About me
This is the webpage for Zongjie’s academic research. I’m a final year Ph.D. student at the HKUST, advised by Prof. Shuai Wang.
Currently, I’m visiting the Advanced Software Technologies (AST) Lab at ETH Zurich (ETH), advised by Prof. Zhendong Su.
News
- [2024-10] Glad to be invited as a PC member at Conference on AI Foundation Models and Software Engineering (FORGE 2025) in ICSE 2025.
- [2024-09] Our paper
Split and Merge: Aligning Position Biases in LLM-based Evaluators
is now accepted by EMNLP 2024 Main Track - [2024-08] Our proposal towards Robustness & Fairness in LLM is now accepted by
OpenAI
. - [2024-06] Our LLM SFT data synthesis paper
API-guided Dataset Synthesis to Finetune Large Code Models
as well as the code is released. - [2024-05] The work of the first student I co-supervised,
NAVRepair: Node-type Aware C/C++ Code Vulnerability Repair
, is now accessible. - [2024-04] Our paper study the multi-lingual bias in large code models is now accessible.
Paper
- [2023-12] Our Visual Reference prompting benchmark
VisualReferPrompt
as well as the framework is released, paper. - [2023-11] Our paper
Evaluating C/C++ Vulnerability Detectability of Query-Based Static Application Security Testing Tools
is now accepted by TDSC - [2023-10] Our paper
On the Feasibility of Specialized Ability Stealing for Large Language Code Models
is now accepted by ICSE’24 (second round) - [2023-10] Make GPT-4 more consistent when using as evaluators, or even replace it with less powerful model is now applicable paper.
- [2023-09] Our paper about Watermark technique for LLM is now accepted by CCS’23, it is planning to be commercialized.
- [2023-08] Enhance binary similarity with low-cost? See our paper in ICSE’24 (first round).
Current Questions
- Why LLM works? How they understand the knowledge?
Is it feasible to use large-scale synthetic data to promote LLM helpfulness and security in specific scenarios?[Solved]Are LLMs good evaluators? To what extent will they impact existing (safety) evaluation systems?[Solved]
Long Vision
Ensuring traceable safety on the path towards AGI