On July 20, 2025, the AI and Large Model Frontier Research Team with an International Perspective from our college successfully completed the pre-departure training and officially kicked off the academic exchange activities at Nanyang Technological University in Singapore. This research expedition was jointly organized by our college and the School of Computer Science and Data Science (CCDS) of Nanyang Technological University. It was led and guided by three teachers from our college: Associate Professor Li Wenbin, Ji Wei, and Undergraduate Secretary Yan Yuting.
The project focuses on large model technology and its applications, covering cutting-edge fields such as multimodal learning, agent development, and reinforcement learning. It aims to deeply explore the latest research achievements and application practices in the field of artificial intelligence. Through this scientific expedition, it not only promotes academic exchanges across cultures but also provides valuable learning opportunities for participating students, which is conducive to enhancing their research capabilities and global perspectives.
Pre-departure training
Lecture One
Professor Dai Wangzhou, the dean of our school's teaching department, delivered the first lecture of the pre-departure training for the expedition team. During the meeting, Professor Dai, starting from the overall development of the college, elaborated in depth on the educational goals and significance of this international expedition, emphasized the importance of team discipline and collective consciousness, and encouraged the students to be proactive and courageous in exploration during the expedition, fully showcasing the outstanding demeanor of the students from Nanjing University's School of Intelligent Science and Technology.
The Second LectureSecond Lecture
Our school's Li Wenbin, Ji Wei and Yan Yuting teachers conducted the second training session for the students before their trip. At the beginning of the lecture, Professor Li Wenbin provided a detailed introduction to the team composition, project theme, cooperation partners and management methods, ensuring that each team member had a comprehensive understanding of the upcoming itinerary and was well-prepared. Subsequently, Professor Ji Wei supplemented the explanations with information about precautions during the scientific expedition based on the local conditions in Singapore. Finally, Teacher Yan Yuting made detailed supplementary explanations regarding the living arrangements during the trip, including transportation, diet and daily necessities, further emphasizing the importance of precautions and personal safety. This training not only enhanced the team members' sense of organization and discipline, but also laid a solid foundation for the successful continuation of the subsequent scientific research activities.
The Third Lecture
Our associate professor Fu Chaoyou, who is on a probationary appointment, delivered a special lecture titled Frontier Developments of Multimodal Large Language Models to the expedition team. He systematically reviewed the development history, core challenges, and future directions of multimodal models. During the interactive session, students actively raised questions on hot topics such as model performance improvement and cross-modal semantic alignment. The atmosphere in the room was lively. This lecture not only helped students systematically understand the core developments in the field of multimodal language models, but also sparked their strong interest in research in this area.
Record of the scientific expedition process
During the scientific expedition, several renowned experts and scholars from the School of Computing and Data Science at Nanyang Technological University delivered rich and diverse specialized lectures to the expedition team members, covering multiple cutting-edge fields such as multi-agent systems, reinforcement learning, large language model technologies, and multimodal applications.During the scientific expedition, several renowned experts and scholars from the School of Computing and Data Science at Nanyang Technological University delivered rich and diverse specialized lectures to the expedition team members, covering multiple cutting-edge fields such as multi-agent systems, reinforcement learning, large language model technologies, and multimodal applications. The experts not only systematically reviewed the theoretical foundations and development trajectories of these fields, but also combined practical cases and the latest research results to guide the students to deeply reflect on research topic selection and explore the paths for technology implementation.
The experts not only systematically reviewed the theoretical foundations and development trajectories of these fields, but also combined practical cases and the latest research results to guide the students to deeply reflect on research topic selection and explore the paths for technology implementation.
Professor An Bo, the dean of the School of Computing and Data Science and the director of the Artificial Intelligence Department at Nanyang Technological University, introduced the theme of this expedition at the opening ceremony and gave a wonderful presentation based on his research in multi-agent systems, game theory, and reinforcement learning.Professor An Bo, the dean of the School of Computing and Data Science and the director of the Artificial Intelligence Department at Nanyang Technological University, introduced the theme of this expedition at the opening ceremony and gave a wonderful presentation based on his research in multi-agent systems, game theory, and reinforcement learning. Professor An reviewed the origin and challenges of multi-agent systems, analyzed the differences between game theory in theory and industrial applications, emphasized the importance of data-driven methods, introduced the development of reinforcement learning and the transformation brought by large-scale language models (LLMs), pointed out that the key to LLM reasoning ability lies in long-term assessment and planning. He stressed that finding the right question is the key to scientific research and encouraged drawing inspiration from the industrial sector.
After the lecture, the students actively raised questions and the on-site discussion was very lively. Professor An reviewed the origin and challenges of multi-agent systems, analyzed the differences between game theory in theory and industrial applications, emphasized the importance of data-driven methods, introduced the development of reinforcement learning and the transformation brought by large-scale language models (LLMs), pointed out that the key to LLM reasoning ability lies in long-term assessment and planning. This sharing pointed out the direction for the students' subsequent research, and ignited their enthusiasm to explore real-world problems and promote the application of AI in society. He stressed that finding the right question is the key to scientific research and encouraged drawing inspiration from the industrial sector.
After the lecture, the students actively raised questions and the on-site discussion was very lively. This sharing pointed out the direction for the students' subsequent research, and ignited their enthusiasm to explore real-world problems and promote the application of AI in society.
