Subscribe to RSS Feed| Start Date: 18 Sep 2025; Start Time: 03:30PM - 05:00PM Title: Exploring structured representations in Vision, Language and Abstract Reasoning Bio: Speaker: Abstract: Recent AI breakthroughs have significantly improved performance across various tasks, yet machine learning models still struggle with scene understanding and abstract reasoning—challenges that humans solve effortlessly. In this talk, we explore how structured representations can enhance AI’s ability to address these problems. First, we examine how organizing multimodal information into structured formats helps summarize scene content and address data imbalances, achieving state-of-the-art results in controllable image captioning. Next, we introduce structured approaches to abstract visual reasoning, leading to interpretable representations, enhanced generative capabilities, and improved generalization both in- and out-of-distribution. Through these advances, we highlight how structured representations can drive more efficient and interpretable AI systems. Location: CBIM 22 Committee: Professor Vladimir Pavlovic Assistant Professor Hao Wang Associate Professor Yongfeng Zhang Professor Olga Russakovsky (external)
|
| Start Date: 22 Sep 2025; Start Time: 10:30AM - 12:00PM Title: Toward Trustworthy Learning-Enabled Systems via Neurosymbolic Programming Bio: He Zhu is an Assistant Professor in the Department of Computer Science at Rutgers University. His research lies at the intersection of programming languages, formal methods, and machine learning, with a focus on building trustworthy learning-enabled systems. He is particularly interested in neurosymbolic programming, which integrates program synthesis with deep learning to enable interpretable, data-efficient, and verifiable AI. Prior to joining Rutgers, he was a research scientist at Galois, Inc., and earned his Ph.D. in Computer Science from Purdue University, where he received the Maurice H. Halstead Memorial Award. Dr. Zhu's work has been recognized with Distinguished Paper Awards at the ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI) and has been supported by the National Science Foundation and the Defense Advanced Research Projects Agency. Speaker: Abstract: Building reliable, learning-enabled systems requires bridging the gap between the interpretability and rigor of programming languages (PL) and formal methods (FM) with the flexibility of modern machine learning (ML). My research advances this integration through neurosymbolic programming (NSP), which frames learning as program synthesis: discovering programs in a domain-specific language that both optimize quantitative objectives and satisfy semantic constraints. While NSP offers a promising path toward trustworthy AI, it faces three fundamental challenges: designing suitable symbolic abstractions, navigating the combinatorial synthesis space, and ensuring correctness in systems that combine discrete programs with continuous neural components.In this talk, I will present new algorithmic frameworks that address these challenges in the context of autonomous robotic systems. First, I will describe reward-guided synthesis techniques that improve the scalability of programmatic reinforcement learning by efficiently searching large program spaces. Second, I will introduce an abstraction refinement method that automatically discovers symbolic elements for neuro-symbolic control programs, alleviating the burden of domain-specific language design. Finally, I will show how deductive reasoning and modular verification can provide correctness guarantees for neurosymbolic programs operating in continuous domains. Together, these contributions illustrate how principles from PL and FM can systematically advance NSP, enabling the construction of interpretable, sample-efficient, and verifiable learning-enabled systems. Location: CoRE 301 Committee: |
| Start Date: 26 Sep 2025; Start Time: 12:00PM - 02:00PM Title: Beyond the Classroom: Careers in Computer Science Lecture Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 16 Oct 2025; Start Time: 10:30AM - 12:00PM Title: Approximation Algorithms: Some ancient, some new - the good, the bad and the ugly Bio: Samir Khuller received his M.S and Ph.D from Cornell University in 1989 and 1990, respectively, under the supervision of Vijay Vazirani. He spent two years as a Research Associate at the University of Maryland, before joining the Computer Science Department in 1992, where he was a Professor for 27 years. He spent several summers at the IBM T. J. Watson Research Center, and also visited the IBM Tokyo Research Lab for several weeks. From 2003 to 2008 he was the Associate Chair for Graduate Education. and he was the first Elizabeth Stevinson Iribe Chair for CS. As chair he led the development of the Brendan Iribe Center for Computer Science and Innovation, a project completed in March 2019. In March 2019, Khuller joined Northwestern University as the Peter and Adrienne Barris Chair for CS. His research interests are in graph algorithms, discrete optimization, and computational geometry. He has published about 200 journal and conference papers, and several book chapters on these topics. He was an editor for the journal Algorithmica, and International Journal on Foundations of Computer Science, problems Editor for ACM Trans. on Algorithms, and currently is a columnist for SIGACT News and Associate Editor for Networks. He has served on several program committees including SODA 1997, APPROX 1999, APPROX 2000 (chair), STOC 2003, PODS 2006, SODA 2007, APPROX 2010, ESA 2010, STOC 2013, SPAA 2017 and SODA 2021. He served on the ESA Steering Committee from 2012-2016 and chaired the 2019 MAPSP Scheduling Workshop. From 2018-2021 he will serve as the Chair of SIGACT. In 2020, he received the CRA-E Undergraduate Research Mentoring Award. He received the National Science Foundation's Career Development Award, several Dept. Teaching Awards, the Dean's Teaching Excellence Award and also a CTE-Lilly Teaching Fellowship. In 2003, he and his students were awarded the "Best newcomer paper" award for the ACM PODS Conference. He received the University of Maryland's Distinguished Scholar Teacher Award in 2007, as well as a Google Research Award and an Amazon Research Award. In 2016, he received the European Symposium on Algorithms inaugural Test of Time Award for his work with Sudipto Guha on Connected Dominating Sets. He graduated at the top of the Computer Science Class from IIT-Kanpur.
