2026 Rising Stars
Autumn 2026 Cohort
-
Marwa Abdulhai
PhD Candidate, University of California, Berkeley -

Amanda Bertsch
PhD Candidate, Carnegie Mellon University -

Luke Bhan
PhD Candidate, University of California, San Diego -

Jacob Carlson
PhD Candidate, Harvard University -

Souradip Chakraborty
Postdoctoral Scholar, Massachusetts Institute of Technology -

Myra Cheng
PhD Candidate, Stanford University -

Gillian Chu
PhD Candidate, Princeton University -

Madison Coots
PhD Candidate, Harvard University -

Riccardo De Santi
PhD Candidate, ETH Zurich -

Zhe Fu
PhD Candidate, University of California, Berkeley -
Katherine Keegan
PhD Candidate, Emory University -

Tzu-Sheng Kuo
PhD Candidate, Carnegie Mellon University -

Nicolas Lanzetti
Postdoctoral Scholar, California Institute of Technology -

Junu Lee
PhD Candidate, University of Pennsylvania -

Shicheng Liu
PhD Candidate, Stanford University -

Xiao Liu
Postdoctoral Scholar, University of Chicago -
Cecile Meier-Scherling
PhD Candidate, Brown University -

Ashley Quiterio
Postdoctoral Scholar, Northwestern University -

Vaidehi Srinivas
Postdoctoral Scholar, Northwestern University -

Huili Sun
Postdoctoral Scholar, Yale University -

Qiaosen Wang
PhD Candidate, University of Chicago -

Qingsong Wang
Postdoctoral Scholar, University of California, San Diego -

Bohan Wu
PhD Candidate, Columbia University -

Xindi Wu
PhD Candidate, Princeton University -

Yifan Wu
Postdoctoral Scholar, Microsoft Research -

Ge Zhang
PhD Candidate, University of Michigan -

Haoran Zhang
PhD Candidate, Massachusetts Institute of Technology -

Xuandong Zhao
Postdoctoral Scholar, University of California, Berkeley
Bio: Marwa Abdulhai is a PhD candidate at UC Berkeley advised by Sergey Levine and Natasha Jaques (UW). Her research focuses on building AI agents for real-world dialogue settings. This includes improving LLM capabilities with multi-turn reinforcement learning, understanding how to protect against deception in AI systems, and exploring how AI can serve as a useful tool for social science research. Her research is supported by the Quad Fellowship, AI Policy Hub, Open AI Research, & Cooperative AI Fellowship. She will be joining Princeton as a Presidential Postdoctoral fellow in Fall 2026, working with Ben Eysenbach, Peter Henderson, & Tom Griffiths.
Talk Title: Multi-Turn Reinforcement Learning for Real-World Dialogue Agents
Abstract: Large Language Models (LLMs) are increasingly deployed as agents in real world interactive settings such as teaching, negotiation, and mental health. However, standard LLMs are optimized for single-turn instruction following, and rarely pursue goal-directed behavior across extended interactions. In this talk, I introduce the tools needed to build such agents. I present multi-turn RL as a framework for training LLMs to plan, gather information, and optimize goals in dialogue. Next, I will argue that consistent human simulations are essential as off-the-shelf LLMs often drift from their assigned personas, contradict earlier statements, or abandon role-appropriate behavior. I show how these tools come together in strategic dialogue tasks such as negotiation, where multi-turn RL enables safe outcomes and the ability for agents to reason hierarchically, generalize across unseen tasks, and adapt to diverse opponents towards cooperative, value-creating outcomes.
Bio: Amanda Bertsch is a final year PhD candidate in the Language Technologies Institute at Carnegie Mellon University, advised by Matt Gormley and Graham Neubig. She is also a student researcher at the Allen Institute for AI. Her research is in natural language processing and focuses on better ways to computationally reason over large quantities of knowledge, particularly by increasing the context length and in-context reasoning abilities of language models. She is a recipient of the NSF Graduate Research Fellowship and a core organizer for Queer in AI. Her work has been published in top NLP and ML venues and won the SAC Award for Language Modeling at NAACL 2025.
Talk Title: Language models for massive-scale information synthesis
Abstract: Large language models are powerful tools. However, they struggle to perform complex tasks that require synthesizing non-obvious insights: for instance, to identify which details in a description are crucial or trivial, see how documents in a corpus are subtly changing in style over time, or find parts of the scientific literature across fields that share useful similarities. In this talk, I will discuss my research on language models that can reason over vast quantities of text. I will first discuss how we enable this capability by extending the context lengths of language models. I will present our work, Unlimiformer, which presents one option for extending language models to arbitrary context lengths. Then, I will show how inference-time capabilities of long context models can be unlocked by architectural improvements, and how these capabilities in turn can inform more efficient architecture design, by discussing the problem of long in-context learning. Finally, I consider what capabilities necessary for reasoning remain underdeveloped in modern language models. Using corpus reasoning as a motivating example, I will show how current methods fail to generalize to corpus reasoning tasks, even when these tasks involve relatively short inputs. I will present our ongoing work on designing new architectures that enable language models to perform corpus reasoning. I conclude by discussing the future of language models as tools for interactive knowledge discovery.
Bio: Luke Bhan is a final-year Ph.D. candidate at UC San Diego and a Department of Energy Computational Science Graduate Fellow. His research combines artificial intelligence, control theory, and optimization to develop learned autonomous decision-makers that are safe and reliable for our societies largest physical systems. He has held research internships at Microsoft, Amazon, and Lawrence Berkeley National Laboratory, working on power grid foundation models, forecasting, and numerical optimization. Prior to his Ph.D., Luke completed an accelerated BS/MS program in computer science, physics, and math at Vanderbilt University.
Talk Title: Neural Operators in Closed Loop Feedback Design: From Theory to Practice
Abstract: Machine learning is often viewed, sometimes justifiably, as a challenge to model-based sciences, including physics-driven modeling and classical control theory. Rather than positioning learning as a replacement for first-principles models, this talk shows how operator learning can be integrated into feedback control design, combining the strengths of both: controllers that are data-efficient and expressive, yet remain real-time computable and provably stabilizing. I develop this through two case studies: nonlinear systems with input delays, as arise in almost all physical systems that require sensing and communication, and boundary control of the PDEs governing phenomena ranging from highway traffic to data center thermodynamics. In each, learned operators replace one part of an otherwise computationally intractable model based design enabling real-time deployment 1000 times faster than numerical solvers while preserving the analytical structure needed to certify closed-loop stability. We showcase two specific applications of these designs where robot manipulators under camera delays complete dynamic tasks such as table tennis and in the context of thermal management showcasing that operator learned PDE solvers can speed up data center cooling while outperforming reduced-order models on hotspot aware performance. Lastly, we conclude by summarizing that the approach is not specific to delays or PDEs: it is a blueprint for combining learned components with control design wherever applications demand both certificates and scalable compute. Accordingly, we outline several future directions across applications such as power systems, humanoid robots, and water cooling designs.
