Events

W375 Westgate Building
10:00AM

Modern users face an unprecedented challenge: critical knowledge is distributed across massive, heterogeneous, and constantly evolving data sources, from the web to scientific literature and structured databases. While search engines excel at retrieving relevant documents and large language models (LLMs) excel at language understanding and reasoning, neither alone can reliably support complex, multi-step information-seeking tasks that require searching, reasoning, and acting in concert. In this talk, I present my research on the science of information-seeking intelligence—a principled framework for building intelligent systems that can autonomously index large-scale data, reason over retrieved information, and take goal-directed actions in complex environments. I will describe three layers of contributions. First, I introduce architecture foundations for modeling heterogeneous, structure-rich data, showing how jointly representing text and structural signals leads to more effective indexing and understanding. Second, I establish the core capabilities required for information-seeking intelligence, including generative retrieval and robust reasoning over noisy or partially relevant evidence. Finally, I present agentic systems trained via reinforcement learning that learn to interleave reasoning, search, and tool use without labeled trajectories, enabling scalable and adaptive behavior in real-world information environments. Across these contributions, I demonstrate how integrating representation learning, retrieval, reasoning, and reinforcement learning yields systems that are both more capable and more interpretable. I conclude by outlining a research agenda toward general-purpose information-seeking agents, with implications for autonomous research assistants, scientific discovery, and reliable AI systems that broaden access to knowledge.

Additional Information:

Bowen Jin is a Ph.D. candidate in Computer Science at the University of Illinois at Urbana-Champaign, advised by Prof. Jiawei Han. His research focuses on the science of information-seeking intelligence, developing agentic systems that can search, reason, and act over large-scale, heterogeneous data by integrating foundation models, information retrieval, and reinforcement learning. Bowen has published extensively in top venues including NeurIPS, ICML, ICLR, KDD, ACL, SIGIR, and WWW, many of which are widely recognized. His research has been adopted in both academia and industry, including deployments at Google Cloud and widespread open-source use. He has received several honors, including the Apple PhD Fellowship, the Yunni & Maxine Pao Memorial Fellowship and NeurIPs top reviewer award.

 

Details...

W375 Westgate Building
10:00AM

AI systems often fail silently, producing confident but incorrect answers, brittle reasoning, or misleading scientific conclusions. This gap motivates a central question: What does it mean for AI to be reliable beyond accuracy? In this talk, I present a unified perspective on reliable AI, grounded in two foundational principles: uncertainty quantification and causal representation learning, and demonstrate how these ideas translate into practical advances in both language models and scientific discovery. I first discuss two uncertainty quantification frameworks for LLMs based on conformal prediction. The first enables uncertainty-aware prediction for API-only LLMs without logit access, producing compact prediction sets with guaranteed coverage. The second extends conformal factuality to multi-step reasoning, introducing a differentiable formulation that preserves formal guarantees while substantially improving claim retention, addressing hallucinations at scale. I then turn to causal representation learning, arguing that reliability also requires modeling how underlying mechanisms change, not just detecting uncertainty. I present TRACE, a causal framework that captures continuous transitions between mechanisms via identifiable mixtures of atomic causal components, enabling robust generalization to previously unseen regimes. Finally, I briefly highlight two applications in AI for science. In the first, retrieval-augmented LLMs are used to extract causal biomarker networks for Alzheimer’s disease. In the second, conformal prediction is used to enable reliable atom-level uncertainty estimates in molecular docking and the first rigorous comparison of reliability across docking methods.

Additional Information:

Lu Cheng is an assistant professor in Computer Science at the University of Illinois Chicago. Her research interests are broadly in AI and data science, with a focus on responsible and reliable AI, causal machine learning, and AI with science and engineering. She is the recipient of the NSF CAREER, PAKDD Best Paper Award, Google Faculty Research Award, Amazon Research Award, Cisco Research Faculty award, AAAI New Faculty Highlights, 2022 INNS Doctoral Dissertation Award (runner-up), 2021 ASU Engineering Dean's Dissertation Award, IBM Ph.D. Social Good Fellowship, among others. She co-authors two books: “Causal Inference and Machine Learning (Chinese)” and “Socially Responsible AI: Theories and Practices”.