Speaker: Abstract: NP-complete problems abound in every aspect of our daily lives. One approach is to simply deploy heuristics, but for many of these we do not have any idea as to when the heuristic is effective and when it is not. Approximation algorithms have played a major role in the last three decades in developing a foundation for a better understanding of optimization techniques - greedy algorithms, algorithms based on Linear Programming relaxations have paved the way for the design of (in some cases) optimal heuristics. Are these the best ones to use in “typical” instances? Maybe, maybe not. In this talk we will focus on two specific areas - one is in the use of greedy algorithms for a basic graph problem called connected dominating set, and the other is in the development of LP based algorithms for a basic scheduling problem in the context of data center scheduling. Location: CoRE 301 Committee: |
| Start Date: 17 Oct 2025; Start Time: 12:00PM - 02:00PM Title: Beyond the Classroom: Careers in Computer Science Lecture Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 27 Oct 2025; Start Time: 12:45PM - 02:30PM Title: Computational Aspects and Implications of Error-correcting Codes Bio: Speaker: Abstract: Error-correcting codes are large collections of strings that are pairwise far apart. Codes, by design, enjoy certain robustness guarantees, making them suitable for applications in communication, pseudorandomness, hardness of approximation, and beyond. At the same time, they give rise to a rich mathematical theory at the intersection of combinatorics, algebra, and geometry.In this talk, I will discuss various computational aspects and implications of error-correcting codes, highlighting three recent results.The first part of the talk will be devoted to permutation codes under the Ulam distance, a metric that has recently garnered attention due to applications in flash memory storage. The main highlight will be a new explicit construction of positive constant-rate codes with relative distance arbitrarily close to 1, overcoming the 1/3-distance barrier of prior constructions.The second part of the talk will concern isometric embeddings of the Hamming metric into the edit metric. Here, the main challenge is to maximize the rate, i.e., the ratio between the lengths of the input and output strings, of such embeddings. The focus will be on the first-ever constant-rate isometric embedding, along with consequences for lower bounds for problems in the edit metric.The final part of the talk will address Folded Reed-Solomon (FRS) codes — a well-studied class of codes known to achieve list-decoding capacity. The main focus will be the first fully polynomial-time algorithm running in poly(1/ε)·n·polylog(n) time for list decoding rate-R FRS codes up to radius 1-R-ε. In addition, we will see a deterministic decoder with running time f(ε)·n·polylog(n) that breaks the longstanding n^{1/ε} bound for deterministic decoding. Location: CoRE 305 Committee: Assistant Professor Karthik Srikanta Assistant Professor Roie Levin Assistant Professor Akash Kumar Sengupta Associate Professor Konstantinos Michmizos |
| Start Date: 29 Oct 2025; Start Time: 03:30PM - 05:00PM Title: Towards Robust and Generalizable Feature Representation Learning Bio: Speaker: Abstract: Feature representation learning focuses on extracting and encoding meaningful information from raw data, such as images or text. These encoded representations form the foundation for a wide range of downstream tasks, such as classification, regression, clustering, and generation. In recent years, the field has witnessed rapid progress, largely driven by advances in self-supervised learning and multimodal modeling. By leveraging large-scale, readily available datasets and designing pre-training tasks that do not rely on human annotation, researchers have been able to learn highly effective and transferable feature representations. Despite these advancements, several key challenges persist. These include: (1) designing pre-training objectives that capture the inherent structure and semantics of data, (2) ensuring learned features generalize across diverse downstream tasks, and (3) learning effectively from noisy, weakly aligned data.This research focuses on addressing these issues. Specifically, the dissertation investigates feature representation learning across multiple modalities, including human skeleton data, visual data (images and videos), and text. It explores the full spectrum of representation learning challenges, covering the design of pre-training tasks, the formulation of representation structures, and the mitigation of noise in large-scale data. To this end, the dissertation presents: (1) a hierarchical encoder combined with a pretext-based self-supervised framework for modeling structured skeleton data; (2) a codebook-based representation strategy that improves alignment between image and text modalities and addresses semantic mismatches during pretraining; and (3) an automatic data filtering system that leverages language models and multi-pathway alignment to filter noisy supervision in video-language data. Together, these contributions offer a unified perspective on learning robust, transferable, and interpretable feature representations across diverse domains and challenges. Location: CBIM 22 Committee: Prof. Dimitris N. Metaxas (advisor/chair) Prof. Hao Wang Prof. Konstantinos Michmizos |
