The 20th China Society for Chinese Information Processing Summer School and Advanced Technology Workshop (CIPS ATT45)

Speaker 1: Junxian He

Speaker:Junxian He
Affiliation:Hong Kong University of Science and Technology
Title:Reinforcement Learning-Centered Deep Reasoning for Large Models
Abstract: This talk will systematically review the fundamental principles and key technologies of deep reasoning for large models. First, we will outline the development trajectory of large model reasoning, with a focus on the paradigm evolution from traditional chain-of-thought to long chain-of-thought. Second, we will focus on the core technology driving this evolution—reinforcement learning, and compare and analyze the basic principles and design ideas of different reinforcement learning algorithms around the two major goals of reasoning performance and efficiency. Finally, we will explore the current development bottlenecks faced by deep reasoning technology and potential future research directions.
Bio: Junxian He is an Assistant Professor in the Department of Computer Science and Engineering at Hong Kong University of Science and Technology. He received his Ph.D. in Natural Language Processing from the School of Computer Science at Carnegie Mellon University in 2022. His recent research focuses on large model reasoning. He serves as an Area Chair for ICLR, ACL, and EMNLP. His representative works include Unify-PEFT, C-Eval, CodeIO, SimpleRL, etc.

Speaker 2: Xiting Wang

Speaker:Xiting Wang
Affiliation:Renmin University of China
Title:Exploring Precise Neuron Control and Fundamental Value Alignment in Large Language Models
Abstract: The complex structure of large language models makes it increasingly difficult for people to understand, predict, and control them. How should we identify bottlenecks in large models, find directions for performance improvement, and ensure that large models are safe and aligned with human intentions in various complex real-world scenarios? What challenges and opportunities does the new paradigm of artificial general intelligence bring to explanation and alignment? This lecture will explore these questions, introduce the latest advances in large model explanation and alignment, focus on highly generalizable concept explanation and how it interprets neuron functions, and discuss macro fundamental value alignment and its possible integration points with sociology and measurement science.
Bio: Xiting Wang is an Associate Professor at the Gaoling School of Artificial Intelligence, Renmin University of China, and a Distinguished Scholar at Renmin University of China. He graduated with a bachelor's degree from Tsinghua University in 2011 and received his Ph.D. from Tsinghua University in 2017. His research focuses on large model interpretability and safety alignment. He has published over 60 papers, twice selected for IEEE TVCG Class A journal cover paper awards, and received the CCF Natural Science Second Prize (fourth contributor). He serves as an Area Chair for IJCAI and AAAI, an editorial board member of the Q1 journal Visual Informatics, and a distinguished senior program committee member of AAAI. Related achievements have been implemented in three Microsoft products.

Speaker 3: Zhongyu Wei

Speaker:Zhongyu Wei
Affiliation:Fudan University
Title:Large Model Agent-Driven Social Simulation: Progress and Challenges
Abstract: Social simulation aims to model the behavioral patterns of target individuals or groups, predict the evolution trends of events, and assist real-world decision-making by constructing references to the real world. With the improvement of large language models' role-playing capabilities, more and more scholars are introducing large language models into social science research, achieving positive results in scenarios such as simulating social surveys, evaluating communication effects, and simulating political behavior. Large model agent-driven social behavior simulation can be divided into three levels: (1) Individual simulation: simulating specific individuals or demographic groups; (2) Scenario simulation: simulating multi-agent interaction and collaboration in specific scenarios; (3) Social simulation: simulating interactions among larger-scale populations, reflecting social dynamics. This workshop will review recent research progress in large model agent-driven social simulation and outline future trends and challenges in related research.
Bio: Zhongyu Wei is an Associate Professor and doctoral supervisor at the School of Data Science, Fudan University, head of the Data Intelligence and Social Computing (Fudan DISC) research group, full-time mentor at Shanghai Chuangzhi Academy, and holds a Ph.D. from The Chinese University of Hong Kong. His main research areas include multimodal large models and social simulation. He has published over 100 academic papers in international journals and conferences such as ICML, ICLR, and ACL. Representative achievements include the multimodal multi-step reasoning large model Volcano and China's first open-source social media simulation framework HISim. He serves as Senior Area Chair (SAC) for ACL 2023, EMNLP 2024, and NAACL 2025, and as Program Committee Chair for YSSNLP 2019, CCAC 2023, and NLPCC 2024. He serves as Secretary-General of the CIPS Affective Computing Special Committee, formerly served as Deputy Director of the CIPS Young Workers Committee Executive Committee, and has received the CIPS Social Media Processing Special Committee Rising Star Award, Shanghai Rising Star Program, and CCF Natural Language Processing Special Committee Rising Scholar Award.

