Biography
Kaixiang is a founding member of the Shopee LLM team. He has participated in training the Shopee proprietary LLMs from scratch, including the 7B, 13B, MoE, Max and Thinker series, whose performance on Southeast Asian languages and e-commerce tasks is among the top. He is primarily responsible for reinforcement learning, long-context extension and tool-use agents. He has participated end-to-end in building the LLM base-model training workflow, covering data, frameworks, pre-training, alignment and evaluation. Before that, he worked on query category classification within the Search team.
He previously led the development of the intelligent customer service systems for multiple WeBank products, including Weilidai, Weiyedai and Wechedai.
He received his PhD. degree in Computer Science in 2018 from the Hong Kong University of Science and Technology, supervised by Prof. Qiang Yang. He received his bachelor degree in computer science from Sun Yat-sen University, Guangzhou. He has published papers in top-tier AI conferences such as SIGKDD, IJCAI and AAAI, and serves as a reviewer for conferences such as AAAI, IJCAI and ACML.
His team won the world championship in the Demographic Prediction Task of the Nokia Mobile Data Challenge 2012.
Research Interest
Large Language Models, Reinforcement Learning, Long-context Modeling, Tool-use Agents, Dialogue Systems, Transfer Learning
Work Experience
-
Shopee, LLM Team & Search and Recommendation Team, Expert Algorithm Engineer | 2021 – Present
-
October 2023 – Present. Participated in training Shopee's 7B, 13B, MoE, Max and Thinker series of proprietary models, primarily responsible for reinforcement learning, long-context extension and tool-use agents.
-
March 2023 – October 2023. Led the iteration of the instruction-following models v0.3, v0.4 and v0.5 based on open-source base models, coordinating and participating in SEA-language vocabulary expansion, data preparation, training framework investigation and improvement, training, long-text extension and evaluation. Built the company's first LLM evaluation system: beyond standard benchmarks, created in-house benchmarks tailored to company scenarios across eight dimensions, an internal SEA-language benchmark, and an MMLU-SEA version. The resulting models reached production-grade performance for internal applications.
-
September 2021 – March 2023. Search keyword category relevance. Through bad case analysis, identified issues with primary and accessory categories in search and discovered numerous mislabels in manual annotations, then proposed a pseudo-labeling method to improve data quantity and quality, resulting in bad case rate -2.5%, OPU +0.28% and CTR +1.67%. Built category relevance models for new countries such as Poland from scratch.
-
-
WeBank, AI Project Team, Senior Researcher | 2018 – 2021
Responsible for the technology and algorithms of the WeBank intelligent customer service system, which provides 24-hour support for WeBank products such as Weilidai, Weiyedai and Wechedai, handling more than 90% of the messages automatically and reducing the workload of human agents by 90%. Also responsible for the SaaS cloud-based intelligent customer service system, serving business partners including Chang Hong, Bank of Shangrao and Bao Sheng Country Bank.
Led the team to bid, design, implement and deliver multiple intelligent customer service projects, including the project for the China National Clearing Center of the People's Bank of China, the Shenzhen bureau of local financial supervision and the Qianhai district bureau of local financial supervision.
-
Tencent TEG, Guangdiantong Intelligent Advertising Platform, Intern | 2014
Improved cold start ads recommendation with image features.
