Junfeng Fang

Timeline

1. 2023

   NeurIPS oral on GNN explanations

   They are first author of an oral paper at NeurIPS 2023, written with Wei Liu, Yuan Gao, Zemin Liu, An Zhang, Xiang Wang and Xiangnan He, on evaluating post-hoc explanations for graph neural networks via robustness analysis.

2. October 2024

   AlphaEdit and null-space editing

   In October 2024, they post AlphaEdit as first author. The method projects each edit's perturbation onto the null space of the preserved knowledge, which keeps the outputs of post-edited models on that knowledge unchanged. On LLaMA3, GPT2-XL and GPT-J, it raises the performance of most locating-then-editing methods by an average of 36.7%.

3. October 2024

   A study of safety attention heads

   In October 2024, they coauthor "On the Role of Attention Heads in Large Language Model Safety". The paper finds that ablating a single safety head lets Llama-2-7b-chat respond to 16 times more harmful queries while modifying 0.006% of the parameters.

4. February 2025

   AnyEdit for long-form knowledge

   In February 2025, they coauthor AnyEdit, an autoregressive editing method that splits long-form knowledge into sequential chunks and edits the key token in each chunk. It outperforms strong baselines by 21.5% on benchmarks including UnKEBench and AKEW.

5. 2025

   ICLR 2025 Outstanding Paper

   AlphaEdit wins an Outstanding Paper Award at ICLR 2025 as one of three Outstanding Paper winners. They are first author, with Houcheng Jiang, Kun Wang, Yunshan Ma, Jie Shi, Xiang Wang, Xiangnan He and Tat-Seng Chua as coauthors.

6. 2025

   Postdoc in the NExT Lab

   In 2025, they hold a first-year postdoctoral position at the National University of Singapore, in the NExT Lab supervised by Chua Tat-Seng. Their research covers trustworthy large language models, including safety alignment and knowledge editing.

7. As of October 2026

   Research fellow at NUS

   They are a research fellow on the post-doc team of the NExT research center, listed under LLM safety, and a postdoctoral researcher at the National University of Singapore. Their research covers trustworthy large language models, including LLM safety and LLM-based agent systems.