference to dataset legitimacy assessment in MLaaS; this work was featured by the Montreal AI Ethics Institute.

Technical Product Manager (Intern)

Baidu PaddlePaddle

• October – December 2021.
• Co-authored a paper on the product design of PaddlePaddle, a deep learning platform for users with technical and non-technical backgrounds.

MEng Research, LINKE Lab

University of Science and Technology of China

• April 2018 – June 2021.
• Implemented subject property inference in federated learning with CycleGAN.
• Implemented lightweight LSTM models on smartwatches for identification and authentication from accelerometer and gyroscope gesture signals.

llm-profiler

• A fine-grained, layer-level LLM profiler that counts arithmetic operations, parameters, temporary buffers, and activations.

ori

• The code of the paper "Data Origin Inference in Machine Learning".

• First Prize Scholarship in University of Science and Technology of China (2019, 2020, 2021)

• Outstanding Graduate Student of Beijing

• Meritorious Winner in Mathematical Contest in Modeling, U.S. (top 10%)

• National Scholarship of China (ranked first in the department)