Education
-
Hong Kong University of Science and Technology, PhD. in Computer Science | 2011 – 2018
Supervisor: Prof. Qiang Yang
Thesis: Transfer Reinforcement Learning for Task-oriented Dialogue Systems
World Champion, Nokia Mobile Data Challenge, 2012
-
Sun Yat-sen University, Bachelor's Degree in Computer Science and Technology | 2007 – 2011
National Scholarship (Top 1%), 2008, 2009
First-Class Scholarship (Top 5%), 2008, 2009, 2010
Publication
-
Shopee LLM Team, Each Prompt Matters: Scaling Reinforcement Learning Without Wasting Rollouts on Hundred-Billion-Scale MoE, Arxiv Preprint, 2025. [Link]
-
Kaixiang Mo, Yuxin Shi, Weiwei Weng, Zhiqiang Zhou, Shuman Liu, Haibo Zhang, Anxiang Zeng, Mid-Training of Large Language Models: A Survey, Arxiv Preprint, 2025. [Link]
-
Shopee LLM Team, Compass-v3: Scaling Domain-Specific LLMs for Multilingual E-Commerce in Southeast Asia, Arxiv Preprint, 2025. [Link]
-
Anxiang Zeng, Haibo Zhang, Kaixiang Mo, Long Zhang, Shuman Liu, Yanhui Huang, Yawen Liu, Yuepeng Sheng, Yuwei Huang, Compass-Thinker-7B Technical Report, Arxiv Preprint, 2025. [Link]
-
Shopee LLM Team, Compass-V2 Technical Report, Arxiv Preprint, 2025. [Link]
-
Shopee LLM Team, Compass-13B Technical Report, Shopee Tech Report, 2024.
-
Shopee LLM Team, Compass: Large Multilingual Language Model for South-east Asia, Arxiv Preprint, 2023. [Link]
-
Shuangyin Li, Yu Zhang, Rong Pan, Kaixiang Mo, Adaptive Probabilistic Word Embedding, The Web Conference 2020 (WWW 2020), April 20-24, 2020 Taipei.
-
Kaixiang Mo, Transfer Reinforcement Learning for Task-oriented Dialogue Systems, HKUST PhD Thesis, 2018. [PDF] [Slides]
-
Kaixiang Mo, Yu Zhang, Qiang Yang, Pascale Fung, Cross-domain Dialogue Policy Transfer via Simultaneous Speech-act and Slot Alignment, Arxiv Preprint, 2018. [Link] [PDF] [Slides]
-
Kaixiang Mo, Yu Zhang, Qiang Yang, Pascale Fung, Fine Grained Knowledge Transfer for Personalized Task-oriented Dialogue Systems, Arxiv Preprint, 2017. [Link] [PDF] [Slides]
-
Weiyan Wang, Yuxiang WU, Yu Zhang, Zhongqi Lu, Kaixiang Mo, Qiang Yang, Integrating User and Agent Models: A Deep Task-Oriented Dialogue System, Arxiv Preprint, 2017. [Link]
-
Wenya Zhu, Kaixiang Mo, Yu Zhang, Zhangbin Zhu, Xuezheng Peng, Qiang Yang, Flexible End-to-End Dialogue System for Knowledge Grounded Conversation, Arxiv Preprint, 2017. [Link]
-
Kaixiang Mo, Qiang Yang, A Survey of Task-Oriented Dialogue Systems, PQE survey, 2017. [PDF] [Slides]
-
Kaixiang Mo, Shuangyin Li, Yu Zhang, Jiajun Li, Qiang Yang, Personalizing a Dialogue System with Transfer Reinforcement Learning, The Thirty-Second AAAI Conference on Artificial Intelligence (AAAI-18), February 2nd - February 7th, 2018, New Orleans, Lousiana, USA. [Link] [Slides]
-
Kaixiang Mo, Bo Liu, Lei Xiao, Yong Li, Jie Jiang, Image Feature Learning for Cold Start Problem in Display Advertising, International Joint Conference on Artificial Intelligence (IJCAI 2015), July 25th - July 31st, 2015, Buenos Aires, Argentina. [PDF]
-
Kaixiang Mo, Erheng Zhong, Qiang Yang, Cross-Task Crowdsourcing, In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD'2013), August, 2013, Chicago, Illinois, USA. [PDF] [Data]
-
Erheng Zhong, Ben Tan, Kaixiang Mo, Qiang Yang. User demographics prediction based on mobile data. Pervasive and Mobile Computing Journal. (Accepted)
-
Kaixiang Mo, Ben Tan, Erheng Zhong and Qiang Yang. Your Phone Understands You. Nokia Mobile Data Challenge Workshop, in conjunction with Pervasive'12, Newcastle, June 2012. [PDF] First Place in Demographic Prediction Task. [Link]
Academic Service
- Reviewer, AAAI / IJCAI / ACML