Ziqi Wang (王子奇)

I am a Ph.D. student at the University of Illinois Urbana-Champaign, advised by Prof. Heng Ji and Prof. Tong Zhang, working on AI reasoning.

I was an intern at Meta GenAI in 2024 Summer, working with Rui Wang. I also spent two summers at Google working with Dr. Crick Wu and Dr. Le Hou. Prior to my Ph.D. study, I obtained a Bachelor's Degree in Computer Science at Tsinghua University, where I was fortunate to work with Prof. Zhiyuan Liu, Prof. Xiaolin Hu, Prof. Minlie Huang, and Prof. Xiang Ren at the University of Southern California.

Email  /  Résumé  /  Google Scholar  /  Twitter  /  Linkedin

profile photo

Research

My research goal is to build AI with strong reasoning capabilities to help human solve real-world problems reliably, and eventually boost science discovery process. My current milestone is to investigate:

(1) the intrinsic bottleneck of AI reasoning, from both model wise and algorithm wise. [Preprint 2024]

(2) Improving AI reasoning through post-training, especially from the aspect of reinforcement learning. [ICLR 2024]

Previously, I worked on information extraction during my undergraduate study.

Selected Publications (See full list on

Google Scholar

)

* denotes equal contribution

Eliminating Position Bias of Language Models: A Mechanistic Approach
Ziqi Wang, Hanlin Zhang, Xiner Li, Kuan-Hao Huang, Chi Han, Shuiwang Ji, Sham M. Kakade, Hao Peng, Heng Ji
Preprint, 2024  
Paper / Twitter

We propose a method to eliminate the position bias in LMs, which help LMs to better conduct reasoning.

Iterative Preference Learning from Human Feedback: Bridging Theory and Practice for RLHF under KL-Constraint
Wei Xiong*, Hanze Dong*, Chenlu Ye*, Ziqi Wang, Han Zhong, Heng Ji, Nan Jiang, Tong Zhang
ICML, 2024; ICLR ME-FoMo, 2024   (Oral Presentation)
Paper / Twitter

We propose Iterative DPO that significantly boosts the post-training performance of DPO.

Enabling Language Models to Implicitly Learn Self-Improvement
Ziqi Wang, Le Hou, Tianjian Lu, Yuexin Wu, Yunxuan Li, Hongkun Yu, Heng Ji
ICLR, 2024  
Paper / Slides / Twitter

A framework that learns self-improvement from preference data that aligns better human intention.

Augmentation with Projection: Towards an Effective and Efficient Data Augmentation Paradigm for Distillation
Ziqi Wang, Yuexin Wu, Frederick Liu, Daogao Liu, Le Hou, Hongkun Yu, Jing Li, Heng Ji
ICLR, 2023  
Paper / Slides / Talk Recording / Poster

We propose a simple data augmentation method that benefit knowledge distillation in a data-efficient regime.

Recognizing Object by Components With Human Prior Knowledge Enhances Adversarial Robustness of Deep Neural Networks
Xiao Li*, Ziqi Wang*, Bo Zhang, Fuchun Sun, Xiaolin Hu
TPAMI, 2023  
Paper / Code

We reveal that human prior knowledge can serve as an important component to improve vision robustness to attacks.

CLEVE: Contrastive Pre-training for Event Extraction
Ziqi Wang*, Xiaozhi Wang*, Xu Han, Yankai Lin, Lei Hou, Zhiyuan Liu, Peng Li, Juanzi Li, Jie Zhou
ACL, 2021  
Paper / Slides / Talk Recording / Code

We show that event-oriented contrastive pre-training can enhance event representations and understanding.

Learning from Explanations with Neural Execution Tree
Ziqi Wang*, Yujia Qin*, Wenxuan Zhou, Jun Yan, Qinyuan Ye, Leonardo Neves, Zhiyuan Liu, Xiang Ren
ICLR, 2020  
Paper / Project Website / Slides / Talk Recording / Code

We propose to use human explanations to build neuro-symbolic networks. The networks can then label pseudo labels to improve semi-supervised training.

NERO: A Neural Rule Grounding Framework for Label-Efficient Relation Extraction
Wenxuan Zhou, Hongtao Lin, Yuchen Lin, Ziqi Wang, Junyi Du, Leonardo Neves, Xiang Ren
WWW, 2020   (Best Paper Runner-Up, 2/1500+)
Paper / Code

We propose to use neural rules to ground data-efficient learning in relation extraction.

HMEAE: Hierarchical Modular Event Argument Extraction
Xiaozhi Wang*, Ziqi Wang*, Xu Han, Zhiyuan Liu, Juanzi Li, Peng Li, Maosong Sun, Jie Zhou, Xiang Ren
EMNLP-IJCNLP, 2019   (Oral Presentation)
Paper / Slides / Talk Recording / Code

We use human prior knowledge to make models aware of the semantic relations between labels and thus perform better in the event argument extraction task.

Education

University of Illinois Urbana-Champaign
Ph.D. in Computer Science, 2021-

Advisor: Prof. Heng Ji
Tsinghua University
B.E. in Computer Science, 2016-2021

Advisor: Prof. Zhiyuan Liu, Prof. Xiaolin Hu, and Prof. Minlie Huang

Service


Reviewer: ICLR, NeurIPS, ICML, ACL, EMNLP, NAACL, Pattern Recognition

Talks


Teaching LMs to Self-Improve by Reinforcement Learning. Cohere AI, 2024. [Slides][Video]
Enabling Language Models to Implicitly Learn Self-Improvement. Objective, Inc., 2023. [Slides]

Miscellanea


I am interested in Physics and Astronomy in my spare time (I was a Physics student before trained as a Computer Science student). I like Repairment and DIY since they help me understand the bottom mechanism. I am a big fan of John Carmack. I also learn a lot of life advice from Elon Musk.

The website is adapted from Jon Barron. Last update: Oct, 2024.