The 1st CogMAEC Workshop

Bio: Dr. Minlie Huang is a professor of Tsinghua University and the deputy director of its Foundation Model Center. He was supported by the National Distinguished Young Scholar project and has won several awards in Chinese AI and information processing societies, including the Wuwenjun Technical Advancement Award and the Qianweichang Technical Innovation Award. His research fields include large-scale language models, language generation, AI safety and alignment, and social intelligence. He authored the Chinese book "Modern Natural Language Generation," published more than 200 papers in premier venues (ICML, ICLR, NeurIPS, ACL, EMNLP, etc.) with over 29,000 citations, and has been named both an Elsevier China Highly Cited Scholar since 2022 and an AI 2000 influential AI scholar since 2020. He has won several best paper awards or nominations at major international conferences (IJCAI, ACL, SIGDIAL, NLPCC, etc.) and was a key contributor to large foundation models such as ChatGLM, GLM-4.5, GLM4.1v-thinking, and CharacterGLM. He serves as an associate editor for TNNLS, TACL, CL, and TBD, has acted as senior area chair of ACL/EMNLP/IJCAI/AAAI more than ten times, and maintains a homepage at http://coai.cs.tsinghua.edu.cn/hml/.

Abstract: Today’s LLM is designed as a machine tool to facilitate the efficiency, productivity, and creativity of human works. However, social intelligence, which is a significant feature of human intelligence, has been largely neglected in current research. Future AGI must have not only machine intelligence but also social intelligence. In this talk, the speaker will discuss how to embrace social intelligence with LLMs for emotion understanding, emotional support, behavior simulation, modeling cognition and theory of mind, and he will also present real-world applications for mental health.

Bio: Dr. Soujanya Poria is an Associate Professor at Nanyang Technological University (NTU), Singapore. His research explores large language models, reasoning, AI safety, embodied AI, multimodal AI, and natural language processing. He completed his Ph.D. in Computer Science at the University of Stirling, UK. Before joining NTU, he worked at the Singapore University of Technology and Design as an Associate Professor and at the Institute of High Performance Computing (IHPC), A*STAR, as a Senior Scientist.

Abstract: Vision-language-action (VLA) models are quickly becoming central to embodied AI, building on the breakthroughs of large language models and vision-language models. Their promise lies in something simple yet profound: the ability to follow natural language instructions and turn them into real-world actions. In this talk, Prof. Poria will walk through ten milestones that mark the progress and challenges ahead for VLA models—ranging from multimodality and reasoning to data, evaluation, generalization across robots, efficiency, whole-body coordination, safety, intelligent agents, and human collaboration. Each of these represents both a technical challenge and a stepping stone toward truly capable embodied systems. He will also highlight emerging trends that are shaping the future: spatial understanding, modeling world dynamics, post-training refinements, and synthetic data generation. Together, these directions point to a roadmap for accelerating VLA models toward real-world deployment and broader societal impact, sparking discussion on how the community can bring VLA models from promising prototypes to widely adopted, trustworthy, and useful embodied intelligence.

Bio: Dr. Amir Zadeh is a Staff ML Researcher at Lambda. He received his Ph.D. in Artificial Intelligence from Carnegie Mellon University with a focus on multimodal machine learning. Dr. Zadeh has published in top machine learning venues including NeurIPS, ICLR, CVPR, and ACL, and has served as an organizer, senior area chair, and committee member for leading conferences and workshops.

Abstract: Autoregressive (AR) models have long dominated the landscape of large language models, driving progress across a wide range of tasks. Recently, diffusion-based language models have emerged as a promising alternative, though their advantages over AR models remain underexplored. In this paper, we systematically study masked diffusion models in data-constrained settings—where training involves repeated passes over limited data—and find that they significantly outperform AR models when compute is abundant but data is scarce. Diffusion models make better use of repeated data, achieving lower validation loss and superior downstream performance. We find new scaling laws for diffusion models and derive a closedform expression for the critical compute threshold at which diffusion begins to outperform AR. Finally, we explain why diffusion models excel in this regime: their randomized masking objective implicitly trains over a rich distribution of token orderings, acting as an implicit data augmentation that AR’s fixed left-toright factorization lacks. Our results suggest that when data, not compute, is the bottleneck, diffusion models offer a compelling alternative to the standard AR paradigm.