Bio: Hi! I’m Jake, a fifth-year PhD candidate in the Department of Economics at Harvard University. I am an econometrician with research that spans topics in high-dimensional inference, semiparametric inference, causal inference, and AI/ML methods. I am particularly interested in developing new machine learning tools and statistical methods for the analysis of unstructured data, such as text, image, and audio data. I graduated from Yale College in 2020, where I pursued a double major in Statistics and Data Science (S&DS) and Ethics, Politics, and Economics (EP&E).
Talk Title: Interpretable Discovery from Unstructured Data: A High-Dimensional Approach
Abstract: We propose an automatic, general-purpose framework for making discoveries from unstructured data (e.g., text data from open-ended surveys of economic beliefs). The framework leverages recent methods from the literature on AI interpretability to transform unstructured datasets into high-dimensional, structured datasets of interpretable concept measurements; specifies concept-level parameters and null hypotheses based on this
transformed dataset; tests these hypotheses using algorithms validated by new results in high-dimensional multiple testing, producing a selected set (“discoveries”); and both generates and evaluates human-interpretable natural language descriptions of these discoveries. The proposed framework has few researcher degrees of freedom, is robust to data snooping, mitigates under-exploration, and facilitates fast and inexpensive sensitivity analysis and replication. We revisit applications to recent descriptive and causal analyses of unstructured data in empirical economics, and find this framework
is able to automatically replicate existing discoveries, add nuance to others, and make entirely new discoveries as well.
Bio: Souradip Chakraborty is a Postdoctoral Researcher at MIT CSAIL, working with Prof. Omar Khattab on the foundations of Pedagogical Reinforcement Learning for AI agents: learning paradigms that let agents improve through interaction, feedback, and long-horizon experience. His broader goal is to build continually learning agents that acquire new capabilities while remaining robust, efficient, and aligned with human objectives combining theoretical and practical insights.
He received his Ph.D. in Computer Science from the University of Maryland, College Park, advised by Prof. Furong Huang and Prof. Dinesh Manocha, where he received the department’s Certificate of Outstanding Achievement. During his doctoral studies, he collaborated closely with Prof. Mengdi Wang, Dr. Mohammad Ghavamzadeh, Prof. Aldo Pacchiano, and Dr. Alekh Agarwal on reinforcement learning and principled foundations of AI alignment, spanning post-training and inference-time optimization. His research experience also includes work at Google Research and JPMorgan Chase and, before his Ph.D., as a Research Scientist at Walmart Global Tech.
Souradip has published at leading venues including ICML, NeurIPS, ICLR, CVPR, ICRA, CoRL, NAACL, and TMLR, and holds multiple US patents in AI and machine learning. His work has been recognized through oral and spotlight presentations at ICML and AAAI, an Outstanding Paper Award at a NeurIPS 2022 workshop, and Outstanding Reviewer Awards at NeurIPS and AISTATS.
Talk Title: Pedagogical Reinforcement Learning: Teaching Models to Teach Themselves with Privileged Information
Abstract: Reinforcement learning uses privileged information to evaluate generated rollouts, leaving the discovery of successful trajectories entirely to the student’s blind exploration. This makes learning slow and highly sample-inefficient, particularly when rewards are sparse. A natural alternative is to use the privileged information as a learning signal from the contrast between tokens the student chose and the one the teachers preferred, as in On policy self distillation approaches. However, we find that such teachers often use the privileged information to cheat rather than reasoning from first principles making those approaches unstable for the student. To address this challenge, we introduce Pedagogical Reinforcement Learning, a framework that trains a self-teacher by optimizing a pedagogical reward which strongly penalizes cheating and rewards trajectories that are closely consistent with an unprivileged student. Further, to allow reasoning and multi-turn tool use for long horizon agentic tasks, we introduce Pedagogical Thinking, where the teacher’s private reasoning tokens are separated from the reasoning tokens the student is meant to follow. We show that Pedagogical RL achieves more than 30% relative improvement across all benchmarks in deep research tasks, coding, and mathematical reasoning.
Bio: Myra Cheng is a final-year PhD candidate in Computer Science at Stanford studying human-centered AI and natural language processing. She develops safe, trustworthy language technologies that promote human agency and wellbeing. Her research translates social science insights into large-scale computational methods for understanding and improving AI’s impacts on individuals and society, including pioneering work on AI sycophancy, overreliance, and anthropomorphism. Her work has been published in Science (cover story), Nature, and leading computer science venues. It has received widespread media coverage, including in The New York Times, The Washington Post, AP, NPR, TIME, and Scientific American. She has been named one of MIT Technology Review’s 35 Innovators Under 35 and is a Stanford Knight-Hennessy Scholar and an NSF Graduate Research Fellow.
Talk Title: Building AI for better societal outcomes
Abstract: Every month, more than a billion people interact with conversational AI systems. Large language models (LLMs) influence people’s beliefs and decisions; social interactions and relationships; and our collective societal well-being. I develop computational approaches to account for these broader human and societal outcomes in LLM development at three interconnected levels: AI model behavior, human-AI interaction dynamics, and societal risks. I identified the widespread issues of AI sycophancy, anthropomorphism, and overreliance, and developed interventions to address these issues, such as algorithms to optimize LLMs for outcomes beyond immediate user satisfaction, and representation-level methods to control the implicit assumptions encoded in model internals. My work shows how we can develop AI that promotes human agency and long-term well-being.
Bio: Gillian Chu is a PhD candidate in the Computer Science department at Princeton University, advised by Professor Ben Raphael. Her research focuses on the design of scalable statistical algorithms to reveal interpretable insights from biological data. Most recently, she has designed several probabilistic models and algorithms for single cell lineage tracing technologies, which measure the developmental process of a single cell becoming a multicellular tissue or organism. Gillian completed her master’s with Professor Tandy Warnow at the University of Illinois Urbana-Champaign, and her Bachelor’s with Associate Professor Priya Moorjani at the University of California, Berkeley. Her work has appeared at RECOMB and ISMB, and she is an NSF Graduate Research Fellow.
Talk Title: Reading history from cells: inference on tree-structured single-cell data
Abstract: Every cell in an organism can be traced, via cell divisions, back to a single fertilized egg. This history of cell divisions is summarized as a cell lineage tree, which forms the basis of scientists’ understanding of how organisms develop, and the baseline against which scientists understand disease. Recently, modern dynamic lineage tracing systems engineer cells to introduce heritable edits, or lineage mutations, as they divide, enabling scientists to trace cell lineage trees in more complex organisms like mice. Sequencing cells in such a population provides measurements of each cell’s lineage mutations and a noisy observation of cell state (e.g. gene expression). Using these lineage mutations to reconstruct the cell lineage tree is a major computational challenge, especially at the scale of tens of thousands of cells.