| Start Date: 31 Oct 2025; Start Time: 11:00AM - 12:30PM Title: Adventures in High-Dimensional Quantum Codes Bio: Speaker: Abstract: Quantum computers have steadily improved over the last decade, but developing fault-tolerant quantum computing (FTQC) techniques, required for useful, universal computation remains an ongoing effort. Key elements of FTQC such as error-correcting codes are supported by a rich bed of stabilizer simulation software such as Stim and CHP, which are essential for numerically characterizing these protocols at realistic scales. Namely, tasks like benchmarking codes' practical thresholds under different error decoding strategies can only be studied via simulation. Recently, experimental groups have built nascent high-dimensional quantum hardware, known as qudits, which have a myriad of attractive properties for algorithms and FTQC. Despite the stage being set for early high-dimensional FTQC, there are no widely available qudit stabilizer simulators. We introduce Sdim, the first open-source realization of such a simulator for all dimensions. We demonstrate its correctness against existing state vector simulations and benchmark its performance in evaluating and sampling quantum circuits. Qudits, while being built out of the same devices as qubits, have access to states on a larger Hilbert space and a potentially richer structure for error correction, but neither current simulation nor experiment have supported a clear affirmative or negative. This simulator is the essential computational infrastructure by which we may systematically answer these questions, and the answers will guide error-correction efforts down the most fruitful paths towards FTQC. Location: CoRE 301 Committee: Prof. Yipeng Huang Prof. Mario Szegedy Prof. Emina Soljanin Prof. Tomasz Imielinski |
| Start Date: 31 Oct 2025; Start Time: 12:00PM - 02:00PM Title: Beyond the Classroom: Careers in Computer Science Lecture Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 04 Nov 2025; Start Time: 10:30AM - 12:00PM Title: Generalizability and Machine Learning in Biomedicine Bio: Speaker: Abstract: There are often performance differences between intra-dataset and cross-dataset tests in machine learning (ML) modeling of biomedical data, as well as those between applying ML to all samples and a subset of them (e.g., all versus older patients). However, reducing these differences may reduce ML performances. It is thus a challenging dilemma to develop models that excel in intra-dataset testing and are generalizable to subset-sample or cross-dataset testing. Therefore, we propose a multi-criteria framework to 1). Improve ML fairness in classifying multicategory cause of deaths in cancer patients; and 2). Understand and improve performance and generalizability of ML in intra-dataset and cross-dataset testing. Among the colorectal cancer patients (n=515) of various age, sex, and racial groups in the TCGA data, all ML models exhibited biases for these sociodemographic groups. Methods to optimize model performance, including testing the model on merged groups and others, show the potential to reduce disparities in model performance for different groups. Importantly, both robust Analysis of Variance (ANOVA) and Kruskal–Wallis tests consistently identified differentially expressed genes as one of the most influential factors in both cancer types. The proposed multi-criteria framework successfully identified the model that achieved both the best cross-dataset performance and similar intra-dataset performance. In summary, generalizing ML performance is challenging, as evidenced by ML biases in classifying a subset of samples and lower performance in cross-dataset testing. We seem able to develop methods to improve ML fairness and generalizable ML performance. Specifically, ML performance distributions significantly deviated from normality, which motivates using both robust parametric and non-parametric statistical tests. We also quantified and provided possible exploitability on the factors associated with cross-dataset performances and generalizability of ML models in two cancer types. A multi-criteria framework was developed and validated to identify the models that are accurate and consistently robust cross datasets Location: CoRE 301 Committee: |
