Hello! Husky Here! I’m a junior student at Shanghai Jiao Tong University(SJTU), majoring in Computer Science. I’m currently collaborating with Professor Jiaqi Ma at UIUC; before this, I had a wonderful research experience at Professor Quanshi Zhang’s XAI Lab.
My research interests include Instruction Following, RLHF, AI explainability, and interesting applications of LLMs, and I warmly welcome any invitations to further discussions!
I am now preparing my MS/PhD applications for Fall 2025, so if interested, please drop me an email!
📅 News
- I will be visiting UCSD as an intern supervised by Zhiting Hu this summer!
📝 Publications
Seeing is not always believing: The Space of Harmless Perturbations
Chen Lu, Shaofeng Li, Benhao Huang, Fan Yang, Yuan Luo
Discovered that we can add a substantial amount of noise to the original image with little impact on the model’s classification accuracy, even when this noise is significant enough that it prevents us from classifying the image correctly as humans.
Defining and extracting generalizable interaction primitives from DNNs
Chen Lu, Siyu Lou, Benhao Huang, Quanshi Zhang
Given different DNNs trained for the same task, we develop a method to extract their shared interactions. I was actively involved in the algorithm design and implementation process, and was responsible for the visualization work in the paper.
🔥 Undergraduate Projects
Chinese Slot Language Understanding
Projects of SJTU-CS3602 Natural Language Processing. We developed a BERT-based pipeline for Chinese slot understanding by incorporating Lexicon information into the BERT backbone and achieved an improved result (from 77.99% to 79.44% in accuracy) on the given noisy dataset.
Metric Transformer | TranBlendNet
Projects of SJTU-CS3324 Digital Image Processing. In the 'Metric Transformer' project, I employed the BLIP model as the backbone and devised a streamlined transformer structure to evaluate AI-generated images. The assessments encompassed Text-Image Image Quality, Authenticity, and Text-Image Correspondence. In the 'TranBlendNet' project, I delved into the impact of text input on saliency prediction tasks. I designed a model capable of dynamically adjusting its attention area within a given image based on different text inputs.
Adaptive Huffman Coding: Analysis and Applications
Information Theory course project: Exploring the application of Adaptive-Huffman. I was responsible for the theoretical exploration of adaptive Huffman coding and the attempt to apply adaptive Huffman coding to incremental learning in the CBOW model.
A website designed to display SJTU classrooms' data, such as seat number, CO2 concentration, temperature, etc. It aims to help student choose a proper self-study classroom. I was responsible for the front-end and back-end development of the website.
📇 Experience
Trustworthy and Reliable AI System Team
Remote Research Internship, 2023.11 - current
Advisor: Jiaqi Ma
Lab for Interpretability and Theory-Driven Deep Learning
Research undergraduate, 2023.04 - 2024.01
Advisor: Quanshi Zhang
🎖 Honors and Awards
- 2022, Shao Qiu Scholarship （10000¥，rank 5/129)
- 2023, Rui Yuan Hong Shan Scholarship (20000¥，rank 3/129)
📖 Educations
- 2021.09 - (now), Shanghai Jiao Tong University, Shanghai, China
- 2018.09 - 2021.07, Zhejiang Ruian High School, Zhejiang, China
🌇 Misc
- I’m a fan of anime, I’ve dabbled in a lot of good anime, such as Attack on Titan
Attack on Titan Wiki , Jujutsu Kaisen
Jujutsu Kaisen Wiki ; But my favourite one is called Pantheon
Pantheon (TV series) , which discussed Uploaded Intelligence(UI) and the relevant ethical problems. - As for sports, I like running and have comleted one-semester orienteering course of SJTU. Besides, I often play badminton and go to gym regularly with my friends.