On December 13-14, 2025, the School of Statistics and Data Science at Shanghai University of Finance and Economics successfully held the "2025 FAIC Artificial Intelligence Foundation Conference". This conference is jointly hosted by the School of Statistics and Data Science at Shanghai University of Finance and Economics, the Big Data Research Institute at Shanghai University of Finance and Economics, and the City of Statistics. The conference brought together nearly 100 experts, scholars and doctoral students from top universities at home and abroad, such as Tsinghua University, Peking University, Shanghai Jiaotong University, Fudan University, Renmin University of China, University of Hong Kong, the Chinese University of Hong Kong (Shenzhen), Xi'an Jiaotong University, as well as industry experts, such as ByteDance, Ming Dynasty Investment, and Digital Qihuan, to focus on the cutting-edge progress in key areas such as big models, artificial intelligence theory, and intensive learning, and to deeply explore the boundaries and future of technology.
The preparation work for the conference was jointly completed by Teng Jiaye from Shanghai University of Finance and Economics, Lv Kaifeng from Tsinghua University, Ma Ziye from City University of Hong Kong, and Wei Taiyun from the Capital of Statistics. The FAIC Artificial Intelligence Foundation Conference originated from the FAI Seminar, an online seminar on artificial intelligence foundations. Previously, FAI Seminar had been successfully held for three years, with over 70 online academic lectures and more than 350000 views. This year, the conference went from online to offline, aiming to further enhance mutual exchanges among young scholars. The conference not only focuses on practical applications in the present, but also cares about the future development of the discipline; Not only discussing the specific engineering implementation, but also questioning the theoretical basis behind it.
On the morning of December 13th, Professor Li Jian from Tsinghua University first presented a report titled "Understanding LLM Behaviors via Compression: Data Generation, Knowledge Acquisition and Scaling Laws". Starting from the classic relationship between compression and prediction, he delved into the Scaling Laws and illusion phenomena of large models. Subsequently, Professor Liu Yong from Renmin University of China delivered a speech titled "Transformers and Architecture Design: Observing from the Perspective of Energy", constructing a new framework from the perspective of statistical physics energy and proposing that the essence of self attention mechanism is the process of minimizing Helmholtz free energy. In his report "Sampling Process and Training Dynamics Analysis of Diffusion Models", Professor Li Shuai from Shanghai Jiao Tong University focused on how to improve the efficiency of the model by starting from the sampling process and training dynamics. Chen Lesi, a student from Tsinghua University, shared the latest research on high-order acceleration of minimax optimization problems and introduced high-order algorithms that break the lower bound of algorithm convergence rate.
On the afternoon of December 13th, Professor Sun Ruoyu from the Chinese University of Hong Kong (Shenzhen) first reported on the characteristics of the Hessian array in neural networks and its implications for the design of large-scale model algorithms. Based on the analysis of the Hessian array, a more efficient Adam mini optimizer was introduced. Professor Yuan Yang from Tsinghua University presented a framework for large-scale software assisted generation based on topological theory, demonstrating how to use topological theory to achieve parallel generation and fine-grained control of large-scale software systems. Professor Zou Difan from the University of Hong Kong revealed in his report "On the Mechanism Interpretability of LLM for Fine tuning and Reasoning" that reinforcement learning and fine-tuning have completely different mechanisms of action in large-scale model inference. Subsequently, Zhong Han from Peking University gave a speech titled "Principled Reinforcement Learning and its Role in Large Language Models", introducing the key role of generalized Eluder coefficients in characterizing the statistical complexity of decision problems.
At the meeting on the morning of December 14th, Professor Chang Xiangyu from Xi'an Jiaotong University first shared "Valuation Method of Data Elements Based on Large scale Cooperative Game Theory" and discussed how to use cooperative game theory to solve the problem of fair distribution of data elements in the MaaS scenario. Professor Zhang Huishuai from Peking University introduced a study titled "AdamS: Momentum Itself Can Be A Normalizer for LLM Pretraining and Post training", proposing an efficient optimizer called AdamS that does not require second-order moment estimation. Professor Zhang Linfeng from Shanghai Jiao Tong University discussed in detail the caching strategy, decoding strategy, and variable length generation method of the diffusion big language model in his article "Reasoning Acceleration Based on the Diffusion Big Language Model". Finally, Dr. Chen Huanran from Tsinghua University reported on "Unveiling the Basin Like Loss Landscape in Large Language Models", which deeply analyzed the formation of the "basin" structure in the loss landscape of large models and its significance for model capability.
This conference fully embodies the interdisciplinary characteristics of basic research in artificial intelligence, providing valuable learning and exchange opportunities for young scholars and doctoral students at home and abroad. The academic atmosphere at the conference was strong, and each presentation sparked in-depth discussions. The organizer hopes to build a higher-level academic exchange platform through this seminar, gather the strength of young people, promote ideological collision, and jointly explore new directions and opportunities for the development of artificial intelligence.
Host: School of Statistics and Data Science, Shanghai University of Finance and Economics Big Data Research Institute, City of Statistics, Basic Research on Artificial Intelligence
Sponsors: ByteDance Seed Team, Mingxuan Investment, Shanghai Digital Qihuan Artificial Intelligence Technology Co., Ltd
Image and Text | FAIC Meeting Group