| Start Date: 06 Nov 2025; Start Time: 10:30AM - 12:00PM Title: Chasing the Constant: Bridging Theory and Practice in Privacy-Preserving Machine Learning Bio: Speaker: Abstract: At the heart of the first large-scale deployments of Google Gboard’s private next-word prediction and Apple’s private federated learning framework lies a simple and perhaps the most fundamental primitive: differentially private counting in the continual release model. This primitive serves as a subroutine not only in federated learning, but also in a wide range of applications, including histogram estimation, non-interactive local learning, graph statistics, stochastic convex optimization, and matrix analysis.In this talk, I will present my recent work establishing deep connections between private continual counting and a concept in operator algebra (factorization norms), which not only advances the foundations of privacy-preserving learning but also improves more than three decades-old results in operator algebra. I will also highlight my results in private graph analysis, where my work resolved an open problem in differential privacy, the first efficient and exact sampling algorithm from a logconcave distribution defined over non-convex sets, and improved results in discrepancy of shortest paths. I will conclude with my future plans. My goal is to design next-generation privacy-preserving deep learning algorithms that adapt to time-varying data sensitivity and system constraints, and to extend these methods into decentralized AI/ML solutions that reduce reliance on large data centers. This vision aims to make privacy-preserving machine learning both theoretically rigorous and practically deployable, paving the way for scalable, sustainable, responsible, and inclusive AI systems. Location: CoRE 301 Committee: |
| Start Date: 07 Nov 2025; Start Time: 05:00PM - 06:30PM Title: Antithetic Noise in Diffusion Models Bio: Speaker: Abstract: We initiate a systematic study of antithetic initial noise in diffusion models. Across unconditional models trained on diverse datasets, text-conditioned latent diffusion models, and diffusion posterior samplers, we find that pairing each initial noise with its negation consistently yields strongly negatively correlated samples. To explain this phenomenon, we combine experiments and theoretical analysis, leading to a symmetry conjecture that the learned score function is approximately affine antisymmetric (odd symmetry up to a constant shift), and provide evidence supporting it. Leveraging this negative correlation, we enable two applications: (1) enhancing image diversity in models like Stable Diffusion without quality loss, and (2) sharpening uncertainty quantification (e.g., up to 90% narrower confidence intervals) when estimating downstream statistics. Building on these gains, we extend the two-point pairing to a randomized quasi-Monte Carlo estimator, which further improves estimation accuracy. Our framework is training-free, model-agnostic, and adds no runtime overhead.Link: https://arxiv.org/pdf/2506.06185 Location: CoRE 305 Committee: Professor Peng Zhang Professor Ruixiang Tang Professor Hao Wang Professor Yipeng Huang |
| Start Date: 14 Nov 2025; Start Time: 12:00PM - 02:00PM Title: Beyond the Classroom: Careers in Computer Science Lecture Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 17 Nov 2025; Start Time: 10:00AM - 12:00PM Title: Almost Sharp Bounds on Weaver's Discrepancy for Gaussian Vectors Across Different Regimes Bio: Speaker: Abstract: Weaver's discrepancy measures the minimum operator norm of signed sums of rank-one matrices and plays an important role in discrepancy theory and applications such as the Kadison-Singer problem, graph sparsification, and experimental design. While the worst-case bound is well understood following the Marcus-Spielman-Srivastava resolution of the Kadison-Singer conjecture (Annals of Mathematics, 2015), the average-case behavior remains less explored.In this talk, we will study Weaver's discrepancy under a probabilistic model where the input vectors are drawn independently from the standard Gaussian distribution in $\mathbb{R}^d$.We will provide almost tight bounds on the discrepancy across a wide range of dimensions $d$ relative to the number of vectors $n$. In the extremely subcritical regime $d = o(n^{1/4})$, we establish almost matching upper and lower bounds that are exponentially smaller than the Marcus-Spielman-Srivastava bound, and our lower bounds are valid in the entire subcritical regime $d = o(n^{1/2})$.In the supercritical