Title: Generalizable Scientific Law Discovery in LLM Agents

Abstract: In this talk, we discuss the recent progress of scientific law discovery with LLM. Having LLM to discover new scientific law is an interesting topic. However, there are several limitations with the existing dataset to benchmark the LLM. We address this issue by a new benchmarking dataset.

Bio: Dr Charles CHEUNG is currently Deputy Director in NVIDIA AI Technology Center HK and Korea. He works on different projects with universities including computer vision, natural language processing, applied mathematics, foundation of machine learning, etc. Prior joining NVIDIA, he led a research team to develop AI solutions for advanced manufacturing companies. He received BSc and PhD degree in applied and computational mathematics from HKBU mathematics department in 2011 and 2016, respectively. His research focuses on computational science, partial differential equation, computer vision and artificial intelligence. He is currently an adjunct associate professor in the department of mathematics at HKBU.

Title: Visualizing and Understanding Multimodal Interactions

Abstract: Building AI systems that learn from multiple sensory inputs such as text, speech, video, real-world sensors, wearable devices, and medical data holds great promise for impact in many scientific areas with practical benefits, such as in supporting human health and well-being, enabling multimedia content processing, and enhancing real-world autonomous agents. Key to these multimodal models is the concept of multimodal interactions -- the study of how modalities interact with each other to give rise to new information for a task. This talk will present new theoretical frameworks to quantify and visualize the interactions present in multimodal datasets and those learned by different models. Their quantification enables users to better understand their datasets and design principled approaches to learn these interactions, especially for scientific applications where interpretability and robustness are crucial.

Bio: Paul Liang is an Assistant Professor at the MIT Media Lab and MIT EECS, where he directs the Multisensory Intelligence research group which advances the foundations of multisensory AI to enhance the human experience. He is a recipient of the MIT Sony Career Development Chair Professorship, Forbes 30 under 30 in Science, CMU School of Computer Science Distinguished Dissertation Award, Siebel Scholars Award, Waibel Presidential Fellowship, Facebook PhD Fellowship, Center for ML and Health Fellowship, Rising Stars in Data Science, and 4 best paper awards. Outside of research, he received the Alan J. Perlis Graduate Student Teaching Award for developing new courses on multimodal machine learning.

Title: Explainability Matters

Abstract: Explainability, the capability of an artificial intelligence system (AIS) to explain its outcomes in a manner that is comprehensible to human beings at an acceptable level, has been deemed essential for critical sectors, such as healthcare. Is it really the case? In this perspective, we consider two extreme cases, ``Oracle'' (without explainability) versus ``AI Colleague'' (with explainability) for a thorough analysis. We discuss how the level of automation and explainability of AIS can affect the determination of liability among the medical practitioner/facility and manufacturer of AIS. We argue that explainability plays a crucial role in setting a responsibility framework in healthcare, from a legal standpoint, to shape the behavior of all involved parties and mitigate the risk of potential defensive medicine practices.

Bio: Andrea Bertolini is Associate Professor of Private Law at the Dirpolis Institute, Scuola Superiore Sant’Anna, and adjunct professor in private law at the University of Pisa. He is the director and scientific coordinator of the Centre of Excellence on the Regulation of Robotics and AI (EURA, www.eura.satannapisa.it), previously funded by the European Commission through the Jean Monnet Action. His published research ranges from private law (contracts, torts, law of obligations) to regulation of robotics and AI and technoethics, with a comparative and law and economics approach, and a focus on alternative liability models, product safety regulation and certification, insurance and risk management, human machine interaction, user manipulation and deception. Dr. Bertolini consults national, International, and European policy makers both with written studies and in person hearings, on a regular basis on issues of robotics and AI regulation. He holds a joint degree from the Scuola Superiore Sant'Anna and the University of Pisa, a Ph.D in private law from the Scuola Superiore Sant’Anna, as well as a LL.M. from Yale Law School. He is an attorney licensed to practice in Italy and New York.

