Jiayi Fu

Jiayi Fu

M.S. student in Fudan University

CV

Jiayi Fu

Location: Shanghai, China
Phone: +86 19370551152
Email: fujy22@m.fudan.edu.cn
GitHub: github.com/PorUna-byte
Homepage: https://poruna-byte.github.io/
Google Scholar: https://scholar.google.com/citations?hl=en&user=V2QlLjkAAAAJ

Education

Fudan University
Master’s Student in Natural Language Processing
Expected Graduation: June 2025

Harbin Institute of Technology
Bachelor of Science in Computer Science and Technology
Graduated: June 2022

Internship Experience

Currently, I have no formal internship experience. However, I have engaged extensively in academic projects and collaborative coding initiatives that have honed my skills and understanding in the field.

Projects

Travel Agent
Primary developer of an intelligent travel planner leveraging ChatGPT. Focused on addressing the challenges of creating travel plans that are reasonable, personalized, and up-to-date. Developed strategies to enhance the reasoning capabilities of base models and integrate user preferences with current global information.

RLHF: Reinforcement Learning From Human Feedback
I am engaged into a project that tries to reproduce the RLHF technique based on Llama2. In this project, my role is to train a reward model. Basically, the reward model is trained via a contrastive loss and a language modeling loss. The reward model is initialized as original Llama2-7B, and substitute the final layer to output a scalar as reward. The final reward for a prompt and a completion is given by the reward model and a KL-divergence between policy model and initial supervised model.

Acadamic Research

GumbelSoft: Diversified Language Model Watermarking via the GumbelMax-Trick
I am the First author of a paper aimed at embedding highly detectable watermarks while maintaining text quality and diversity into machine generated texts. Our paper is based on Aarason’s Exponential watermark and explores several ways to enhance the generation diversity of his exponential watermark.

P4: Plug-and-Play Discrete Prompting for Large Language Models Personalization
In this paper, we propose a plug-and-play prompting method to manipulate the LLMs’ personality traits. Specifically, we append discrete personalized suffixes, automatically generated through an aggregated gradient-based search method, to the user query or dialog histories and induce LLMs to respond with target personalities.

Skills

Extracurricular Activities

Fudan Football Team
Position: Left Forward, March 2024 – Present
Key striker, contributing significantly to team dynamics and success.