Hi:wave:, I am Zhaoyi Li, a computer science PhD student under the collaboration programm of University of Science and Technology of China (School of Computer Science and Technology, adviced by Prof.Defu Lian) and City University of Hong Kong (Department of Computer Science, co-adviced by Prof.Ying Wei and Prof.Linqi Song). The most fortunate thing for my early PhD life is that I met up with my brilliant professors! Previously, I got my Bachelor degree in Computer Science and Technology from University of Science and Technology of China in 2022, under the supervision of Prof.Cheng Li.
My research interest includes machine learning and natural language processing. Specifically, I work on making human beings and neural network systems better understand each other (i.e., generalization and interpretability). I currently focus on investigating the compositionality of language models from the perspectives of generalization, reasoning pattern, implicit structure and so on. I would love to discuss on related topics should anyone is interested (always feel free to drop me an email: lizhaoyi777[AT]mail.ustc.edu.cn)!
🔥 News
- 2024.05:  🎉🎉 Two papers was accepted by ACL’2024 (1$\times$main conference and 1$\times$findings).
- 2023.05:  🎉🎉 One paper was accepted by ACL’2023 (main conference).
đź“ť Publications
($*$ indicates the co-first authorship)
Understanding and Patching Compositional Reasoning in LLMs (ACL’2024 Findings)
Zhaoyi Li, Gangwei Jiang, Hong Xie, Linqi Song, Defu Lian and Ying Wei
- We interpret the internal mechanism of compositional step-by-step reasoning on multi-hop factual knowledge inside large language models and propose a light-weight method based on model editing to patch their compositional reasoning errors.
Benchmarking and Improving Compositional Generalization of Multi-aspect Controllable Text Generation (ACL’2024)
Tianqi Zhong$^*$, Zhaoyi Li$^*$, Quan Wang, Linqi Song, Ying Wei, Defu Lian and Zhendong Mao
- We propose CompMCTG, systematically unveiling the shortcoming of current multi-aspect controllable text generation methods on compositional generalization. Besides, we propose MetaMCTG, a meta-learning based training framework to improve the compositional generalization performance of current methods.
Learning to Substitute Spans towards Improving Compositional Generalization (ACL’2023 Oral)
Zhaoyi Li, Ying Wei and Defu Lian
- We propose L2S2, a adversarial learning based text data augmentation scheme to improve the compositional generalization performance of neural sequence models (including LSTMs, Transformers and Pre-trained LMs).
🎖 Honors and Awards
- 2022.06 Honored as The Outstanding Bachelor Graduate of USTC.
- 2022.04 YuanQing (CEO of Lenovo Co.) Research Student Scholarship.
- 2019 ～ 2022 First prize, USTC Excellent Undergraduate Scholarship.
- 2017.10 First prize, National (Senior) High School Mathematical Competition of China
- 2014.5 First prize, National (Junior) High School Mathematical Competition of China
đź“– Educations
- 2023.09 - now, PhD student, Department of Computer Science, City University of Hong Kong.
- 2022.09 - now, PhD student, School of Computer Science and Technology, University of Science and Technology of China.
- 2018.09 - 2022.07, Undergraduate, School of Computer Science and Technology, University of Science and Technology of China.
đź’» Internships
- 2021.05 - 2021.09, Summer Research Internship Programm, Department of Computer Science, Hong Kong University.
🎖 Services
- Invited Reviewer: EMNLP’2023, NeurIPS’2024.
- Teaching Assistant: Data Structure@USTC’2021, Data Engineering@CityU’2023, Introduction to Computer Programming@CityU’2024.