I will present a generative model of dynamic lineage tracing systems and an efficient inference algorithm scalable to modern datasets. I will demonstrate how the cell lineage trees inferred in cancerous tissues and developing organisms reveal the migration history of cancer cells and the spatial organization of cells over tissue development, opening up new modeling opportunities to improve our understanding of tissue development and cellular spatial organization. Finally, while cell state changes are usually difficult to study from a single snapshot of cells, I will show how lineage trees improve our models of when and how cells change state over development.
Bio: Madison Coots is a Ph.D. candidate in Public Policy at Harvard Kennedy School and a Stone Ph.D. Scholar in Inequality and Wealth Concentration. Her research uses computational methods to evaluate and build AI systems that improve organizational decisions, services, and public policy. Her work draws on techniques from data science, engineering, and social science and spans several domains, including health care, lending, criminal justice, and education. Previously, Madison earned a B.S. in Management Science and Engineering and an M.S. in Computer Science from Stanford University, and has worked as a data scientist in government, academia, and the private sector.
Talk Title: Algorithmic Bias in Lending: Evidence from a Fintech Audit
Abstract: Algorithmic lending has transformed the consumer credit landscape, with machine learning models commonly facilitating underwriting decisions. To comply with fairlending laws, these algorithms exclude legally protected characteristics, such as race and gender. Yet algorithmic underwriting can still inadvertently favor certain groups,prompting concerns about whether lending algorithms exhibit discriminatory behavior. Using proprietary loan-level data from a major U.S. fintech platform, we audit lending decisions across approximately 80,000 personal loans. We find that loans made to men and Black borrowers yielded lower profits than loans to other groups, suggesting that men and Black borrowers benefited from relatively favorable pricing. We trace these disparities to miscalibration in the platform’s underwriting model, which overestimates risk for women and underestimates risk for Black borrowers. We then show that one could correct this miscalibration—and the corresponding disparities—by including race and gender in underwriting models, illustrating a tension between competing notions of fairness.
Bio: Riccardo De Santi is a PhD student in Computer Science and an ETH AI Center Doctoral Fellow at ETH Zurich, advised by Andreas Krause, Niao He, and Kjell Jorner. His research centers on out-of-distribution generative discovery: from foundations of reinforcement-learning exploration, recognized with an ICML Outstanding Paper Award, to scalable methods for expanding flow and diffusion models beyond the data. More broadly, he aims to understand and develop the principles that underlie discovery systems at the intersection of generative modeling, optimization, and decision-making under uncertainty, ultimately contributing to the foundations of a science of generative discovery. Following a research visit at Caltech, he continues to collaborate with Yisong Yue and Frances H. Arnold’s lab on connecting generative discovery systems with wet-lab validation. He has also conducted research visits at the University of Oxford and Imperial College London.
Talk Title: Generative Discovery Beyond the Data
Abstract: Flow and diffusion models are powerful priors over scientific design spaces, but a prior is not yet a discovery system. Standard training and reward guidance largely concentrate probability within regions already represented by data. Scientific discovery poses a harder question: how can a generative model reach valid, useful regions that are rare, underrepresented, or absent from its training distribution, within limited experimental budget? I will present a research program on the foundations of generative discovery, progressing from distributional control of flow models to expansion beyond the data and natural-language goal adaptation. Flow Density Control starts from the premise that discovery requires controlling the full output distribution, including rare high-value outcomes and low-probability modes that average-reward fine-tuning can miss. Yet distributional control remains bounded by a model’s existing generative reach. This motivates the generable-set view: rather than evaluating a discovery model only by how well it matches observed data, we ask which valid regions it can reliably generate. Active Flow Expansion combines verifier feedback with active exploration in learned flow representations to enlarge this set with finite-sample coverage guarantees. To guide expansion toward high-level goals more sample-efficiently, LLM-to-Flow Distillation distills scientific knowledge and reasoning from frontier LLMs into domain-specific generators, enabling adaptation to natural-language discovery goals without task-specific reward proxies. With applications spanning molecular, peptide, and protein design, these developments connect mathematical foundations and scalable methods toward systems for sample-efficient generative discovery in the sciences.
Bio: Zhe Fu is a Stanford Energy Fellow co-hosted by Prof. Eric Darve and Marco Pavone. Her research centers on complex systems where a few controllable agents shape a large population of uncontrolled ones under a shared physical constraint, with field data as the arbiter. She builds learning-enabled control for mixed human–autonomy systems, where physics provides the inductive bias: control and optimization for guarantees, scientific machine learning for models that respect conservation laws, and field data decides what is true. Her work covers the full research arc, ranging from physics-informed neural PDE solvers to a 100-automated-vehicle deployment on I-24 highway, the largest traffic-smoothing experiment to date, with industry partners including Nissan, GM, and Toyota. She aims to extend this framework to agentic AI and autonomous scientific discovery.
Zhe earned her Ph.D. in Transportation Engineering and M.S. in EECS from UC Berkeley, advised by Prof. Alexandre Bayen and affiliated with the BAIR Lab. She is a Rising Star in Data Science (UChicago), EECS (MIT), Cyber-Physical Systems (NSF), and Mechanical Engineering (CMU); a 2025 Eno Fellow; and ITS Berkeley’s 2026 Outstanding Ph.D. Student. She won first place in the 2023 INFORMS Poster Competition and was runner-up in the 2025 Berkeley Grad Slam.
Beyond research, Zhe initiated the IEEE Technical Committee on Automated Mobility in Mixed Traffic and has mentored over 25 students, earning EDGE in Tech’s Athena Award and Berkeley’s Dean’s Award for Inclusive Excellence. Learn more at fu-zhe.com.
Talk Title: From Learning Dynamics to Shaping Them: Physics-Grounded Data Science for Real-World Systems
Abstract: Black-box machine learning models often predict well within their training distribution, yet ignore the physical structure of the systems they describe: conservation of mass and momentum, wave and shock formation, and the fact that deployment itself changes the data such models later see. My research develops physics-grounded data science methods that preserve this structure while learning from data, and that extend from prediction into safe, closed-loop decision-making.The methodological anchor is the Neural Finite Volume Method, which embeds a trainable neural flux inside a conservative finite-volume update for nonlinear hyperbolic conservation laws. The framework supports supervised training on entropy solutions and unsupervised training from a weak-form PDE residual, and outperforms classical numerical schemes and standard learning architectures across multiple flux functions and heterogeneous real-world spatiotemporal datasets.Learning structure is not sufficient once a model is used to intervene. I extend these methods into closed-loop decision-making: kernel-based speed planning and imitation learning that convert learned dynamics into real-time control, and Pareto control barrier functions that certify safety under input constraints. I validate this pipeline in CIRCLES, the largest scientific traffic field experiment to date with 100 automated vehicles operating on a live interstate highway, where deployed control produced measurable fuel savings under real sensing noise, communication delay, and human-driver adaptation.My future research extends this structure-aware, closed-loop approach toward scientific foundation models for dynamical systems and toward reliable learning under feedback and distribution shift, with applications spanning energy systems, autonomy, and other cyber-physical infrastructure beyond transportation.
