I am a master's student in Information and Communication Engineering at Nantong University, focusing on explainable AI (XAI) and large language model post-training. I am currently seeking Ph.D. opportunities.
AAAI 2026 Oral
Authors: Runyu Wang, Peng Ping, Zhengyu Guo, Xiaoye Zhang, Quan Shi, Liting Zhou, Tianbo Ji
[Paper] [Code]
LoKI is a low-damage fine-tuning approach grounded in mechanistic interpretability. It combines knowledge-vector attribution with layer-wise balancing to identify critical parameters, improving downstream performance while substantially mitigating catastrophic forgetting.
Catastrophic Forgetting · Parameter-Efficient Fine-Tuning · Knowledge Localization · Mechanistic Interpretability
EMNLP 2026 Main
Authors: Runyu Wang, Bo Liu, Xiaxin Zhang, Yu Han, Jiawei Cao, Xiaoye Zhang, Zhe Zhang, Yifan Yang, Peng Ping
[Paper] [Code]
RACE efficiently identifies LLM neurons that exhibit functional consistency within representative corpus domains. By analyzing model residuals, it provides a scalable and cost-effective tool for large-scale model auditing.
Knowledge Localization · Mechanistic Interpretability · Model Auditing · Residual Analysis