Title: Trustable XAI for Healthcare

Abstract: This talk will outline the value of explainable AI for clinical applications and the inherent challenges in developing these methods, including the need to balance explanation quality, model performance, and usefulness. To address these challenges, the first part of the presentation centers on rigorous frameworks for evaluating interpretability in the context of time-series data. The second part focuses on moving beyond model characterization toward actionable, concept-driven explanations.

Bio: Hugues Turbé is a researcher in trustworthy and explainable AI for healthcare applications. His work focuses on two key barriers to clinical AI adoption: i) leveraging real-world clinical data, with an important focus on cardiovascular data, and ii) developing explainable models.

Title: Language Models Do Not Embed Numbers Continuously

Abstract: Recent research has extensively studied how large language models manipulate integers in specific arithmetic tasks, and on a more fundamental level, how they represent numeric values. These previous works have found that language model embeddings can be used to reconstruct the original values, however, they do not evaluate whether language models actually model continuous values as continuous. Using expected properties of the embedding space, including linear reconstruction and principal component analysis, we show that language models not only represent numeric spaces as non-continuous but also introduce significant noise. Using models from three major providers (OpenAI, Google Gemini and Voyage AI), we show that while reconstruction is possible with high fidelity (\(R^2 \geq 0.95\)), principal components only explain a minor share of variation within the embedding space. This indicates that many components within the embedding space are orthogonal to the simple numeric input space. Further, both linear reconstruction and explained variance suffer with increasing decimal precision, despite the ordinal nature of the input space being fundamentally unchanged. The findings of this work therefore have implications for the many areas where embedding models are used, in-particular where high numerical precision, large magnitudes or mixed-sign values are common.

Bio: Alex O. Davies is an EPSRC Doctoral Impact Fellow at the University of Bristol, where he completed his PhD in the Interactive Artificial Intelligence Centre for Doctoral Training in November 2025. With a background in physics at Durham University, his research addresses fundamental challenges in generating and representing graph-structured data, developing generation frameworks and methods for learning unified representations across diverse graph domains. His doctoral thesis has been nominated for the University of Bristol's Outstanding Thesis Award. At AAAI 2026, he presents "Language Models Do Not Embed Numbers Continuously", showing that while embedding models preserve numerical information for reconstruction, they do not represent numbers as truly continuous values, with the majority of embedding dimensions encoding information orthogonal to numerical content.

Title: What is Your Force Field Really Learning? Gaining Scientific Intuition with A Dual-Level Explainability Framework

Abstract: Machine learning force fields (MLFFs) now rival quantum methods in accuracy, yet their internal logic remains opaque—a critical barrier to both trust and scientific discovery. We introduce DUAL-X, a Dual-Level Explainability Framework that bridges model reasoning with human understanding. DUAL-X unites two complementary perspectives: a model-centric level identifying where in the atomic structure the model focuses its attention, and a human-centric level revealing what physically meaningful interactions it prioritizes. Implemented with Grad-CAM for spatial attribution and SHAP-on-SOAP for physical interpretation, DUAL-X provides a general, human-in-the-loop paradigm for interpretable scientific AI. Applied to dopant migration—a challenging task for MLFFs—DUAL-X reveals that training strategy governs the emergence of chemical intuition. Multi-temperature fine-tuned models exhibit over \(10^2 \times\) stronger selectivity for Cr–Cr \(f\)-type (\(l=3\)) angular correlations than scratch-trained ones, emphasizing complex 3D coordination motifs essential for accuracy. Models with sharper Grad-CAM focus also display coherent SHAP importance for dopant clustering, revealing consistent internal reasoning across scales. By aligning model logic with human physical knowledge, DUAL-X transforms opaque predictors into interpretable scientific partners—advancing trustworthy, explainable, and insight-driven AI for materials discovery.

Bio: Yi Cao is a PhD candidate in Chemical and Biomolecular Engineering at Johns Hopkins University, advised by Prof. Paulette Clancy. Her research focuses on integrating AI with molecular simulations to advance data-efficient, reliable, and closed-loop discovery of next-generation materials.