Bio: I am a PhD student and DOE Computational Science Graduate Fellow at Emory University, where I am fortunate to be advised by Lars Ruthotto. My research focuses on adapting state-of-the-art generative modeling paradigms to equality-constrained settings, making them more amenable to scientific practitioner adoption. I have worked previously in both applied and theoretical capacities at Berkeley Lab, RIKEN Center for Computational Science, and Frontier Development Lab. I received my undergraduate degree in Applied Mathematics at age 18 from Mary Baldwin University in 2022 and am expected to receive my PhD in Computational Mathematics in 2027.
Talk Title: Techniques in Equality-Constrained Generative Modeling
Abstract: Generative modeling techniques such as diffusion models, flow matching, and normalizing flows have found massive success in a variety of commercial applications. However, their utility in scientific applications is often hindered by the challenges of incorporating hard constraints, such as those arising from domain knowledge or physical laws. In this talk, I will provide an overview of continuous-time generative modeling techniques, including score-based diffusion models and normalizing flows, and describe some common pathologies they face when modeling equality-constrained data. I will then discuss some of my recent and ongoing work in hard-constrained generative modeling, with an emphasis on combating or quantifying the probability measure degeneracy imposed by strict equality constraints.
Bio: Tzu-Sheng Kuo is a final-year PhD candidate at Carnegie Mellon University. As a human-computer interaction (HCI) researcher, he creates interactive systems that empower people to harness their collective intelligence for AI alignment. Conversely, his systems leverage AI to catalyze human collaboration, creating a virtuous cycle where collective human intelligence and artificial intelligence strengthen each other. His first-author publications have earned a CHI Best Paper Award and a NeurIPS Spotlight, and he has been named both a Rising Star in Data Science and a Rising Star in ML & Systems. His research has directly informed policy reports from the Open Data Institute and the European Parliament.
Talk Title: Enabling Democratic AI through Collective Intelligence
Abstract: AI is transforming society’s digital infrastructure, impacting vital decisions from how public agencies allocate services to how online communities moderate content. But who gets to shape this technology? I envision a future where AI is shaped by the people it serves, who best understand their own context, rather than by external developers who lack the domain knowledge to address local needs. I formalize this vision as community-centered AI.
While people seek more agency to shape the AI deployed in their communities, realizing the vision of community-centered AI is challenging in practice. In particular, members of the same community frequently have conflicting views on how their AI should be designed. Even when they reach consensus, they often lack the technical expertise to implement their decisions. My research overcomes these challenges by creating systems that surface disagreements, facilitate deliberation, and translate collective decisions into AI artifacts. In this talk, I showcase three such systems that enable groups of non-AI experts to collaborate on shaping AI policies, agents, and datasets.
Bio: Nicolas Lanzetti is a PIMCO Postdoctoral Fellow in Data Science at Caltech, working with Adam Wierman, Eric Mazumdar, and Steven Low. He received his PhD from ETH Zurich, advised by Florian Dörfler (main advisor) and Alessio Figalli (second advisor). He has held visiting positions at Stanford, MIT, and EPFL. As AI increasingly drives consequential decisions in society, engineering, and business, his research develops the mathematical and algorithmic foundations for deployable AI in strategic, uncertain, and dynamic environments. To this end, he blends and advances tools from operations research, game theory, optimal transport, machine learning, and control theory.
Talk Title: Strategically robust decision-making
Abstract: In many multi-agent settings, AI agents must make decisions against the uncertain behavior of other AI agents and humans. This uncertainty arises from multiple sources, e.g., incomplete information or bounded rationality. While each source could be modeled individually, their joint presence makes this task daunting. It is natural to seek protection against deviations around the emergent behavior itself. In the first part of the talk, we propose that each agent makes decisions in the face of the worst-case behavior contained in an ambiguity set of tunable size, centered at the emergent behavior so implicitly defined, giving rise to a novel equilibrium notion we call strategically robust equilibrium. When judiciously operationalized via optimal transport, strategically robust equilibria (i) interpolate between Nash and security strategies; (ii) come at no additional computational cost compared to Nash equilibria; (iii) often lead to better decisions and higher payoffs. Through experiments in bimatrix games, congestion games, and Cournot competition, we show that strategic robustness protects against uncertainty in the opponents’ behavior and, surprisingly, results in higher equilibrium payoffs for all players, an effect we call coordination-via-robustification. In the second part of the talk, we turn strategic robustness into a scalable training mechanism for AI agents, by introducing a fictitious agent that acts adversarially while staying close to the others’ policies. Across standard multi-agent reinforcement learning environments, warehouse operations, and fine-tuning of LLMs, strategically robust agents generalize to unseen partners, do not free ride, and sometimes achieve higher rewards than standard agents, demonstrating coordination-via-robustification at scale.
Bio: Junu Lee is a fifth-year Ph.D. student in the Department of Statistics and Data Science at the University of Pennsylvania, advised by Dr. Zhimei Ren and Dr. T. Tony Cai. Previously, he received an A.B. in Mathematics and Statistics and an S.M. in Computer Science from Harvard University. His research focuses on developing methods for multiple hypothesis testing and selective inference that can accommodate the complexities of modern datasets. Much of his work involves model- and distribution-free statistical tools (such as e-values, knockoffs, and conformal inference), contributing both novel methodology and a deeper theoretical understanding of their properties. His research also includes formulating statistically principled designs of agentic data science systems to leverage the capabilities of generative AI for large-scale data-driven discovery without sacrificing statistical validity.
Talk Title: Compound p-values for competition-based testing
