I am a PhD student at University of Manchester, supervised by Prof. Sophia Ananiadou. I worked as a DL4Code and DL4CyberSecurity researcher at Tencent Keen Lab from 2020 to 2022, where I designed deep learning models and methodologies to support cybersecurity researchers in reverse engineering binary code. I graduated from Shanghai Jiao Tong University in 2020 with a Master’s Degree, supervised by Prof. Gongshen Liu, and in 2017 with a Bachelor’s Degree.

I believe mechanistic interpretability is crucial for achieving trustworthy AGI. Without understanding the reasoning behind a model’s decisions, pursuing AGI would be extremely dangerous. Therefore, my current research focuses on:

a) Identifying important neurons and understanding the neuron-level information flow in LLMs. I have investigated the mechanisms underlying factual knowledge, arithmetic, in-context learning in LLMs, and visual question answering in multimodal-LLMs.

b) Leveraging interpretability insights to identify structural weaknesses in LLMs, design safer and more effective models/methods, and enhance the performance of downstream tasks.

Please feel free to contact me at zepingyu@foxmail.com for discussions. I am also open to potential collaborations and career opportunities.

🔥 News

2024.12: I’ve compiled a paper list for those interested in sparse auto-encoder (SAE).
2024.11: New preprint: Understanding Multimodal LLMs: the Mechanistic Interpretability of Llava in Visual Question Answering. This work explores the mechanism of Llava in visual question answering. [CODE]
2024.09: Our work is accepted by EMNLP 2024 (main): Interpreting Arithmetic Mechanism in Large Language Models through Comparative Neuron Analysis. This work explores the neuron-level information flow of arithmetic mechanism in LLMs. [CODE]
2024.09: Our work is accepted by EMNLP 2024 (main): Neuron-Level Knowledge Attribution in Large Language Models. This work introduces how to identify important neurons in LLMs, and explores the neuron-level information flow of factual knowledge mechanism. [CODE]
2024.09: Our work is accepted by EMNLP 2024 (main): How do Large Language Models Learn In-Context? Query and Key Matrices of In-Context Heads are Two Towers for Metric Learning. This work explores the mechanism of in-context learning in LLMs. [CODE]
2024.04: I’ve compiled a paper list for those interested in exploring the mechanisms of LLMs.