Professor Tao Dacheng, a visiting professor at the School of Computing and Data Science at Nanyang Technological University, presented a lecture in the form of a fun quiz, explaining the basic principles of the Transformer model and its wide range of applications in various tasks in an easy-to-understand manner.Professor Tao Dacheng, a visiting professor at the School of Computing and Data Science at Nanyang Technological University, presented a lecture in the form of a fun quiz, explaining the basic principles of the Transformer model and its wide range of applications in various tasks in a clear and concise manner. He focused on key questions such as Why use the Decoder-only structure? How to determine the number of attention blocks? Will positional encoding cause pattern repetition? and guided the students to think about the engineering considerations and theoretical basis behind model design. He focused on key questions such as Why use the Decoder-only structure? How to determine the number of attention blocks? Will positional encoding cause pattern repetition? and guided the students to think about the engineering considerations and theoretical basis behind the model design. In addition, Professor Tao also analyzed the advantages of Layer Normalization in sequence tasks, the nonlinear expression ability of the GELU activation function, and the memory mechanism behind the powerful modeling ability of the Transformer, stimulating the students' strong interest in deep learning architectures. In addition, Professor Tao also analyzed the advantages of Layer Normalization in sequence tasks, the nonlinear expression ability of the GELU activation function, and the memory mechanism behind the powerful modeling ability of the Transformer, stimulating the students' strong interest in deep learning architectures.
Wang Wenya, an assistant professor at the School of Computer Science and Engineering of Nanyang Technological University, systematically reviewed the development history, training methods, and application scenarios of large language models (LLMs), providing students with a solid theoretical foundation. Professor Wang explained the decoding strategies of generative models and analyzed their applicable scenarios in text generation, deeply delved into the development of pre-training and fine-tuning technologies, from the early Word2Vec to the evolution of modern models, and introduced the differences and respective advantages of masked language models (MLMs) and causal language models (CLMs). At the application level, Professor Wang introduced the retrieval-enhanced generation (RAG) technology and its advanced pipeline design, and discussed the performance of LLMs in deductive, inductive, and abductive reasoning. Finally, she illustrated the effect of instruction fine-tuning using FlanT5 and introduced the application of RLHF in optimizing the response quality of the model. The lecture content was comprehensive and provided important guidance for students' subsequent research.
Dr. Zheng Longtao from the School of Computing and Data Science at Nanyang Technological University shared his numerous achievements in the field of agents (intelligent entities) systems, particularly in web agents based on large language models (LLM). He elaborated on key technical issues such as how to filter irrelevant information from web code, improve the observation efficiency of agents, and achieve longer historical information modeling. He also discussed core considerations in model design, such as why agents need grounding, why only one face should be retained, and why information should be processed by concatenation rather than addition. Additionally, he introduced the exploration of agents in terms of having memory capabilities and self-improvement mechanisms, inspiring students to have in-depth thinking about the future development direction of intelligent agents.
Professor Lin Guosheng, an associate professor from the School of Computing and Data Science at Nanyang Technological University, delivered a captivating lecture on generative AI to the students.Professor Lin Guosheng, an associate professor from the School of Computing and Data Science at Nanyang Technological University, delivered a captivating lecture on generative AI to the students. Focusing on its fundamental principles, technological advancements, and application scenarios, he conducted a detailed explanation through on-site demonstrations and video presentations, covering the evolution process from text to multimodal generation systems. Focusing on its fundamental principles, technological advancements, and application scenarios, he conducted a detailed explanation through on-site demonstrations and video presentations, covering the evolution process from text to multimodal generation systems. Professor Lin began by using a code assistant as an example, showcasing the practical application of generative AI in software development, such as how GitHub Copilot automatically generates code snippets based on the context. Professor Lin began by using a code assistant as an example, showcasing the practical application of generative AI in software development, such as how GitHub Copilot automatically generates code snippets based on the context. Subsequently, he introduced the core technologies for image, audio, and video generation, including GAN, VAE, and diffusion models, and demonstrated the powerful capabilities of AI in image generation, style transfer, and speech synthesis through presentations. Subsequently, he introduced the core technologies for image, audio, and video generation, including GAN, VAE, and diffusion models, and demonstrated the powerful capabilities of AI in image generation, style transfer, and speech synthesis through presentations. The lecture was cutting-edge and diverse in form, deepening the students' understanding of generative AI technology and inspiring their enthusiasm for thinking about and exploring future AI applications. The lecture was cutting-edge in content and diverse in form, deepening the students' understanding of generative AI technology and inspiring their enthusiasm for thinking about and exploring future AI applications.
Dr. Xue Zhenghai from the School of Computing and Data Science at Nanyang Technological University delivered a profound and insightful presentation on the latest research advancements in reinforcement learning (RL). Drawing on his own research, he delved into the issue of strategy optimization in scenarios where human interaction is imperfect. Dr. Xue pointed out that current research on reinforcement learning in simulators has become relatively mature, but its application in real robot systems still faces many challenges. He also analyzed the differences in the application of reinforcement learning and supervised learning in the field of robotics, and shared his thoughts on future research directions. During the interactive session, Dr. Xue patiently answered questions from the students regarding algorithm design, experimental verification, and research methods, encouraging them to pay attention to the latest trends, such as relevant public accounts, academic communities, and the latest achievements of domain experts, to broaden their horizons and grasp the trends. The entire presentation was cutting-edge and highly inspiring, providing valuable ideas and direction guidance for the students to conduct related research.
In the latter part of the scientific expedition project, our teachers and students will continue to conduct in-depth discussions, exploring further in areas such as the application implementation of multimodal large models, the actual deployment of intelligent agent systems, and the optimization strategies of reinforcement learning in real-world scenarios.