Abstract: Standard competition-based multiple testing methods—such as the knockoff filter and the target-decoy procedure—are typically formulated as head-to-head competitions between a test unit and its assigned negative control unit. Motivated by real-world applications, which leverage any available negative controls, we study this competition-based framework in more intricate settings where competitions vary in size and structure. We propose new methods based on compound p-values—p-values that are superuniform only on aggregate—that enable powerful multiple testing with broader competition structures, via the Bonferroni and Benjamini–Hochberg procedures. The former provides exact familywise error rate control, whereas the latter may incur a false discovery rate inflation factor, for which we derive upper bounds across a range of settings. We showcase the performance of our general competition-based testing methodology through extensive simulations and real-data applications, including analyses of virtual spatial transcriptomics and CRISPR screen datasets that exhibit precisely the complex competition structures that motivate our methods.
Bio: Shicheng Liu is a final-year Ph.D. candidate in Computer Science at Stanford University, advised by Professor Monica S. Lam. His research develops knowledge agents and data-driven deep research systems that enable the scalable extraction, exploration, and synthesis of information from structured, unstructured, and hybrid sources. His work has appeared at ICLR, COLM, ACL, NAACL, EMNLP, and ICSE. He has been awarded the 2025–2026 IBM Ph.D. Fellowship, the 2024 Brown Institute Magic Grant, and won first place in the 2026 Databricks Grounded Reasoning Cup.
Talk Title: From Heterogeneous Knowledge to Automated Discovery: Building Next-Generation Data-Driven Deep Research Systems
Abstract: Large language models promise to transform how people investigate data, yet real-world knowledge rarely arrives in a single clean format. It is distributed across webpages, databases, knowledge graphs, and free text, making exploration and discovery inherently difficult. This talk presents a four-stage research agenda aimed at building LLM-based knowledge agents that automatically explore heterogeneous sources and surface insights.First, SCRIBES (ICLR 2026) uses layout-aware reinforcement learning to extract semi-structured web data at scale by generating reusable scripts for groups of similar pages. Second, Structured and Unstructured Query Language (NAACL 2024) extends SQL with free-text primitives, providing a composable interface for querying hybrid data while retaining the precision and efficiency of database execution. Third, SPINACH (EMNLP 2024) and DataTalk introduce agentic reasoning and planning methods that navigate large knowledge bases and complex datasets, enabling expert question answering and verifiable analysis for users such as Wikidata contributors and investigative journalists. Finally, DataSTORM (COLM 2026) moves beyond answering predefined questions toward automated exploratory research. Grounded in exploratory data analysis and data storytelling, it autonomously conducts deep research across both large-scale structured databases and internet sources. Building on DataSTORM, I will discuss ongoing collaborations with Databricks and Stanford Medicine to develop the next generation of data-driven discovery systems for complex scientific databases.Together, these systems outline a path from raw, heterogeneous information to automatic insight discovery. This line of research advances data-driven discovery by enabling people to answer complex questions, generate new hypotheses, and uncover meaningful connections in data.
Bio: Xiao Liu is a postdoctoral scholar at the Data Science Institute, University of Chicago, working with Prof. Chenhao Tan and Prof. James Evans. She obtained her Ph.D. from Peking University in 2025. Her research interests lie primarily in advancing the complex reasoning abilities (especially causal reasoning) of large language models, and applying these capabilities to support reliable decision-making and scientific discovery. She has published more than ten first-author and co-first-author papers in top AI and NLP conferences and journals, accumulating over 1,500 citations.
Talk Title: Toward Causally Intelligent AI
Abstract: Large language models can generate compelling explanations and accurate predictions, yet still misunderstand why an outcome occurs or what would happen under an intervention. Closing this gap requires causal reasoning: the ability to connect observations to underlying mechanisms, anticipate the consequences of actions, and reason about alternative possibilities. How can we develop this ability as a general capability of AI systems?
In this talk, I will present my research on advancing causal reasoning in large language models through three complementary directions: representation, tools, and training. First, I will show how programming code makes causal relationships and counterfactual branches explicit, providing both a medium for eliciting causal reasoning and a source of transferable training signals. Second, I will examine the gap between knowing causal concepts and applying them to data, and introduce a framework that transforms external references into executable tools for knowledge-intensive reasoning. Third, I will present a unified approach to training causal reasoners using diverse problems spanning association, intervention, and counterfactual inference. By aligning symbolic formulations, executable programs, and natural-language narratives, this approach grounds answers in exact causal inference and supports generalization across tasks and scenarios.
I will conclude by briefly discussing how these advances support trustworthy decisions and rigorous scientific discovery—from assessing whether evidence supports a conclusion to synthesizing findings and actively testing hypotheses. Together, these directions advance a vision of causally intelligent AI that explains outcomes, reasons faithfully from evidence, and investigates what it does not yet understand.
Bio: I am a PhD candidate at the Center for Computational Molecular Biology at Brown University, advised by Dr. Lorin Crawford and Dr. Jeffrey Bailey. My research develops machine learning models, spanning Bayesian inference and multimodal learning, that extract reliable signal from real world heterogeneous biological and clinical data. Working closely with biologists, clinicians, and public health scientists, I design models that don’t just perform well in silico but directly improve real decisions in medicine and public health. My work has been cited in the WHO World Malaria Report 2025. I am a Blavatnik Family Graduate and Frontera Computational Science Fellow.
Talk Title: Learning What the Data Doesn’t Say: Multimodal, Uncertainty-Aware Biomarker Prediction from Clinical and Wearable Data
Abstract: Machine learning for health increasingly draws on multiple data modalities such as clinical records, genomics, and wearables. Yet real-world cohorts rarely have all modalities. For large cohorts, missingness is an inconvenience. For small rare-disease communities, where predictions guide irreversible decisions and no larger dataset exists, it is the central obstacle. My research confronts this problem directly. In this talk I present a framework that predicts blood biomarkers by integrating clinical and wearable data into a single shared representation. Because clinically meaningful signals, such as how sleep and physical activity jointly relate to health, emerge only when modalities are modeled together, the framework learns across them jointly. On the All of Us Research Program, a large, diverse cohort, I show that this improves biomarker prediction while highlighting which modalities and features drive the predictions. This work is motivated by a collaboration with a patient advocacy group for a rare inherited metabolic disease, where cohorts are small. Scarcity shapes the problem in two ways. The model must transfer knowledge from where data is plentiful to where it is not. Additionally, with no large dataset to validate against, it must highlight what data drove each prediction and how confident the model is in its prediction. I conclude by outlining a future direction that treats missingness itself as structure to be modeled, inferring absent modalities probabilistically rather than imputing point estimates, so that a model accurate across a population can be trusted for a single patient.
Bio: Ashley earned a PhD in Learning Sciences from Northwestern University. As a data scientist and learning scientist, she incorporates theories about learning, youth agency, and data processes to study the teaching and learning of data science education. With a specialization in K-12 contexts, Ashley develops and implements contextual approaches to data pedagogies that situate the relationships between people and data. She mixes critical quantitative, computational, and qualitative methodologies to study these relationships and interactions underlying data-abundant learning environments. Our society’s futures depend on people’s data skills, and Ashley contributes to this by investigating and supporting youths’ understandings of data.
Talk Title: Relational Data Pedagogies for Elementary Data Science Education
Abstract: People routinely produce and use data in their daily lives to make decisions for themselves and others. Given data’s ubiquity, there is an urgent need to teach youth early and often how to interact with data. Through design-based implementation research methods, I co-design and study contextualized data approaches to youth-generated movement data activities for elementary school students. The design relies on sports and embodied knowledge as the domain for youth to incorporate their existing knowledge into their data practice.With 15 teachers, 44 undergraduate students, and 5 researchers, we implemented an eight week curriculum across 32 classrooms. In each session, youth generate, collect, transform, and interpret their embodied movement data. Using pre and post survey responses from over 400 fifth graders, I explore (1) connections between youth’s beliefs about data agency, identity, and confidence and (2) changes in these beliefs over time. I validate the survey instrument and identify key factors in data beliefs. I find that learners’ beliefs about data are intertwined with their beliefs about the domain of sports. I offer principles to design better data science curriculum for learners of all ages and scaffolds to support educators to teach these concepts in ways that acknowledge the relationships between learners and data. Beyond the impact for learners, I theorize on the role of domain knowledge in learning data science broadly. My study has implications for how we can create opportunities for youth to develop their data agency in the classroom and apply this knowledge to their daily lives.
Bio: Vaidehi Srinivas is a postdoc working in theoretical computer science and the foundations of machine learning at MIT. She is interested in providing provable guarantees for machine learning tasks, particularly in beyond worst-case settings like smoothed analysis, algorithms with prediction, and conformal prediction. Before MIT, she did her Ph.D. at Northwestern University, was a Fulbright visiting student at the University of Vienna and earned her B.S. in Computer Science at Carnegie Mellon University.
Talk Title: Volume-Optimal Conformal Prediction
Abstract: I will discuss a new framework for designing volume-optimal conformal prediction methods. Conformal prediction is an emerging and popular paradigm for ensuring reliability of black box models, by replacing the model prediction with a prediction set, such that the set contains the true outcome with at least some desired probability. A major bottleneck in the uptake of these methods is ensuring that these prediction sets are small enough to be informative. We design the first methods, to the best of our knowledge, that provably optimize for set size, while ensuring reliability. We approach this by designing new methods for learning confidence sets, which is a fundamental task with connections to many problems, like support estimation, robust estimation, and estimating density level sets. This talk is based on joint work with Chao Gao, Liren Shan, and Aravindan Vijayaraghavan.
Bio: I’m currently a Postdoctoral Researcher in the Complex Systems Group at Yale working with Dr. Dani S. Bassett. My research leverages developmental neuroscience, machine learning, and network science to characterize how developing brain networks support cognitive function in early life. Beyond the human brain, I am extending this framework to artificial networks, asking how network reconfiguration gives rise to intelligence across biological and artificial systems. My long-term goal is to establish an independent research laboratory developing data-driven frameworks to understand how intelligence emerges through network development and learning processes.
Talk Title: Developing energy landscapes of intelligence: from infant brains to artificial networks
Abstract: How does a network develop to support intelligent behavior and to do so efficiently? I address this question by treating both the developing human brain and artificial neural networks as controllable systems, using network control theory to quantify the energy required to move a network between functional states. Across scales, networks develop toward increasingly efficient energy landscapes. In the human brain, efficiency is present from the start: even at birth, structural networks reach canonical activity states at lower control energy than degree- and strength-preserving null models. Through infancy, structurally estimated control energy comes to match functional energy demand, as maturing white-matter architecture lets dynamics exploit low-cost structural pathways. Across early childhood, the energy needed to reach the majority of cognitive states declines, where sensory functions maturing early and fast, higher-order functions following a slower trajectory. Strikingly, the relative cost of transitioning between cognitive states — a map of “mental gear-shifting” — is set early and stays stable even as absolute costs fall. Artificial networks reveal parallels and contrasts. As recurrent neural networks train on cognitive tasks, their control energy falls and their landscape becomes modular; edge pruning during the network training, cutting roughly 50% of cross-family connections versus 9% within families, mediates the coupling between lower energy and higher performance. By comparing systems that are pre-programmed against those built from scratch, my series of work uncovers shared principles by which developing networks come to control their own dynamics efficiently, and points toward a common understanding of how intelligence emerges.
Bio: Qiaosen Wang is a Ph.D. candidate in Statistics at the University of Chicago, advised by Prof. Cong Ma and Prof. Chao Gao. His research lies at the intersection of statistical machine learning and information theory. He studies the fundamental limits of statistical learning problems and develops optimal and adaptable methods that remain effective under limited information and adversarial contamination. His work spans robust statistics, adaptive estimation and inference, and statistical foundations of large language models. Prior to Chicago, he received his B.S. in Mathematics from the School of the Gifted Young at the University of Science and Technology of China.
Talk Title: Simple Yet Instance-Optimal Location Estimation Adapted to Log-Concave Tails
Abstract: Location estimation can be much easier for some distributions than for others. Regular location families like Gaussian and Laplace typically exhibit a root-n rate, whereas compactly supported families such as the uniform and triangle distributions may allow faster and boundary-driven rates. This raises a natural adaptation question: can a single location estimator attain the instance-wise minimax rate for every underlying distribution without knowing its shape?We answer this question affirmatively for unknown location families with symmetric log-concave tails. We identify a critical quantity determined entirely by dyadic quantile gaps and show its equivalence, up to universal constants, to the critical Hellinger modulus governing Le Cam’s two-point testing rate — the minimax benchmark when the underlying density is known. Because dyadic quantile gaps can be approximated with distribution-free order statistics, we construct an adaptive estimator using a data-dependent convex combination of sample mid-summaries, namely averages of symmetrically placed order statistics. The resulting estimator simultaneously attains the instance-wise optimal rate for every unknown underlying density with no adaptation cost, revitalizing the classical and simple art of order-statistics-based estimation.
