Yihao Quan（全奕豪）

Yihao Quan (全奕豪)

I'm a fourth-year ungraduate student at Beijing Jiaotong University, where I major in Management Information system.

I will be pursuing my Ph.D. in Computer Science at Rutgers University, starting in Fall 2025, under the supervision of Prof. Ryan Tang.

Research Interests

My research primarily focuses on MultiModal Learning and Trustworthy & Interpretability AI, with a emphasis on the following areas:

Vision + Language: Bridging the gap between our knowledge of language and vision representation, providing a foundation for more interpretable and controllable multimodal systems.

Unify Multimodal Understanding and Generation: Bridging architectural differences between autoregressive and diffusion models via mutual promotion of understanding and generation.

Mechanistic Interpretability: Understanding the internal mechanisms of LLMs and MLLMs through techniques like sparse autoencoder, circuit analysis, causal tracing, and neuron analysis.
Hallucination, Factuality and Safety: Using the interpretability findings to help downstream tasks (e.g. factual knowledge, enhancing reasoning, reducing hallucinations, model editing), and design safer models.
Representative papers are highlighted, * = Equal Contribution

Mar 8th 2025: Our paper "From Redundancy to Relevance: Information Flow in LVLMs Across Reasoning Tasks" got selected as NAACL 2025 Oral!!!
Jan 23rd 2025: Our paper "From Redundancy to Relevance: Information Flow in LVLMs Across Reasoning Tasks" got accepted by NAACL 2025 Main track.
Dec 10th 2024: Our paper "Enhancing LVLMs’ Complex Reasoning via Similarity Computation" got accepted by AAAI 2025.

	Seeing Clearly by Layer Two: Enhancing Attention Heads to Alleviate Hallucination in LVLMs Xiaofeng Zhang, Yihao Quan, Chaochen Gu, Chen Shen, Xiaosong Yuan, Shaotian Yan, Jieping Ye arXiv Under Review A plug-and-play and training-free method showing significant hallucination-mitigating performance on different VLMs and metrics.
	From Redundancy to Relevance: Information Flow in LVLMs Across Reasoning Tasks Xiaofeng Zhang, Yihao Quan, Chaochen Gu, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Hao Tang, Jieping Ye Github / arXiv Accepted by NAACL'25 Main (Oral) A novel perspective to enhance understanding of LVLMs and their functioning, particularly for complex reasoning tasks.
	Enhancing LVLMs’ Complex Reasoning via Similarity Computation Fanshuo Zeng, Xiaofeng Zhang, Yihao Quan, Zheng Hui, Jiawei Yao Github / arXiv Accepted by AAAI'25 A novel image token reduction method, Simignore, designed to enhance the complex reasoning capabilities.

Selected Awards and Honors

2022-2024: Second Class Scholarship

2022-2024: Merit Student of Beijing Jiaotong University

2022-2023: Dean’s List of Rochester Institute of Technology

2022: Second Prize of National College Student Mathematical Modeling Competition

2022: Kaggle Sliver Medal (Top 5%): Feedback Prize - Evaluating Student Writing

Feel free to steal this website's source code.