Details...

W375 Westgate Building
10:00AM

ABSTRACT Foundation models are rapidly becoming capable assistants for knowledge work, but their deployment in real settings is limited by three gaps: they do not transfer reliably across environments, their internal reasoning is opaque, and their behavior is hard to precisely control. In this talk, I argue that these limitations are not only about model size — they are fundamentally about whether learning captures and leverages the underlying structure of the data-generating process. I use causal thinking as a practical lens to model what is invariant, what changes, and what can be intervened on, and I further show how this leads to learning principles that improve trustworthiness. I will first present methods for learning invariant mechanisms from heterogeneous data, across domains and modalities, to enable reliable transfer and controllable generation. Next, I will show how structured concepts can be recovered even from seemingly unstructured data, by analyzing and improving self-supervised objectives (such as masking and diffusion) through hierarchical latent-variable models. These concept structures can then be used to interpret generative models and support targeted, multi-level edits. Finally, I connect these two threads to generalization beyond the training distribution. I will discuss natural conditions for extrapolation and a compositional generation framework that improves prompt following for novel concept combinations. I will conclude with a brief outlook on self-improving world models and AI-assisted scientific discovery.

Additional Information:

BIOGRAPHY

Lingjing Kong is a Ph.D. candidate in the Computer Science Department at Carnegie Mellon University. His research focuses on Causal AI for transferable, interpretable, and controllable systems, with an emphasis on understanding and exploiting the structure of real-world data to make foundation models actionable and more reliable. He develops identification principles and scalable algorithms for learning unified models from heterogeneous data, uncovering hierarchical concept structures in unstructured data (e.g., images and text), and generalizing beyond training support through compositionality and extrapolation. His work has appeared in top ML venues including ICML, NeurIPS, CVPR, ICLR, and EMNLP, and has been prototyped and applied in industry through research internships (including Amazon).

Details...

Hammond Building, Room 220
1:25 – 2:35 p.m.

Details...

W375 Westgate Building
10:00AM

ABSTRACT Spectral methods such as principal component analysis and canonical correlation analysis are foundational tools in statistics for extracting low-dimensional structure from high-dimensional data, and their streaming and stochastic variants play a central role in modern large-scale inference. At the same time, many emerging problems in science and engineering require estimating the spectral structure of infinite-dimensional operators on function spaces over high-dimensional domains, including differential operators and conditional expectation operators. In these regimes, classical Rayleigh-quotient-based formulations, which rely on explicit orthogonality constraints, often become unstable and difficult to scale, particularly in parametric or learned settings. In this talk, I present a reformulation of spectral estimation based on unconstrained variational objectives that implicitly encode ordered spectral structure through nested optimization. This approach removes the need for explicit orthogonality constraints while preserving spectral ordering, and is naturally compatible with large-scale stochastic optimization. The resulting framework provides a unified and computationally efficient approach to learning structured representations in high-dimensional problems, without requiring explicit orthogonalization or projection steps. I illustrate this framework in three domains: eigenvalue problems for differential operators such as the Schrödinger equation in quantum chemistry, spectral (Koopman) analysis of nonlinear dynamical systems including molecular dynamics, and structured representation learning with deep neural networks. Together, these examples demonstrate how principled reformulations of classical spectral objectives can enable scalable, reliable AI methods for scientific discovery.

Additional Information:

BIOGRAPHY

Jongha (Jon) Ryu is a postdoctoral associate in the Department of Electrical Engineering and Computer Science at the Massachusetts Institute of Technology. He received his Ph.D. in Electrical and Computer Engineering from the University of California San Diego. His research develops statistical and mathematical foundations for artificial intelligence, with a focus on spectral learning, generative modeling, and uncertainty quantification for scientific and engineering systems.

Details...