I am currently a researcher on the Pre-training Team at Tencent Hunyuan. Prior to this, I received both my B.S. and M.S. degrees from Beihang University, under the supervision of Prof. Qinghong Yang. I also interned at Microsoft STCA and the BAAI.
My research interests lie in the interpretability of large language models and efficient Transformer training. Outside of work, I enjoy coffee and have been hitting the gym for over five years — and I share my life with a cat named Caiyuan.
Warning
Problem: The current name of your GitHub Pages repository ("Solution: Please consider renaming the repository to "
http://".
However, if the current repository name is intended, you can ignore this message by removing "{% include widgets/debug_repo_name.html %}" in index.html.
Action required
Problem: The current root path of this site is "baseurl ("_config.yml.
Solution: Please set the
baseurl in _config.yml to "Education
-
Beihang UniversityM.S. in Software EngineeringSep. 2021 - Jun. 2024 -
Beihang UniversityB.S. in Software EngineeringSep. 2017 - Jun. 2021
Experience
-
Researcher2024 - present -
-
Research Intern · Advised by Zengxuan Wen and Zhiqiang Zuo2023
Honors & Awards
-
National Graduate Scholarship2023
-
Academic Excellence Scholarship2022
-
Outstanding Undergraduate Graduate Award2021
-
Outstanding Student Leader Award2020
-
National College Innovation and Entrepreneurship Competition Award2020
News
Selected Publications
AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities
Zhongzhi Chen, Guang Liu, Bo-Wen Zhang, Fulong Ye, Qinghong Yang, Ledell Wu
Findings of the Association for Computational Linguistics (ACL) 2023
We present a simple and effective method to train a strong bilingual/multilingual multimodal representation model by altering the text encoder in CLIP with XLM-R, achieving new state-of-the-art on ImageNet-CN, Flicker30k-CN, COCO-CN and XTD.
AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities
Zhongzhi Chen, Guang Liu, Bo-Wen Zhang, Fulong Ye, Qinghong Yang, Ledell Wu
Findings of the Association for Computational Linguistics (ACL) 2023
We present a simple and effective method to train a strong bilingual/multilingual multimodal representation model by altering the text encoder in CLIP with XLM-R, achieving new state-of-the-art on ImageNet-CN, Flicker30k-CN, COCO-CN and XTD.
Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuning
Zhongzhi Chen, Xingwu Sun, Xianfeng Jiao, Fengzong Lian, Zhanhui Kang, Di Wang, Cheng-Zhong Xu
AAAI Conference on Artificial Intelligence (AAAI) 2024
We introduce Truth Forest, a method that enhances truthfulness in LLMs by uncovering hidden truth representations using multi-dimensional orthogonal probes, improving Llama-2-7B truthfulness from 40.8% to 74.5% on TruthfulQA.
Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuning
Zhongzhi Chen, Xingwu Sun, Xianfeng Jiao, Fengzong Lian, Zhanhui Kang, Di Wang, Cheng-Zhong Xu
AAAI Conference on Artificial Intelligence (AAAI) 2024
We introduce Truth Forest, a method that enhances truthfulness in LLMs by uncovering hidden truth representations using multi-dimensional orthogonal probes, improving Llama-2-7B truthfulness from 40.8% to 74.5% on TruthfulQA.
Adaptive Activation Steering: A Tuning-Free LLM Truthfulness Improvement Method for Diverse Hallucinations Categories
Tianlong Wang, Xianfeng Jiao, Yinghao Zhu, Zhongzhi Chen, Yifan He, Xu Chu, Junyi Gao, Yasha Wang, Liantao Ma
The Web Conference (WWW) 2025
We introduce ACT (Adaptive Activation Steering), a tuning-free method that adaptively shifts LLM activations toward the "truthful" direction at inference time, addressing diverse hallucination categories and improving truthfulness significantly across LLaMA, Alpaca, Vicuna and other models.
Adaptive Activation Steering: A Tuning-Free LLM Truthfulness Improvement Method for Diverse Hallucinations Categories
Tianlong Wang, Xianfeng Jiao, Yinghao Zhu, Zhongzhi Chen, Yifan He, Xu Chu, Junyi Gao, Yasha Wang, Liantao Ma
The Web Conference (WWW) 2025
We introduce ACT (Adaptive Activation Steering), a tuning-free method that adaptively shifts LLM activations toward the "truthful" direction at inference time, addressing diverse hallucination categories and improving truthfulness significantly across LLaMA, Alpaca, Vicuna and other models.
Exploiting Global and Local Hierarchies for Hierarchical Text Classification
Ting Jiang, Deqing Wang, Leilei Sun, Zhongzhi Chen, Fuzhen Zhuang, Qinghong Yang
Conference on Empirical Methods in Natural Language Processing (EMNLP) 2022
We propose HBGL, a hierarchy-guided BERT model that jointly exploits both global (static label structure) and local (per-sample dynamic) hierarchies for hierarchical text classification, achieving significant improvements over state-of-the-art on three benchmark datasets.
Exploiting Global and Local Hierarchies for Hierarchical Text Classification
Ting Jiang, Deqing Wang, Leilei Sun, Zhongzhi Chen, Fuzhen Zhuang, Qinghong Yang
Conference on Empirical Methods in Natural Language Processing (EMNLP) 2022
We propose HBGL, a hierarchy-guided BERT model that jointly exploits both global (static label structure) and local (per-sample dynamic) hierarchies for hierarchical text classification, achieving significant improvements over state-of-the-art on three benchmark datasets.