Quzhe Huang
Email:huangquzhe [AT] pku [DOT] edu [DOT] cn

I am a fifth-year Ph.D. student at Wangxuan Institute of Computer Technology, Peking University, advised by Prof. Yansong Feng and Prof. Dongyan Zhao .

I am expected to graduate in July and looking for postdoc or other job opportunity now!

| CV | Google Scholar | Github |

  Research Interests

My research interest lies in designing efficient LLM and adapting LLM to different modality and domains, with the goal of building practical AI systems:

  • Efficient LLM: build small but competitive long-context LLM from scratch(Tech Report) and reduce the cost of both training and inference by dynamically allocating resources(ACL 2024).
  • Domain-Specific LLM: adapt LLM to specific domain by integrating domain knowledge(Tech Report) and evaluate whether models reason as domain experts(EMNLP 2022).
  • Multimodal LLM: design unified image-video-language pre-training framework(ICLR 2024, ICML 2024) and explore the ability of encoding global information(COLING 2024).

I am also interested in long context reasoning(ICLR 2024), document-level information extraction(ACL 2021_1 , ACL 2021_2, ACL 2022, ACL 2023, EMNLP 2023), and low-resource languages ( ACL 2024)

  • [June. 2024] I am expected to graduate in July and looking for postdoc or other job opportunity now!
  • [May. 2024] Our two papers, which are about efficient MoE and minority languages in China, are accepted in ACL 2024!
  • [May. 2024] Our work about unified video-language pre-training, Video-LaVIT, is accepted in ICML 2024!
  • [Apr. 2024] Our work about explore the multimodal LLM's ability of encoding global information is accepted in COLING 2024!
  Publications (Selected)

Harder Tasks Need More Experts: Dynamic Routing in MoE Models (ACL 2024)
Quzhe Huang*, Zhenwei An*, Nan Zhuang*, Mingxu Tao, Chen Zhang, Yang Jin, Kun Xu, Liwei Chen, Songfang Huang, Yansong Feng

TL:DR: We present a dynamic routing mechanism for Mixture of Experts(MoE) models that improves efficiency and performance by adjusting the number of activated experts based on input difficulty.
| paper | code |

Lawyer LLaMA: Enhancing LLMs with Legal Knowledge (Arxiv)
Quzhe Huang, Mingxu Tao, Zhenwei An, Chen Zhang, Cong Jiang, Zhibin Chen, Zirui Wu and Yansong Feng

TL:DR: We have designed a framework to adapt a general LLM into a specific domain and built a legal model, Lawyer LLaMA, based on the framework.
| paper | code & data |

Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization (ICML 2024)
Yang Jin, Zhicheng Sun, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, Kun Gai, Yadong Mu

TL:DR: Our proposed framework efficiently decomposes videos into keyframes and temporal motions, enabling unified generative pre-training for videos, images, and text.
| paper | code | website |

LaVIT: Unified Language-Vision Pretraining with Dynamic Discrete Visual Tokenization (ICLR 2024)
Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu, Chengru Song, Xiaoqiang Lei, Di Zhang, Wenwu Ou, Kun Gai, Yadong Mu

TL:DR: Our method introduces a visual tokenizer that translates images into discrete tokens, allowing LaVIT to handle and generate multi-modal content seamlessly within a unified generative learning paradigm.
| paper | code |

Probing Multimodal Large Language Models for Global and Local Semantic Representations (COLING 2024)
Mingxu Tao, Quzhe Huang, Kun Xu, Liwei Chen, Yansong Feng, Dongyan Zhao

TL:DR: The topmost layers of multimodal LLMs may excessively focus on local information, leading to a diminished ability to encode global information.
| paper | code |

More than Classification: A Unified Framework for Event Temporal Relation Extraction (ACL 23)
Quzhe Huang, Yutong Hu, Shengqi Zhu, Yansong Feng, Chang Liu, Dongyan Zhao

TL:DR: Determining the temporal relation between events based on their start and end time points, instead of treating this problem as a simple classification task.
| paper |

Do Charge Prediction Models Learn Legal Theory? (Findings of EMNLP 22)
Zhenwei An*, Quzhe Huang*, Cong Jiang, Yansong Feng, Dongyan Zhao

TL:DR: An empirical study of whether charge prediction models make judgments corresponds to the decision logic of human judges.
| paper |

Does Recommend-Revise Produce Reliable Annotations? An Analysis on Missing Instances in DocRED (ACL 22)
Quzhe Huang, Shibo Hao, Yuan Ye, Shengqi Zhu, Yansong Feng, Dongyan Zhao

TL:DR: An analysis of recommend-revise annotation scheme, where existing models or knowledge bases are used to recommend candidate instances and annotators revise the recommendations. We find that in the revision stage, annotators cannot supplement adequate missing instances, which might cause bias in the constructed dataset and the models trained on the new data
| paper | code & data |

Three Sentences Are All You Need: Local Path Enhanced Document Relation Extraction (ACL 21)
Quzhe Huang, Shengqi Zhu, Yansong Feng, Yuan Ye, Yuxuan Lai, Dongyan Zhao

TL:DR: Empirically showed that determining the relation between entities in long texts only requires limited evidence, and proposed a method based on co-reference and multi-hop reasoning to select evidence sentences for document-level RE.
| paper | code |

Exploring Distantly-Labeled Rationales in Neural Network Models (ACL 21)
Quzhe Huang, Shengqi Zhu, Yansong Feng, Dongyan Zhao

TL:DR: Proposed a method to prevent the model from ignoring the potential important non-rationale words and not distinguishing the importance of different rationale words, when incorporating human rational.
| paper |

  • ACL Rollings, area chair, 2023.10
  • ACL Rollings, reviewer, since 2021.9
  • ACL, reviewer, 2022,2023
  • EMNLP, reviewer, 2022,2023
  • COLING, reviewer, 2022,2023
  • EACL, reviewer, 2023
  • AAAI, reviewer, 2023
  • President Scholarship, Peking University, 2022
  • Uniqlo Scholarship, Peking University, 2017
  • Panasonic Scholarship, Peking University, 2016

Wangxuan Institute of Computer Technology, Peking University
No. 128 Zhongguancun North Street,
Haidian District, Beijing, 100871

Website design: