October 2025 - talks - lists.cs.princeton.edu

Fwd: Tomorrow's ORFE Optimization Seminar: Thursday, October 9, 2025 at 4:30 PM: Antoine Deza, McMaster University
by Emily C. Lawrence 08 Oct '25

08 Oct '25

Tomorrow’s ORFE Optimization Seminar Date & Time: Thursday, October 9, 2025 at 4:30 PM Location: 101 - Sherrerd Hall Speaker: [ https://orfe.princeton.edu/events/2025/antoine-deza-mcmaster-university | Antoine Deza, McMaster University ] Title: Integrality properties of kissing polytopes distances Abstract: A lattice (d,k)-polytope is the convex hull of a set of points in dimension d whose coordinates are integers ranging from 0 to k. We investigate the smallest possible distance between two disjoint lattice (d,k)-polytopes. A pair of polytopes achieving this minimal distance are called kissing polytopes. This question arises in various contexts where the minimal distance between such polytopes appears in complexity bounds for optimization algorithms. We provide nearly matching lower and upper bounds for this distance and propose an algebraic model. Our formulation yields explicit formulas in dimensions 2 and 3, and allows for the computation of previously intractable values. We also discuss related results, such as the Alon–Vu bounds for flat simplices — that is, the minimum distance between a vertex of a lattice (d,1)-simplex and the affine space spanned by the remaining vertices. Finally, we observe that all the known squared distances between kissing polytopes are inverses of integers, and we ask whether this observation holds in general. Based on joint-work with Shmuel Onn (Technion), Sebastian Pokutta (Zuse Institute Berlin and TU Berlin), and Lionel Pournin (Université Paris 13).

1 0

Howard Yen will present his General Exam "Long-Context Language Models for Long-Horizon Agents" on Friday, October 10, 2025 at 2:00 PM in CS 401.
by CS Grad Department 07 Oct '25

07 Oct '25

Howard Yen will present his General Exam " Long-Context Language Models for Long-Horizon Agents " on Friday, October 10, 2025 at 2:00 PM in CS 401 . Committee Members: Danqi Chen (advisor), Tri Dao , Karthik Narasimhan Abstract: Agentic systems that autonomously interact with their environments enable powerful large language models (LLMs) across numerous downstream applications, such as deep research and software engineering. Due to the complex, long-horizon nature of these tasks, agentic systems need to effectively scale to extremely long trajectories with many steps. Thus, the building block of these systems—the underlying LLMs—must be both capable of processing long sequences and interacting with environments through tool use. In this talk, I will first establish the foundation for these systems by developing new evaluations for long-context language models (LCLMs) and retrieval systems, a key tool in many agentic systems. I introduce HELMET, a holistic and robust benchmark for LCLMs, and show that diverse evaluations provides more reliable signals for developing LCLMs and how they perform in real settings. Then, I propose reasoning-intensive retrieval, which extends the information retrieval task beyond lexical and semantic matching. Reasoning-intensive retrieval requires in-depth reasoning to identify relevant documents and enables agentic systems to tackle more complex problems. I will conclude with context management for long-horizon agentic search, where agents must exhaustively search and reason over many sources to solve complex tasks. I first develop an automated analysis pipeline and error taxonomym to gain a better understanding of long-horizon agentic systems and how they fail, where one of the major failure modes is context mismanagement. Then, I design a simple framework, SLIM, that allows these systems to scale effectively to long trajectories with effective context management. Reading List: https://docs.google.com/document/d/1XH4QblNAg9rEaeFWuCDBTcgl6iGbRJ6jgLRaT4R… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

Siyang Wu will present his General Exam "How to evaluate monocular depth estimation?" on Wednesday, October 8, 2025 at 3:00 PM in CS 105
by CS Grad Department 03 Oct '25

03 Oct '25

Siyang Wu will present his General Exam "How to evaluate monocular depth estimation?" on Wednesday, October 8, 2025 at 3:00 PM in CS 105 Committee Members: Jia Deng (advisor), Adam Finkelstein, Zhuang Liu Abstract: Monocular depth estimation is an important task with rapid progress, but how to evaluate it remains an open question, as evidenced by a lack of standardization in existing literature and an unhelpfully large selection of evaluation metrics whose trade-offs and behaviors are not well understood. This paper contributes a novel, quantitative analysis of existing metrics in terms of their sensitivity to various types of perturbations of ground truth, emphasizing comparison to human judgment. Our analysis reveals that existing metrics are severely under-sensitive to curvature perturbation such as making flat surfaces wavy. To remedy this, we introduce a new metric based on relative surface normals, along with new depth visualization tools and a principled method to create composite metrics with better human alignment. Reading List: https://docs.google.com/document/d/1bX49rxxy3z-RrqDJHKuIKXfQAySIYS7e7mZ5Pgj… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

QCB Seminar with Su-In Lee: Monday Oct 6 at 3pm
by Emily C. Lawrence 03 Oct '25

03 Oct '25

[ https://lsi.princeton.edu/events/2025/qcb-seminar-su-lee-university-washing… | QCB Seminar with Su-In Lee, University of Washington ] Monday, October 6, 2025 3:00 – 4:00 p.m. Carl Icahn Lab, 101 “ Explainable AI for health: where we are and how to move forward. ” The first part of my talk explores various research initiatives undertaken by my lab, emphasizing the principles and techniques of explainable AI and their application across a range of biomedical fields. I will demonstrate how explainable AI can elucidate novel scientific inquiries, with a primary emphasis on understanding neurodegenerative diseases and biological age. In the second part, we will explore the evolving landscape of explainable AI, uncovering its potential to chart new scientific directions in biomedicine, exemplified by our recent work in dermatology, emergency medicine and precision cancer medicine. This discussion aims to shed light on the necessary enhancements for explainable AI to effectively tackle a wide array of real-world challenges in biomedicine. BQ_BEGIN Professor Su-In Lee is the Boeing Endowed Professor of Computer Science at the University of Washington (UW). She earned her Ph.D. from Stanford University in 2009 under Professor Daphne Koller and joined UW in 2010 after serving as a Visiting Assistant Professor at Carnegie Mellon University. She is renowned for her groundbreaking research at the intersection of AI, biology, and medicine, and widely recognized as a pioneer in explainable AI (XAI). Among her seminal contributions is the SHAP framework (Lundberg and Lee, NeurIPS’17 oral; cited over 40,000 times), which has transformed the interpretation of machine learning models across disciplines. She has been honored with major awards, including the Samsung Ho-Am Prize in Engineering (the “Korean Nobel Prize,” as its first woman recipient in 34 years), the ISCB Innovator Award, and the NSF CAREER Award. She is an American Cancer Society Research Scholar, an AIMBE Fellow, and an ISCB Distinguished Fellow. Her recent work advances fundamental principles of XAI and applies them to biomedicine—from uncovering molecular drivers of disease to auditing clinical AI systems—fundamentally reshaping how AI is integrated into biomedical research and healthcare. This integration has enabled novel discoveries and produced [ https://aims.cs.washington.edu/su-in-lee | numerous awards ] and [ https://aims.cs.washington.edu/publications | highly cited publications ] spanning AI, molecular biology, and clinical medicine. Hosted by Yuri Pritykin and Mona Singh BQ_END BQ_BEGIN BQ_END

1 0

Next Week's ORFE Department Colloquia: Tuesday, October 7, 2025 at 4:30 PM: Milind Tambe, Harvard University
by Emily C. Lawrence 03 Oct '25

03 Oct '25

Next Week’s ORFE Department Colloquium Date & Time: Tuesday, October 7, 2025 at 4:30 PM Location: 104 - Computer Science Speaker: [ https://orfe.princeton.edu/events/2025/milind-tambe-harvard-university | Milind Tambe, Harvard University ] Title: Optimizing Social Impact: Field Deployments, Bandit Algorithms, and the Role of LLM Agents Abstract: For nearly two decades, my team’s work on AI for Social Impact (AI4SI) has focused on optimizing limited resources in critical areas like public health, conservation, and public safety. In this talk, I will highlight field test results from our deployed work in India, demonstrating measurable improvements in effectiveness for the world's two largest mobile health programs for maternal and child care. We have leveraged innovative restless and collaborative bandit algorithms to achieve these gains, revealing new technical directions in the process. Additionally, I will present our ongoing research on network-based HIV testing in South Africa, which is modeled as a branching bandit problem. These projects, along with our other health-related projects in Africa, illustrate a core challenge in our field. Deploying end-to-end AI4SI systems requires us to repeat three labor intensive steps: understanding stakeholder challenges, building tailored models, and rigorous field testing. I'll share initial results on how we can leverage current LLM-based agents and multiagents to accelerate this AI4SI process, significantly improving the speed and scale of social impact applications that focus on resource optimization. Bio: Milind Tambe is Gordon McKay Professor of Computer Science; concurrently, he is also Principal Scientist and Director at Google Deepmind. Prof. Tambe and his team have developed pioneering AI systems that deliver real-world impact in public health (e.g., maternal and child health), public safety, and wildlife conservation. He is recipient of the AAAI Award for Artificial Intelligence for the Benefit of Humanity, AAAI Feigenbaum Prize, IJCAI John McCarthy Award, AAAI Robert S. Engelmore Memorial Lecture Award, AAMAS ACM Autonomous Agents Research Award, INFORMS Wagner prize for excellence in Operations Research practice, Military Operations Research Society Rist Prize, Columbus Fellowship Foundation Homeland security award and commendations and certificates of appreciation from the US Coast Guard, the Federal Air Marshals Service and airport police at the city of Los Angeles. He is a fellow of AAAI and ACM. * Please join us for a meet and greet with some light refreshments from 4:00pm – 4:30pm in the Sherrerd Hall Atrium before the talk.

1 0

Hee Seung (Will) Hwang will present his General Exam "COMPACT: COMPositional Atomic-to-Complex Visual Capability Tuning" on Monday, October 6, 2025 at 1:00 PM in CS 402.
by CS Grad Department 01 Oct '25

01 Oct '25

Hee Seung (Will) Hwang will present his General Exam " COMPACT: COMPositional Atomic-to-Complex Visual Capability Tuning " on Monday, October 6, 2025 at 1:00 PM in CS 402. Committee Members: Olga Russakovsky (advisor), Zhuang Liu , Manoel Horta Ribeiro Abstract: Visual Instruction Tuning (VIT) data of Multi-modal Large Language Models (MLLMs) have continuously scaled in volume, with “the more the merrier” mentality when it comes to data. However, a natural question to ask is whether this blind scaling can be wasteful, and whether leveraging more informative samples can enable faster training and perhaps even more accurate models. In this paper, we present COMPACT (COMPositional Atomic-to-complex Visual Capability Tuning), a VIT data recipe that explicitly controls training sample complexity by combining multiple atomic visual capabilities in a single training example. Concretely, we synthesize rich and informative text questions for each image, allowing us to significantly reduce the number of training examples required for effective visual instruction tuning. COMPACT demonstrates superior data efficiency compared to existing data reduction methods. When applied to the LLaVA-665K VIT dataset, COMPACT reduces the data budget by 90% while still achieving 100.18% of the full VIT performance (compared to only 97.47% by state-of-the-art) across ten multimodal benchmarks. Further, training on this COMPACT data even improves performance compared to training with full-scale data on particularly complex benchmarks such as MM-Vet (8.6%) and MMStar (2.9%). COMPACT offers a scalable and efficient synthetic data generation recipe to improve on visual language tasks. Reading List: https://docs.google.com/document/d/1lhdxFBNmEkr1Kw_D6Te-4Bp4Btp7Zp7u1UAI38h… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0