CSoNet 2026Learning on Graphs in Changing Environments: From Topology to Physical Dynamics
Dr. Yi He
Department of Data Science at William & Mary.
Abstract. Modeling evolving processes on networks is essential to applications such as traffic forecasting and hydrologic prediction. Although graph neural networks (GNNs) provide a flexible framework for capturing spatial dependencies, they can struggle when predictions are governed by directional flows and physical interactions that extend beyond the observed graph. In particular, conventional message passing may suppress direction-sensitive information, unobserved forcing at network boundaries can introduce substantial prediction errors, and these errors can accumulate during long-horizon forecasting.
This talk presents our recent research on topology-aware and physics-guided graph learning to address these challenges. We first explore how directional difference operators and physical constraints can help GNNs capture flow dynamics that conventional graph aggregation tends to overlook. We then examine how unobserved external forcing affects prediction accuracy at network boundaries and discuss boundary-consistent learning methods that infer these missing influences. Building on these insights, we investigate physics-refined spatiotemporal forecasting for open-boundary hydrologic systems, with an emphasis on reducing error accumulation over time. Through examples from water-flow and traffic networks, these studies demonstrate the benefits of incorporating physical knowledge into graph learning.
Biography. Yi He is an Assistant Professor in the Department of Data Science at William & Mary, within the School of Computing, Data Sciences, and Physics. Prior to joining William & Mary, he was an Assistant Professor in the Department of Computer Science at Old Dominion University from 2021 to 2024. He received his Ph.D. from the Center for Advanced Computer Studies at the University of Louisiana at Lafayette in 2020, under the supervision of Prof. Xindong Wu, and his B.E. from Harbin Institute of Technology in 2013. His research focuses on learning from evolving and networked data, with interests spanning data mining, streaming algorithms, graph learning, and AI applications in environmental and conservation sciences. He is a recipient of the NSF CAREER Award (2026), NSF CRII Award (2023), and IEEE TCII Volunteer Award (2022). He is a Senior Member of the IEEE.
Computer Science of Language Models
Dr. Xinrun Wang
School of Computing and Information Systems, Singapore Management University.
Abstract. Large language models have become general-purpose problem solvers, a shift as consequential as the arrival of the digital computer. Yet the theoretical scaffolding we built for classical computing, complexity classes, algorithm analysis, formal correctness, doesn't quite fit them. In this talk, I argue that language models deserve a theoretical discipline of their own, which I call the Computer Science of Language Models (CSLM). I'll start from seven ways LMs depart from traditional computation: they generate, solve, and verify rather than merely solve and verify; they compute in tokens rather than operations; they are probabilistically reliable and prone to hallucination rather than deterministically correct; they understand multimodal input natively; they are general-purpose rather than problem-specific; they depend on a coupled training–inference process; and they carry inherent knowledge. Each of these breaks an assumption that classical theory takes for granted, and together they motivate new complexity measures and evaluation methods. At the center of the proposal is token complexity — a candidate analogue to time and space complexity for the LM era. I'll connect it to active threads in scaling laws, memory, tool use, and fine-tuning, and lay out the open problems I find most pressing: what an “LM-complete” problem might be, where reliability bounds come from, how emergence should be characterized, and how to reason about multimodal complexity. Complexity theory took fifty years to mature; I'll close with a five-year roadmap (2026–2030) for getting CSLM off the ground and an argument for why moving from empirical scaling to principled, theory-grounded engineering would change how we understand both these models and the problems we ask them to solve.
Biography. Xinrun Wang is an Assistant Professor and Lee Kong Chian Fellow at the School of Computing and Information Systems, Singapore Management University, where he has led the DIGA research group since July 2024. Previously, he was a Research Assistant Professor at Nanyang Technological University (NTU). He received his Ph.D. from NTU in 2020. His research focuses on building capable and reliable agents through reinforcement learning, multi-agent learning, and multimodal large language models. His interests span foundation models and agents, the computer science of language models, and AI for science, with applications in finance, urban security, and materials discovery. His recent work explores computer-use and autonomous research agents, world models for agent decision-making, and formal verification for scientific computing.
Causal-Aware Agentic Architectures for Multimodal Regulatory Compliance in High-Stakes Digital Ecosystems
Assoc. Prof. Dr. Nguyen Thanh Binh
University of Science, Vietnam National University Ho Chi Minh City (VNU-HCM), Vietnam.
Abstract. Governing high-stakes information in multimodal digital networks poses a profound challenge to established moderation frameworks. As specialized services, from medical diagnostics to regulated financial advice, increasingly permeate social discourse, the gap between textual disclaimers and the visual content they accompany has reached a critical threshold. Existing moderation systems, predominantly predicated on veracity-based filtering, lack the structural depth to resolve the cross-modal semantic discrepancies inherent in modern digital obfuscation. This talk introduces a paradigm shift towards authorization-aware regulatory reasoning. We delineate a causal-aware agentic framework that enforces strict liability constraints in complex, multimodal environments. By integrating inference-time causal interventions with dynamic regulatory mapping, this architecture enables systems to isolate information gain from visual evidence while mitigating deceptive alignment and spurious modality correlations. The presentation will examine deploying these agentic architectures in high-stakes governance, emphasizing the need for systemic transparency, zero-day adaptability, and robust evasion resistance. Grounded in empirical validations on multimodal regulatory benchmarks, this session provides a rigorous roadmap for architecting next-generation compliance systems. We conclude by evaluating the future trajectory of autonomous agents as critical infrastructure for maintaining the integrity and stability of global digital ecosystems against increasingly sophisticated regulatory circumvention.
Biography. Dr. Binh T. Nguyen is an Associate Professor of Computer Science and Head of the Department of Computer Science at the University of Science, Vietnam National University Ho Chi Minh City (VNU-HCM), Vietnam. With over 15 years of experience in artificial intelligence and machine learning, he has authored and co-authored more than 200 research publications, including papers presented at leading international conferences such as NeurIPS, ICLR, AAAI, and ACM Multimedia. He is also an inventor on four patent applications filed in the United States and Canada and three national patents. Beyond academia, Prof. Nguyen has extensive experience leading AI research and development initiatives, building high-performing R&D teams, and bridging cutting-edge research and real-world applications. He has collaborated with technology companies and startups to translate AI innovations into practical, impactful products and solutions. His research contributions have been recognized through numerous national and international honors, including the Golden Globe Award in Science and Technology (Vietnam, 2021), the Mathematical Research Award (2020) from the Vietnam Institute for Advanced Study in Mathematics, the Best Paper Award at SoMeT 2018, and Best Paper Runner-up Awards at KSE 2022 and PACLIC 2022. He also led teams that won championship titles at the ACM ICMR Lifelog Search Challenge in both 2020 and 2021.