Bio: Qingsong Wang is a postdoctoral researcher at the Halıcıoğlu Data Science Institute at the University of California San Diego, working with Mikhail Belkin and Yusu Wang. His research aims to develop mathematical foundations for understanding and controlling generative AI, particularly diffusion and flow-matching models. Drawing on geometry and topology, he studies how training data, initial noise, and internal representations shape model behavior, with a focus on generalization, memorization, and control. He combines mathematical analysis with experiments to understand the generation process and develop methods for steering it. His long-term goal is to make generative AI a reliable tool for scientific and mathematical discovery. Previously, he was a postdoctoral researcher at the University of Utah from 2022 to 2024. He received his PhD in mathematics from The Ohio State University in 2022 under the supervision of Facundo Mémoli.
Talk Title: Understanding and Controlling Diffusion Models
Abstract: A diffusion model turns an initial noise sample into an image along a trajectory. What organizes this trajectory, and where can its behavior be changed? This talk develops a geometric view of diffusion generation around these two questions.Our analysis of flow-matching ODEs provides the overall picture. The denoiser guides each trajectory through a global-to-local progression: first toward the data mean, then toward local clusters, and finally toward the data support. We also prove terminal convergence when the data lie on low-dimensional manifolds. Within this picture, the initial condition matters in a structured way. In the image models we studied, low-frequency variations in initial noise are systematically related to global composition. Memorization also has a distinct location: it is strongest in a data-dependent intermediate-noise regime.These findings lead to interventions matched to the relevant part of generation. Editing selected components of initial noise changes composition, changing the training emphasis in the intermediate regime reduces memorization, and directions learned from internal activations steer selected features during sampling. The broader message is that diffusion is an organized dynamical process whose geometry can explain model behavior and guide effective control.
Bio: Bohan Wu is a PhD candidate in Statistics at Columbia University, advised by David Blei. Bohan develops methods and theory for large-scale probabilistic inference, with recent work on scaling and generalizing empirical Bayes, approximate posterior sampling via optimal transport, and learning causality from heterogeneous data. Before Columbia, Bohan received a BS in Mathematics and Statistics from Rice University and worked with Daniel Kowal. In June 2026, Bohan joined Netflix’s Machine Learning and Inference Research team as a research intern. Bohan has presented work on empirical Bayes at the Flatiron Institute’s Bayes Reading Group and received an ISBA Junior Travel Award to attend the 2026 ISBA World Meeting in Nagoya.
Talk Title: Empirical Bayes for Data-Driven Content Decisions
Abstract: Creators and digital platforms routinely make decisions about new content based on engagement metrics from pilot studies. While pilot studies are often limited in sample size or subject to selection bias, firms also possess abundant historical data from previously released content that can inform decisions about new products. This talk presents empirical Bayes (EB) methods that leverage historical data to improve inference and decision making for new content.The first part develops an AI-assisted empirical Bayes method that combines pilot data with a calibrated language-model prior. A large language model predicts the relative engagement of a new item from its content, and its predictions are calibrated using historical engagement data from released content to construct an informative prior. This prior is then combined with pilot data in a Bayesian discrete-time survival model to estimate user engagement. The resulting estimator substantially improves prediction accuracy compared with using pilot data alone.The second part addresses selection bias in pilot audiences. We model user heterogeneity through frailty in a survival model and develop a nonparametric empirical Bayes method to estimate the frailty distribution from a panel of historical engagement curves. The estimated frailty distribution enables us to characterize distributional shifts between pilot and full-release audiences and correct engagement estimates for these shifts.
Bio: Xindi Wu is a Ph.D. candidate in Computer Science at Princeton University, advised by Olga Russakovsky. Her research studies how data shapes the capabilities of multimodal AI systems and world models. She develops methods to trace model behavior back to training examples, design data and representations that support compositional and physical reasoning, and preserve useful experience through memory. Through this data-centric perspective, her broader goal is to build AI systems that learn efficiently from experience, reason reliably about the physical world, and continually expand their capabilities.
Her research has appeared at ICML, ICLR, NeurIPS, and CVPR. She received an Outstanding Paper Honorable Mention at ICML 2026 and an Apple Scholars in AIML PhD Fellowship. She has also received Best Paper Awards at the ICML 2026 From Frames to Stories Workshop, the ICLR 2026 Test-Time Update Workshop, and the ECCV 2024 Dataset Distillation Workshop. She co-organizes workshops at ICCV, ECCV, and NeurIPS on data-centric learning and world models.
Talk Title: Building data foundation for world models
Abstract: World models are opening new possibilities for simulating environments, planning actions, and understanding the physical world. In this talk, I will discuss my research on understanding and advancing these models, centered around how they learn from and use data. First, I will describe our efforts to trace model behavior back to the training examples responsible for it, showing how identifying the videos that teach a particular motion enables targeted improvements with a fraction of the training data. Next, I will discuss how the structure of training examples shapes what models can learn, demonstrating that deliberately composing capabilities within examples improves generalization to more complex combinations. I will then introduce our work on memory for video world models, showing that retaining past observations is insufficient unless models can still access them, and describing how compact, addressable memory supports consistent predictions over long interactions. Finally, I will outline a vision for world models that improve continually from experience, focusing on choosing informative data, preserving the physical structure needed for prediction, and learning what to remember.
Bio: Yifan Wu is a postdoctoral researcher in the Economics and Computation group at Microsoft Research New England. She earned her Ph.D. in Computer Science from Northwestern University in 2025 and her B.S. in Computer Science from Peking University’s Turing Class in 2020.
Her research develops theoretical foundations for trustworthy AI using ideas from game theory and statistical decision theory, with a focus on calibration, information elicitation, and human–AI decision making. Her work combines theory with empirical studies of real-world AI systems and human-AI experiments.
Wu was a runner-up for the 2025 ACM SIGecom Doctoral Dissertation Award and received a Best Paper Award at FORC 2025. She was also selected as a 2026 Rising Star in Theoretical Computer Science.