Speaker 4: Wenxi Li, Weidong Zhan, Weiwei Sun

Speakers:Wenxi Li, Weidong Zhan, Weiwei Sun
Affiliations:Minzu University of China, Peking University, University of Cambridge
Title:Language Models and Linguistic Theory
Abstract: In recent years, with the rapid development of large language models, their capabilities in language learning and use have continuously expanded our scientific understanding of language. Language models and linguistic theory are forming unprecedented deep interactions. This lecture focuses on the interdisciplinary research between language models and linguistic theory, aiming to explore how linguistic theory can provide new cognitive foundations and modeling paradigms for language models, and how language models can react upon linguistic theory to deepen understanding of language. We will systematically explore possible ways of integrating the two from the following three research perspectives:
• In the resource evaluation dimension, how to design evaluation data based on linguistic theory to reveal the performance of language models in grammar, semantics, pragmatics, and other aspects?
• In the theory verification dimension, how to treat language models as controllable and repeatable experimental platforms to implement, refine, and verify linguistic theoretical hypotheses?
• In the engineering practice dimension, how to integrate linguistic theory into the design and training of language models to make them closer to human language processing mechanisms?
Bio: Wenxi Li holds a Bachelor's degree in Chinese Language and Literature from Peking University and a Ph.D. in Modern Chinese (Chinese Information Processing) from Peking University, with joint Ph.D. training at the Department of Computer Science and Technology, University of Cambridge. She is currently an Assistant Professor at Minzu University of China. Her research direction is computational linguistics, mainly focusing on multilingual syntactic, lexical, and semantic computational modeling. She has published multiple papers in domestic and international academic journals such as Computational Linguistics and Language Sciences, as well as international conferences such as ACL and NAACL.
Weidong Zhan is a Professor and doctoral supervisor in the Department of Chinese Language and Literature at Peking University. He is the Deputy Director of the Center for Chinese Linguistics at Peking University and the Deputy Director of the Institute of Computational Linguistics at Peking University. He is mainly engaged in research on modern Chinese formal grammar, language knowledge engineering and Chinese information processing, and language and script applications. Representative achievements include "Research on Modern Chinese Phrase Structure Rules for Chinese Information Processing," the national language and script standard "Usage of Numbers in Publications" and its supporting reader "Interpretation of 'Usage of Numbers in Publications'." He has co-edited multiple textbooks including "Introduction to Computational Linguistics," "Natural Language Processing," and "Modern Chinese." He has published multiple papers in domestic and international academic journals. In recent years, his research interests have mainly focused on the construction of modern Chinese construction resource databases and the evaluation of machine language understanding capabilities for cognitive intelligence.
Weiwei Sun is an Associate Professor and doctoral supervisor in the Department of Computer Science and Technology at the University of Cambridge. She mainly explores how to use computational models to explain the structure and changes of language, and is committed to building multilingual computational models with both theoretical depth and application value. She has published over forty papers in top conferences and journals in the field of computational linguistics. She has served multiple times as Program Chair and Area Chair for international conferences such as ACL, EMNLP, and CCL.

Speaker 5: Yixin Cao

Speaker:Yixin Cao
Affiliation:Fudan University
Title:Model Utility Laws: Towards Generalizable Evaluation
Abstract: In recent years, the proposal of numerous evaluation benchmarks has greatly promoted the training and application of large models, but these benchmarks still inevitably face problems such as rapid obsolescence, data contamination, and unfair comparison, gradually becoming core bottlenecks in the "second half" of artificial intelligence. This report first summarizes two evaluation paradigm shifts brought by large models: from task-based to capability-based evaluation, and from manual to automatic evaluation. Second, based on this, it deeply analyzes the fundamental limitations of traditional evaluation paradigms in the era of large models: driven by scaling laws, model capabilities can continuously leap through expanding data, parameters, and computing power, while evaluation datasets cannot expand synchronously due to cost and efficiency considerations. Thus, we are forced to use limited samples to measure nearly infinite model capabilities, the so-called "evaluation generalizability" dilemma. Finally, to break this dilemma, the goal of the generalizable evaluation paradigm is to predict and measure the potential capabilities that models have not yet explicitly demonstrated. To this end, we propose the Model Utilization Index (MUI), introducing mechanistic interpretability methods to complement traditional performance metrics, comprehensively evaluating model capabilities from a "utility" perspective (including potential capabilities beyond datasets). Large-scale experiments show that MUI has an inverse relationship with performance, from which we summarize the "Utility Law" that universally exists in mainstream LLMs. Based on this law, we further derive four corollaries, addressing key challenges including training determination, data contamination issues, model comparison fairness, and data diversity.
Bio: Yixin Cao is a Young Researcher and doctoral supervisor at Fudan University. She received her Ph.D. from Tsinghua University and has successively held positions as postdoctoral researcher, research assistant professor, and assistant professor at the National University of Singapore, Nanyang Technological University, and Singapore Management University. She is a recipient of the National Youth Talent Program and the Shanghai Youth Leading Talent Program. Her research areas include natural language processing, knowledge engineering, and multimodal information processing. She has published over 80 papers in internationally renowned conferences and journals, with over 8,600 Google Scholar citations, and has been selected multiple times for oral presentations at top international conferences in the field. Her research achievements have received two international conference best paper awards and nominations. She has received honors including the Lee Kong Chian Fellowship, Google South Asia & Southeast Asia Awards, and AI 2000 Most Influential Scholar Award nomination, and is listed as Elsevier 2024 Global Top 2% Scientists. She serves as demonstration program chair, area chair for multiple international conferences, and reviewer for international journals.