Xiaoyu Xu (许潇予)

I am a final-year Software Engineering undergraduate at Fuzhou University (GPA rank: 1/89).

I study the internal mechanisms of language and vision-language models, focusing on failures in multi-step reasoning and multimodal hallucinations. My goal is to use these findings to improve model reliability. I previously interned at Shanghai AI Lab.

海到无边天作岸，山登绝顶我为峰。

News

• 2026.04: First-author paper accepted — ACL 2026 Main Conference.
• 2026.01: First-author paper accepted — IEEE Transactions on Consumer Electronics.
• 2025.10: Received the National Scholarship (4 of 1,200 students).

Publications

Reasoning Fails Where Step Flow Breaks

Xiaoyu Xu, Yulan Pan, Xiaosong Yuan, Zhihong Shen, Minghao Su, Yuanhao Su, Xiaofeng Zhang

ACL 2026 Main Conference (CCF-A)

Step-level saliency reveals recurring information-flow failures in long reasoning traces and guides a training-free repair.

Paper arXiv Code

Manifold-guided Geometric Enhancement for Reliable Multimodal Interaction in Consumer Electronics

Xiaoyu Xu, Yulan Pan, Xiaofeng Zhang, Xuhang Chen, Kim-Fung Tsang

IEEE Transactions on Consumer Electronics (JCR Q1)

Studies how multimodal representations change across layers and uses their geometry to reduce hallucinations.

Paper

Education

• 2023.08 - 2027.06 (Expected), B.E. in Software Engineering, Fuzhou University (Maynooth International Engineering College).

Experience

• 2025.07 - 2025.10, Research Intern, Shanghai AI Lab — Trustworthy AI Center.