regime $d = \omega (n^{1/2})$, we show that the discrepancy is $\Omega(\sqrt{dn})$ with high probability, matching known upper bounds and demonstrating that the Marcus-Spielman-Srivastava bound is tight even for Gaussian random inputs. Our results give a nearly complete characterization of the average-case discrepancy for Gaussian vectors and clarify the critical threshold that separates different asymptotic regimes.Related Publications:Ziyi Cai, Qing Chen & Peng Zhang. (2025). WEAVER'S DISCREPANCY FOR GAUSSIAN RANDOM VECTORS. SIAM Journal on Discrete Mathematics, 39(3), 1418-1447. https://doi.org/10.1137/24M1678878Qing Chen, Yunwei Ren & Peng Zhang. ALMOST SHARP BOUND ON WEAVER'S DISCREPANCY FOR GAUSSIAN VECTORS ACROSS DIFFERENT REGIMES. In submission. Location: CoRE 305 Committee: Assistant Professor Peng Zhang Professor Jie Gao Assistant Professor Kangning Wang Assistant Professor Hao Wang |
| Start Date: 18 Nov 2025; Start Time: 10:30AM - 12:00PM Title: The Many Facets of Monte Carlo: Quantum-Inspired Averaging and Antithetic Diffusion Sampling Bio: Speaker: Abstract: Monte Carlo methods play a crucial role in statistics, computer science, and physics. I will present two such facets. The first is a quantum-inspired averaging process on graphs, a simplified abstraction of Google’s quantum sampling experiment, where local random interactions drive convergence toward equilibrium. We analyze its mixing time and establish universal lower bounds on how fast such distributed sampling dynamics can approach uniformity. The second comes from diffusion-based generative models, where pairing each Gaussian noise with its negation induces strong negative correlation. This simple antithetic design sharply reduces sampling variance and improves reliability in downstream estimation tasks. Location: CoRE 301 Committee: |
| Start Date: 19 Nov 2025; Start Time: 03:30PM - 04:30PM Title: CS TeaTime Meet and Mingle Bio: Speaker: Abstract: Location: CoRE 350 Committee: |
| Start Date: 21 Nov 2025; Start Time: 11:00AM - 12:30PM Title: From Reasoning to Acting: A Program Synthesis Approach to Reinforcement Learning Bio: Speaker: Abstract: The pursuit of artificial agents that can autonomously learn and act in complex environments is a central ambition of Artificial Intelligence (AI). Within this pursuit, Reinforcement Learning (RL) has emerged as a core computational framework for creating agents that can bridge the gap from high-level reasoning to effective real-world acting. The dominant Deep RL paradigm, however, faces fundamental challenges with policy opacity, which precludes formal verification, and compositional brittleness, which leads to poor sample efficiency. This dissertation argues that the principles of Program Synthesis provide a unified framework to address these challenges by creating policies that are structured, verifiable, and grounded in data. We develop this thesis through three interconnected contributions that span the spectrum from formal logic to learned behavior. First, to enable high-level reasoning, we introduce GCRL-LTL, a deductive synthesis framework where agents generate behavioral plans that are provably correct with respect to complex temporal logic specifications, enabling zero-shot generalization by decoupling high-level planning from low-level control. To translate these plans into structured policies, we then develop π-PRL and π-HPRL, a search-based synthesis approach that automatically discovers the optimal, interpretable program structure for an agent's policy through a differentiable relaxation of the program search space. Finally, to ensure these policies are grounded in experience and can learn to act robustly, we introduce PREFORL, an inductive synthesis method that learns effective behaviors from static, offline datasets of examples by bypassing direct value function estimation via a novel contrastive learning objective. Collectively, this work demonstrates a complete pathway from reasoning to acting within RL, showcasing how the diverse strengths of program synthesis can be leveraged to create a new generation of autonomous agents that are not only high-performing, but also logical in their reasoning, transparent in their planning, and robustly grounded in their actions, representing a meaningful step towards more reliable and Trustworthy AI. Location: CoRE 301 Committee: Professor He Zhu Professor Ruixiang Tang Professor Abdeslam Boularias Professor Abhinav Verma (external)Professor Abhinav Verma (external) |
