Abstract: The emergence of large language models (LLMs) has introduced a new paradigm in data modeling. These models replace specialized models designed for individual tasks with unified models that are effective across a broad range of problems. In scientific domains, this shift not only transforms approaches to handling natural language tasks (e.g., scientific papers) but also suggests new methods for dealing with other data types (e.g., molecules, proteins, pathology images). In many fields, LLM has already shown great potential to accelerate scientific discovery. In this talk, I will present our recent work on LLMs, especially in the context of science and engineering research.
Bio: Wei Wang is the Leonard Kleinrock Chair Professor in Computer Science and Computational Medicine at University of California, Los Angeles and the director of the Scalable Analytics Institute (ScAi). She is also a member of the UCLA Jonsson Comprehensive Cancer Center, Institute for Quantitative and Computational Biology, and Bioinformatics Interdepartmental Graduate Program. Dr. Wang received the IBM Invention Achievement Awards in 2000 and 2001. She was the recipient of an NSF Faculty Early Career Development (CAREER) Award and was named a Microsoft Research New Faculty Fellow in 2005. She was honored with the 2007 Phillip and Ruth Hettleman Prize for Artistic and Scholarly Achievement. She was recognized with an IEEE ICDM Outstanding Service Award in 2012, an Okawa Foundation Research Award in 2013, and an ACM SIGKDD Service Award in 2016.
Bio: Marinka Zitnik is an Associate Professor at Harvard in the Department of Biomedical Informatics. Dr. Zitnik is Associate Faculty at the Kempner Institue for the Study of Natural and Artificial Intelligence, Broad Institute of MIT and Harvard, and Harvard Data Science. Dr. Zitnik investigates foundations of AI to enhance scientific discovery and facilitate individualized diagnosis and treatment in medicine. Her algorithms and methods have had a tangible impact, which has garnered interests of government, academic, and industry researchers and has put new tools in the hands of practitioners. Some of her methods are used by major biomedical institutions, including Baylor College of Medicine, Karolinska Institute, Stanford Medical School, and Massachusetts General Hospital.
Abstract: AI has tremendous potential for accelerating science---helping users access the literature, execute experiments, analyze data, and more. In this talk, I will present Ai2 ScholarQA, a long-form scientific question answering system recently released by the Allen Institute for AI as a first step toward building a comprehensive AI-powered scientific assistant. While the system and others like it are often helpful, I’ll discuss how realizing AI’s full potential for long-form QA will require tackling important unsolved challenges, including evaluation, obtaining supervision at scale, and new methods for providing personalized and proactive assistance.
Bio: Doug Downey is a Director of Semantic Scholar Research at the Allen Institute for AI (AI2). He is currently on leave from Northwestern University, where he is an Associate Professor of Computer Science. His research focuses on information extraction, natural language processing, and machine learning. Outside of work, he enjoys spending time with family, exploring the outdoors, and watching movies.
Abstract: Three-dimensional imaging in astronomy seems counterintuitive: we typically assume 3D information comes from multiple views (e.g., CT in medical imaging), and in most of astronomy (outside of our galactic neighborhood), we only have access to a single view point and cannot rely on parallax cues for 3D reconstructions. Nonetheless, in many astronomical systems we find hidden 3D cues embedded in the time or spectral axes. While these signatures are not as strong as multiple views, in principle, they enable recovering some 3D information (or “2.5D”). Cosmological redshift is a classic example of getting 3D positions at the cosmic scale. Doppler shit gives information about the projected velocities of disks. Fast temporal evolution of an orbiting system enables reconstructing the 3D structure through observations over time. Modern telescope resolutions are now able to resolve many of those axes (spectral, temporal, and polarimetric) to unprecedented accuracy, which calls for new algorithms for 3D reconstructions to be developed. In the last few years, neural radiance fields (NeRFs) have completely changed the landscape of 3D imaging within the computer vision and graphics communities. One of the reasons behind their success is the simplicity of constraining a neural representation through ray tracing. The focus of much of the research within the vision community is on developing better representations with efficiency and scalability for complex 3D scenes in mind. In astronomy (and more broadly in the natural sciences), the scenes are not very complex; nonetheless, the radiative transfer physics are often a bottleneck in terms of complexity and runtime, preventing the solution of inverse problems. In this talk, I will show how neural fields constrained by ray tracing and complex physics enable 3D imaging in astronomical environments: black holes, protoplanetary disks, and dark matter.
