Tutorials

Abstract

The current mainstream paradigm of big language models highly relies on the brute force stacking of computing power and massive corpora, but in multilingual, especially low resource language scenarios, this method encounters data bottlenecks and performance ceilings. Due to a lack of understanding of the internal micro mechanisms of the model, existing cross language transfer and fine-tuning often lead to catastrophic forgetting or concept drift. The billions of parameters inside the large model are not chaotic black boxes. How do feedforward neural networks, as key value memory units of knowledge, encode concepts with single neurons or even neural circuits? How is factual knowledge in one language projected onto another language at the microscale? Which neurons are language independent abstract semantic centers, and which are language specific grammatical surfaces? We aim to break away from the homogenized training optimization and agent pipeline, and delve deeper into the micro neuron level of large models. We will analyze the operating mechanism of multilingual models at the scale of single neurons and subnetworks based on the latest advances in mechanism interpretability, and use this as a scientific basis to speculate optimization solutions for multilingual and low resource languages.

Biography

Shaolin Zhu is currently an Associate Research Professor and Ph.D. Supervisor at Tianjin University. His research interests include artificial intelligence and natural language processing. He has published more than 50 papers in leading international conferences and journals, including ACL, AAAI, KDD, Science China information science, and TASLP. He has served as an Area Chair or Senior Area Chair for several major international conferences, including ACL and COLING, and received the Outstanding Paper Award at COLING 2024.

Abstract

As large language models evolve into agents that plan, use tools, and interact with dynamic environments, evaluating their capabilities requires more than measuring answer accuracy. This tutorial introduces three complementary perspectives on LLM agent evaluation: outcome-based evaluation, process-level evaluation, and evaluation of open-ended tasks. Case studies of Memento and Research Claw Arena illustrate how these perspectives apply to agents that learn from experience and support scientific research, including idea generation and experimental design. Throughout the tutorial, we highlight the limitations of individual metrics and explain how complementary evaluation methods can reveal failures hidden by aggregate success rates. Participants will gain a practical framework for designing agent evaluations that connect measurable performance with reliable behavior and meaningful user value.

Biography

Linyi Yang is an Assistant Professor and Ph.D. supervisor at Southern University of Science and Technology (SUSTech), where he leads the GenAI Lab. He previously conducted research at University College London (UCL) and Westlake University. His research focuses on AI scientists and intelligent agents. His work has received over 14,000 citations on Google Scholar, and he serves as an Area Chair for leading conferences, including NeurIPS, ICLR, ICML, and ACL.

Abstract

Foundation models are opening new possibilities for AI systems that perceive their surroundings, act in physical environments, and interact with people and other agents. These developments connect Embodied AI with social intelligence, bringing new opportunities for building agents that can adapt to changing environments and participate in increasingly complex interactions.

This tutorial examines foundation-model applications along this progression. We begin with Embodied AI, exploring how multimodal understanding and reasoning support navigation, visual tracking, and action planning. We then extend the discussion to social intelligence, examining how agents interpret intentions, model behavior, and communicate and cooperate with others. Across these settings, we discuss the roles of memory, environmental feedback, and learning from interaction, together with challenges in generalization, long-horizon decision-making, and evaluation.

To connect these ideas with practice, the tutorial draws on representative applications and research platforms, including UnrealZoo, which provides photorealistic virtual environments for studying embodied perception and multi-agent interaction. These examples illustrate how simulation can support the development and evaluation of interactive agents and help investigate the relationship between physical competence and social behavior. The tutorial aims to give researchers and practitioners a practical perspective on foundation models in action and identify research opportunities at the intersection of Embodied AI and social intelligence.

Biography

Hao Chen is an assistant professor at the Faculty of Data Science, City University of Macau. Before that, he worked as a postdoctoral research fellow at The Hong Kong Polytechnic University. He received his Ph.D degree from Beihang University, China in 2022. He has published over 40 papers, such as TKDE, NeruIPS, ACL, SIGKDD, WWW, and SIGIR, etc. He received the Best Paper Award Honorable Mention at SIGIR 2023. His research interests include Embodied AI, Robotics, and Billion-scale Recommendation Systems.