| Start Date: 24 Nov 2025; Start Time: 09:00AM - 10:00AM Title: Applying LLMs in Creative Design Bio: Speaker: Abstract: Automating creative design workflows poses significant technical challenges in tool integration, knowledge retrieval, and controlled content generation. Recent progress in Large Language Models (LLMs) offers new opportunities, but it remains unclear how to enable LLMs to (1) coordinate multiple computational tools, (2) exploit large design libraries in a structured manner, and (3) generate novel artifacts that extend beyond existing data. This dissertation investigates these questions through the domain of automated card-game prototyping, a setting that demands novel game mechanics design, consistent code generation, and scalable gameplay AI.List of Publications: Danrui Li, Sen Zhang, Samuel S. Sohn, Kaidong Hu, Muhammad Usman, and Mubbasir Kapadia. 2025. Cardiverse: Harnessing LLMs for Novel Card Game Prototyping. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 29723–29750, Suzhou, China. Association for Computational Linguistics. Location: CBIM 22 Committee: Professor Mubbasir Kapadia Professor Vladimir Pavlovic Professor Yongfeng Zhang Professor Peng Zhang |
| Start Date: 24 Nov 2025; Start Time: 05:30PM - 06:30PM Title: Towards Efficient and Reliable Adaptation of Foundational AI Models Bio: Speaker: Abstract: Foundation AI models must adapt efficiently and reliably to new tasks and evolving data distributions, yet current adaptation techniques often suffer from overfitting, unstable generalization, and an inability to quantify uncertainty. This qualifying exam presents a unified agenda toward efficient and reliable adaptation through both principled Bayesian learning and theoretically grounded continual adaptation. First, we investigate reliable post-training adaptation through Bayesian parameter-efficient methods. BLoB (Bayesian Low-Rank Adaptation by Backpropagation; NeurIPS 2024) introduces a variational low-rank framework that jointly learns posterior means and covariances of weight updates, substantially improving generalization and uncertainty estimation while retaining the efficiency of LoRA-style adapters. Building on this, TFB (Training-Free Bayesianization; NeurIPS 2025) provides a simple yet theoretically grounded approach for converting any pre-trained LoRA adapter into a Bayesian one without additional training, enabling reliable uncertainty quantification at negligible computational cost. Complementing these efforts on post-training reliability, UDIL (A Unified Approach to Domain Incremental Learning with Memory; NeurIPS 2023) develops the first principled theory unifying replay-based domain-incremental learning methods, showing that existing algorithms optimize a common error bound with fixed coefficients, and introducing adaptive coefficients that achieve tighter bounds and superior performance across sequential domain shifts. Together, these contributions form a coherent methodology for adapting foundation models that is computationally efficient, uncertainty-aware, and theoretically principled, paving the way for robust, scalable, and trustworthy foundation AI systems. Location: CoRE 305 Committee: Professor Hao Wang Professor Chengzhi Mao Professor Ruixiang Tang Professor David Pennock. |
| Start Date: 01 Dec 2025; Start Time: 09:00AM - 10:00AM Title: From Efficiency to Experience: Human-Centered Cyber-Physical Systems for Urban Logistics Bio: Speaker: Abstract: Cyber-Physical Systems (CPS) tightly integrate sensing, computation, and control to couple the physical and cyber worlds, enabling data-driven modeling and decision-making that enhance physical operational performance. Driven by the expansion of e-commerce and accelerating urbanization, urban logistics, where couriers are the key agents, is becoming a critical and fast-evolving domain for human-centered CPS. This dissertation aims to understand human behaviors in urban logistics and then enhance the delivery system through intelligent decision-making policies. To this end, we present three works: i) REDE, a relay-based last-mile delivery system, predicts delivery couriers’ mobilities and optimizes relay courier scheduling to improve delivery efficiency; ii) WORD, a workload-constrained order assignment system, measures courier energy consumption of various behaviors (i.e., cart pulling, walking, and stair climbing) and optimizes order assignment considering efficiency and courier experience; iii) HeroCS, a heterogeneous task assignment system, optimizes real-time courier scheduling for delivery, pick-up, and customer expansion tasks, considering efficiency and courier experience. These works are implemented and evaluated with real-world data from one of the largest logistics companies in China. The promising results pave the way towards human-centered CPS for future urban logistics. Location: CoRE 305 Committee: Professor Desheng Zhang Professor Yongfeng Zhang Professor Dong Deng |