Talk Title: Calibration Error for Decision Making
Abstract: In modern AI workflows where prediction and decision-making are separated, an upstream prediction team must serve unknown or heterogeneous downstream decision makers. Calibration, a property of predictive models, enables this separation. It requires reported probabilities to be conditionally unbiased, so they can be trusted and directly used by any downstream decision maker without task-specific correction.We propose Calibration Decision Loss (CDL) as a decision-theoretic calibration error. CDL is defined as the largest improvement in expected payoff obtainable by recalibrating a predictor across all bounded decision problems. By definition, CDL captures the worst-case decision loss from miscalibration and therefore directly measures whether a predictive model is trustworthy for heterogeneous downstream users.Our results highlight a fundamental difference between decision-theoretic error and known calibration errors. CDL counts only biases that can change decisions, whereas standard metrics such as Expected Calibration Error (ECE) average over all prediction biases, regardless of their impact on decisions. In an online setting, this perspective leads to a prediction algorithm with near-optimal expected CDL, improving upon decision-making guarantees obtained by optimizing canonical calibration errors.
Bio: Ge (Ashley) Zhang is a Ph.D. candidate at the University of Michigan School of Information, advised by Professor Steve Oney. Her research lies at the intersection of human-computer interaction (HCI), programming education, and human-centered AI. She designs interactive systems that make complex human behaviors visible and actionable, with the goal of empowering educators to better understand and support students’ learning processes.
Her research explores how to capture, analyze, and visualize programming activities at scale, understand human–AI interactions during programming, and develop AI-powered tools that enable instructors to provide timely, personalized feedback. Through this work, she aims to bridge advances in AI and data-driven techniques with human-centered approaches to teaching and learning.
Her work has been published at leading venues, including ACM CHI, UIST, Learning at Scale (L@S), and IEEE VL/HCC, and has received multiple Best Paper and Honorable Mention awards. She previously worked as a research intern at Adobe Research.
Her long-term research vision is to build human-centered AI systems that augment human capabilities, preserve human agency, and make personalized learning support accessible at scale.
Talk Title: Scaling Instructional Decisions: Turning Programming Behavior into Actionable Intelligence
Abstract: Whenever someone builds something, they leave behind a record of how they did it, revealing more insights than the finished work. Software now captures that record in remarkable detail: edits, tests, errors, and increasingly questions asked of an AI assistant. But detail is not understanding. A single edit means little on its own, and a five-minute pause may signal careful thinking or complete confusion. The same scale that makes these traces hard to interpret is what makes them valuable, and turning behavioral data at this scale into intelligence people can act on remains a fundamental challenge. I study this problem in programming education, where it is especially hard. An introductory course can enroll hundreds of students supported by a handful of instructors, who must infer what each student understands from traces far larger than anyone can inspect. Code makes the inference harder still: it varies widely, it changes with every keystroke, and its meaning depends on how it runs, not on how it looks. Yet these traces are rich evidence of how students think, and they arrive while the class is still in progress, in time for an instructor to act. In this talk, I show what follows from studying programming not as a static artifact but as a process that unfolds through edits, executions, debugging, and AI interactions: representations that make a whole classroom readable at a glance, and mechanisms that turn an instructor’s own criteria into support delivered at a scale no instructor could reach alone.
Bio: Haoran Zhang is a postdoctoral scholar in the Department of Computer Science at Stanford University, advised by Sanmi Koyejo. His research focuses on evaluating and improving the reliability of machine learning models in healthcare, by tackling issues such as robust performance under distribution shift, learning under label noise, and designing evaluations that reflect real-world clinical use. His work has appeared in venues including Nature Medicine, The Lancet Digital Health, NeurIPS, ICML, and ICLR, and has been featured by media outlets such as Wired, VICE, and The Boston Globe. Haoran received his PhD in EECS at MIT in 2026 where he was advised by Marzyeh Ghassemi.
Talk Title: Why Clinical ML Models Fail in the Wild and How to Fix Them
Abstract: Machine learning models are increasingly embedded in clinical workflows. However, the real-world utility of such models has been limited, with many high profile cases of their failures. In this talk, I ask: why do clinical ML models fail in the wild, and how can we design systems that remain reliable under real-world conditions? Drawing on examples from predictive models and large language models, I will highlight challenges arising from label bias, distribution shift, and the use of model outputs in clinical decision-making.
Focusing on distribution shift, I will introduce a framework to attribute performance gaps across sites and subgroups to addressable distribution shifts. Through a clinical case study, I will show how these explanations can guide targeted data collection and model adaptation. I will conclude with directions for evaluating, selecting, and learning models for new clinical environments, with an emphasis on their effects on downstream clinical decisions.
Bio: Xuandong Zhao is a postdoctoral researcher at UC Berkeley, affiliated with the Berkeley Artificial Intelligence Research (BAIR) Lab, where he works with Prof. Dawn Song. He received his Ph.D. in Computer Science from UC Santa Barbara, advised by Prof. Yu-Xiang Wang and Lei Li. His research spans machine learning, natural language processing, and AI safety, with the goal of building generative AI systems that are capable, reliable, and responsible. Xuandong focuses on establishing the provenance of AI-generated content through watermarking, improving model alignment and robustness to mitigate safety and security risks, and protecting privacy, copyright, and intellectual property throughout model training and deployment. Xuandong has published more than 50 papers in leading ML, security, and NLP venues and has served as an area chair for ICML, NeurIPS, and ACL. He is a recipient of UCSB’s Chancellor’s Fellowship and Rising Star awards in adversarial machine learning and AI.
Talk Title: Building Foundations for Safe and Secure Generative AI
Abstract: As generative AI reshapes science, creativity, and industry, it raises pressing concerns about safety, security, and societal trust. A central challenge of our era is ensuring these systems are not only capable, but also safe, accountable, and aligned with human values. In this talk, I will present my research on foundations for trustworthy generative AI, with an emphasis on provable watermarking and robust alignment. First, I will introduce provable content attribution, including Unigram-Watermark for LLMs with robustness guarantees and PRC-Watermark, a cryptographically undetectable scheme for diffusion models. Second, I will discuss AI safety under adversarial manipulation, showing how “weak-to-strong” jailbreaking exposes shallow safety boundaries in current LLMs, and I will present DOOR, a dual-objective framework that significantly improves safety from the defense side. I will conclude with my vision for responsible AI, including scalable alignment for superintelligence, security for agentic systems, and trustworthy deployment in high-stakes domains.