Title: Statistics-Powered AI

Time: 10:30am 

Venue: Room 301, 3/F, Run Run Shaw Building

Speaker(s): Prof. Chengchun Shi

Remark(s): 

Abstract

We have definitely entered an era of generative artificial intelligence (AI), where large language models (LLMs) are increasingly reshaping our daily lives. Their impact is everywhere -- from education and academia to professional work and everyday life. In this talk, I will present two recent NeurIPS papers on statistics-powered AI, focusing on how statistical methodologies can enhance AI's performance in (1) aligning LLM's model outputs with human feedback, and (2) detecting LLM-generated content with rigorous guarantees. Open-source Python implementations are available at https://github.com/Mamba413/AdaDetectGPT and https://github.com/DRPO4LLM/DRPO4LLM.

Title: Deep Model Fusion

Time: 10:00am 

Venue: HW312, Haking Wong Building

Speaker(s): Prof. Dacheng Tao

Remark(s): 

Abstract

In recent years, we have witnessed a profound transformation in the learning paradigm of deep neural networks, especially in the applications of large language models and other foundation models. While conventional deep learning methodologies maintain their significance, they are now augmented by emergent model-centric approaches such as transferring knowledge, editing models, fusing models, or leveraging unlabelled data to tune models. Among these advances, deep model fusion techniques have demonstrated particular efficacy in boosting model performance, accelerating training, and mitigating the dependency on annotated datasets. Nevertheless, substantial challenges persist in the research and application of effective fusion methodologies and their scalability to large-scale foundation models. In this talk, we systematically present the recent advances in deep model fusion techniques. We provide a comprehensive taxonomical framework for categorizing existing model fusion approaches, and introduce our recent developments, including (1) weight learning-based model fusion and data-adaptive MoE upscaling, (2) subspace learning approaches to model fusion, and (3) enhanced multi-task model fusion incorporating pre- and post-finetuning to minimize representation bias between the merged model and task-specific models.

About the speaker

Prof. Dacheng Tao is the Distinguished University Professor and the Inaugural Director of the Generative AI Lab in the College of Computing and Data Science at Nanyang Technological University. He was an Australian Laureate Fellow and the founding director of the Sydney AI Centre at the University of Sydney, the inaugural director of JD Explore Academy and senior vice president at JD.com, and the chief AI scientist at UBTECH Robotics. He mainly applies statistics and mathematics to artificial intelligence, and his research is detailed in one monograph and over 300 publications. His publications have been cited over 160K times and he has an h-index 180+ in Google Scholar. He received the 2015 and 2020 Australian Eureka Prize, the 2018 IEEE ICDM Research Contributions Award, 2020 research super star by The Australian, the 2019 Diploma of The Polish Neural Network Society, and the 2021 IEEE Computer Society McCluskey Technical Achievement Award. He is a Fellow of the Australian Academy of Science, ACM and IEEE.

Title: The Wild Robot: A Journey Toward Long-Horizon Agentic Intelligence

Time: 11:00am 

Venue: HW312, Haking Wong Building

Speaker(s): Prof. Ivor W Tsang

Remark(s): 

Abstract

Long-horizon planning in robotic manipulation demands translating abstract goals into precise, executable actions while maintaining spatial, temporal, and physical consistency. However,
 language model-based planners often fail to handle extended task decomposition, constraint satisfaction, and adaptive recovery from errors. We present The Wild Robot, a framework for autonomous, feedback-driven reasoning that constructs and refines symbolic instruction graphs to guide code generation in robotic tasks. The system dynamically decomposes complex goals into coherent subtasks and generates executable control programs accordingly. When execution failures occur, it analyzes environmental feedback to induce and propagate new constraints, enabling targeted refinement without restarting the planning process. This structured, interpretable approach fosters resilience, adaptability, and transparency, significantly enhancing performance in long-horizon and constraint-sensitive robotic benchmarks. The Wild Robot represents a step toward truly agentic intelligence capable of robust, self-correcting decision-making in complex, real-world manipulation scenarios.

About the speaker

Professor Ivor W. Tsang is the Director of the A*STAR Centre for Frontier AI Research (CFAR) and an Adjunct Professor at the College of Computing and Data Science, NTU, Singapore. Since January 2022, he has led Singapore’s national initiative on Trustworthy Foundation Models under the National Multimodal LLM Programme. He also drives research on Agentic World Models and oversees major national initiatives such as the AI Singapore Materials Design Grand Challenge and the Maritime AI Programme. Under his leadership, CFAR has secured over S$23 million in strategic research funding, strengthening Singapore’s frontier AI ecosystem.

His research spans transfer learning, deep generative models, and big data analytics involving ultra–high-dimensional data. His influential work has earned international recognition, including the ARC Future Fellowship (2013), the ICCM Best Paper Award (2019), and recognition as the AI 2000 AAAI/IJCAI Most Influential Scholar in Australia (2020). An IEEE Fellow, he has made distinguished contributions to large-scale and transfer learning. He also serves on editorial boards of leading AI journals and top AI conference committees.

Title: From Statistical Physics to Privacy: Phase Transitions, Sampling and Differential Privacy

Time: 10:30am 

Venue: CB 308

Speaker(s): Prof. Jingcheng Liu

Remark(s): 

Abstract

Classical interacting particle systems studied in statistical physics are intimately connected to constraint satisfaction problems studied in computer science. Much progress has been made in the recent decade in understanding the "computational" phase transition in Gibbs sampling, where a sharp transition in computational tractability coincides precisely with the underlying physical phase transition in many models. I'll give a survey of my research along this line, and also highlight how these developments also lead to new perspectives and applications in differentially private optimizations.

About the speaker

Jingcheng Liu is an Associate Professor in the Theory Group of the School of Computer Science at Nanjing University. He is broadly interested in theoretical computer science, which includes randomized algorithms, computational phase transition, and differential privacy. Before that, he completed undergrad at SJTU (ACM Honors class) and PhD at UC Berkeley, and he was a Wally Baer and Jeri Weiss postdoctoral scholar at Caltech.