Abstract

Large language models (LLMs) are reshaping how information extraction (IE) is performed and prompting a reassessment of its role. As general-purpose models can perform a wide range of extraction tasks and process documents directly, a central question is where dedicated IE research can continue to create value. This tutorial examines IE in the LLM era through three questions. First, how do LLMs change established assumptions about tasks, schemas, supervision, and benchmarks? We review representative evidence on the capabilities and limitations of general-purpose extraction. Second, when is explicit structure worth constructing and maintaining? Drawing on studies of structured retrieval, memory, and domain-specific applications, we examine when extracted representations provide benefits over using text directly, and what trade-offs accompany these benefits. Third, what research opportunities emerge from these changes? We discuss how these developments reshape the research priorities of IE and open new directions for its application. The tutorial aims to provide researchers and practitioners with a broader perspective on the changing role of IE, helping them identify meaningful research questions and opportunities for applying IE in the LLM era.

Biography

Xian-Ling Mao is a professor at Beijing Institute of Technology, specializing in natural language processing and big data mining. He has published over 80 papers in top‑tier conferences and journals, including SIGIR, WWW, SIGMOD, ACL, TOIS, and TKDE, and has led the development of scientific intelligent systems such as Hammer Scholar and Hammer PDF. He has received 2 Best Paper Awards and 5 Provincial/Ministerial-Level Science and Technology Progress Award, and has led or participated in over 10 national research projects, including National Key R&D Program of China and the Key Program of the National Natural Science Foundation of China.

Abstract

Generative retrieval (GR) and generative recommendation (GenRec) are reshaping information access by replacing the traditional index–retrieve–then–rank pipeline with a single sequence-to-sequence model that directly generates document or item identifiers from a query or user context. In GR, the corpus is consolidated into model parameters, and a query is mapped to its relevant docids autoregressively. In GenRec, the same paradigm is applied to recommendation: items are encoded as semantic IDs—often via residual quantization—and the model generates the next item or a slate of items directly from a user’s history, eliminating cascading errors and enabling end-to-end optimization. However, both paradigms face a fundamental optimization gap: the autoregressive generation is typically trained with token-level likelihood or exact-match supervision, which does not directly align with the retrieval or recommendation objective. Reinforcement learning (RL) has emerged as a principled remedy. In GR, RL is used to optimize docid generation toward retrieval metrics and downstream utility, rather than mere token accuracy. In GenRec, RL addresses the multi-answer nature of recommendation by shaping serendipity rewards, mitigating reward sparsity and aligning generation with potential new user preference. This tutorial provides a unified treatment of reinforced generative retrieval, covering the foundations of GR and GenRec, the limitations of likelihood-based training, and the design of RL objectives—reward shaping, policy optimization, and stability—that make generative retrieval and recommendation systems more effective, robust, and aligned with their end goals.

Biography

Xin Xin is now a tenured Associate Professor in the School of Computer Science and Technology of Shandong University, as a member of the General AI Lab (GAIL). He also serves as the Dean Assistant for the School of Artificial Intelligence, Shandong University. Before that, he got his PhD degree in computing science from University of Glasgow, under the supervision of Prof. Joemon Jose. Formerly, he got his master degree from School of Software Engineering in SJTU and his bachelor degree from XJTU. His research interests span recommender systems, reinforcement learning, information retrieval, NLP, etc. He has published more than 50 papers in top-tier ML & IR conferences including SIGIR, WSDM, CIKM, IJCAI, ACL, KDD, etc. His research is supported by the National Science Foundation of China. He also conducted collaborative research with companies, such as Tecent and Kuaishou.

Abstract

Syntactic parsing makes sentence structure explicit and has long been central to computational language understanding. With large language models (LLMs) now handling many language tasks without explicit parsing pipelines, the role of syntax deserves renewed attention. This tutorial traces the evolution of parsing from rule-based grammars through statistical and neural approaches to the LLM era. Drawing on foundational parsing research and recent developments, it introduces the core ideas behind constituency and dependency parsing, then examines how LLMs perform syntactic analysis, support resource construction, and broaden parsing across languages and domains. It also explains why fluent generation does not guarantee structurally correct analysis. This tutorial provides a historical perspective on parsing, an overview of current methods and limitations, and an understanding of how explicit syntax support the evaluation, interpretation, and control of LLMs.

Biography

Yuanhe Tian is currently an Assistant Professor at Zhongguancun Academy. He received his Ph.D. in Linguistics from the University of Washington in 2025. His research interests include the integration of linguistic knowledge into natural language processing, large language models, multimodal understanding and generation, and AI agents, with particular applications to social media and healthcare. He has published more than 70 papers in leading international conferences and journals, including ACL, EMNLP, NAACL, IEEE TNNLS, and IEEE TMI. His publications have received over 2,500 citations on Google Scholar, with an h-index of 27. He has served as an Area Chair for ACL, EMNLP, NAACL, and EACL, and as a reviewer for journals including IEEE TPAMI, TMM, TNNLS, and TMI.