Bio: Aviad Levis is an assistant professor in the Departments of Computer Science and Astronomy and Astrophysics at the University of Toronto. He is an associated faculty member at the Dunlap Institute for Astronomy and Astrophysics. His research focuses on scientific computational imaging and AI for science. Prior to that, he was a postdoctoral scholar in the Department of Computing and Mathematics at Caltech, supported by the Zuckerman and Viterbi postdoctoral fellowships, working with Katie Bouman on imaging the galactic center black hole as part of the Event Horizon Telescope collaboration. He received his Ph.D. (2020) from the Technion and B.Sc. (2013) from Ben-Gurion University. His Ph.D. thesis into tomography of clouds has paved the way for an ERC-funded space mission (CloudCT) led by his Ph.D. advisor Yoav Schechner.
Abstract: This talk explores how Artificial Intelligence (AI) can enhance three critical phases of the scientific research lifecycle. First, we examine how AI transforms scientific writing from a time-intensive drafting process to a collaborative refinement approach, mirroring the professor-student relationship while addressing concerns about language barriers and time constraints. Second, we tackle the growing challenge of literature surveys in an era of exponential publication growth, proposing specialized AI research assistants that can help identify relevant works, synthesize knowledge across disciplines, and uncover "sleeping beauties" in scientific literature. Finally, we address the peer review crisis by demonstrating how AI can streamline meta-reviews through effective summarization, allowing human reviewers to focus on conflict resolution and nuanced judgment. Throughout, we emphasize a human-AI collaboration that maintains research integrity while increasing efficiency and accessibility.
Bio: Jinho Choi is an associate professor of Computer Science, Quantitative Theory and Methods, and Linguistics at Emory University. He is the founder and the director of the Natural Language Processing Research Laboratory at Emory University. He has presented many state-of-the-art NLP models that automatically derive various linguistic structures from plain text. These models are publicly available in the NLP Toolkit called ELIT. He has also led the Character Mining project and introduced novel machine comprehension tasks for explicit and implicit understanding in multiparty dialogue. For the application side, he has developed innovative Biomedical NLP models by collaborating with several medical fields such as radiology, neurology, transplant, and nursing. His latest research focuses on building the conversational AI-based chatbot called Emora that aims to be a daily companion of everyone's life. With Emora, his team won the 1st-place at the Alexa Prize Socialbot Grand Challenge 3 that came with $500,000 cash award.
Abstract: A grand challenge in artificial intelligence is developing systems capable of open-ended learning and autonomous scientific discovery. This talk highlights recent progress toward fully autonomous AI-driven science, beginning with The AI Scientist, which automates the entire scientific process, from hypothesis generation through experimentation. We then discuss Automated Design of Agentic Systems (ADAS), demonstrating how meta-agents autonomously design and optimize large language model architectures, achieving strong performance in complex reasoning and problem-solving tasks. Lastly, we introduce Automated Capability Discovery (ACD), a method for systematically evaluating the extensive capabilities of foundation models. ACD employs one model to autonomously generate open-ended tasks to probe another model's abilities, uncovering numerous surprising capabilities and limitations in models such as GPT, Claude, and Llama. The talk concludes by exploring future directions and the potential of autonomous scientific exploration.
Bio: Cong Lu is a Research Scientist at Google DeepMind on the Open-Endedness team. He is interested in developing autonomous agents that are safe, curious, and capable of open-ended learning—especially given recent advances in foundation models and deep reinforcement learning. Previously, he was a postdoctoral research and teaching fellow at the University of British Columbia and the Vector Institute, supervised by Prof. Jeff Clune. During this time, they developed The AI Scientist, the first agent to automate the entire scientific process (from forming hypotheses and conducting experiments to visualizing results, writing a paper, and reviewing it). They work has been featured by Science News, Nature News, VentureBeat, Ars Technica, WIRED, IEEE Spectrum, Forbes, and Air Street Press. He also discussed how AI is transforming science on CBC's Quirks & Quarks.