| Start Date: 01 Dec 2025; Start Time: 10:00AM - 11:00AM Title: Learning Human Behavior in Human-centric Cyber-Physical Systems Bio: Speaker: Abstract: Cyber-Physical System (CPS) integrates sensing, computation, and control with the physical process, which has transformed how the physical world interacts with the digital world. With billions of devices interconnected worldwide and the rapid progress of AI models such as large language models, CPS has greatly enhanced the efficiency and automation of physical systems, such as healthcare and logistics. However, as CPS evolves toward increasingly interactive and adaptive architectures, humans are no longer passive recipients of system outputs but active participants in sensing, learning, and decision-making loops. This transformation introduces a new paradigm, Human-Centric Cyber-Physical Systems (HCPS), where human behavior plays a critical role in shaping system intelligence. This dissertation investigates how to model, learn, and leverage human behavior to improve the intelligence and performance of large-scale HCPS. Specifically, we will introduce three representative systems. (i) GenHAR learns fine-grained human motion dynamics from wearable IMU sensor data, enabling cross-domain behavior recognition without target-domain supervision. By modeling sensor attention in the frequency domain, GenHAR achieves 9.97% higher accuracy and a significant reduction in computational costs compared to state-of-the-art baselines. (ii) CoMiner learns coarse-grained human mobility patterns from large-scale GPS trajectories of over delivery couriers, achieving 95.1% accuracy and over 20% improvement compared to existing methods. (iii) Finally, CarbonSense demonstrates the societal impact of HCPS by quantifying human behavior-driven carbon emissions across nationwide logistics operations. These systems are evaluated with real-world datasets and have shown promising results in modeling large-scale human behavior and supporting practical applications. Location: CoRE 305 Committee: Professor Desheng Zhang Professor Dong Deng Professor Yongfeng Zhang |
| Start Date: 01 Dec 2025; Start Time: 03:30PM - 05:30PM Title: Controllable and Efficient Generative Models: Methods and Applications in Medical Imaging Bio: Speaker: Abstract: Deep generative models have revolutionized visual computing, achieving unprecedented realism in synthesizing images, videos, and 3D scenes. However, as the quality of generation matures, the critical challenge shifts towards achieving precise, user-driven control and practical, efficient deployment, especially in high-stakes domains like medical imaging. This dissertation addresses two fundamental obstacles hindering this transition. The first is the "controlled generation" in emerging generative architectures, such as discrete diffusion and next-scale visual autoregressive models, whose non-invertible sampling mechanisms prevent the recovery of latent codes necessary for high-fidelity editing. The second is the "practicality gap" in clinical applications, where a confluence of data scarcity (few-shot learning), data heterogeneity (non-IID distributions), stringent privacy constraints (requiring Federated Learning), and communication bottlenecks impede the collaborative development of robust AI models.This dissertation presents a cohesive body of work that bridges these gaps, progressing from foundational algorithmic innovations to their application in real-world medical imaging. The primary contributions are fourfold: (1) DICE, a pioneering inversion framework that enables, for the first time, controllable editing for discrete diffusion and masked generative models; (2) VARIN, the first noise inversion-based editing technique for next-scale visual autoregressive models, which introduces a novel pseudo-inverse for the argmax operator; (3) DMCVR, a morphology-guided diffusion model that solves the clinical problem of 3D cardiac volume reconstruction from sparse MRI by leveraging explicit anatomical conditioning; and (4) a comprehensive suite of Federated Learning frameworks that integrate few-shot learning, dual knowledge distillation, and parameter-efficient fine-tuning (LoRA) to enable robust, private, and communication-efficient medical image analysis on decentralized data. Collectively, this research provides a validated blueprint for building the next generation of visual computing systems that are not only powerful but also controllable, efficient, and trustworthy. Location: CoRE 301 Committee: Prof. Dimitris N. Metaxas (Chair) Prof. Konstantinos Michmizos Prof. Hongyi Wang |
| Start Date: 09 Dec 2025; Start Time: 10:00AM - 11:00AM Title: Metric Distortion of Matching: A Stochastic Perspective Bio: Speaker: Abstract: Bipartite matching is a fundamental problem with a wide range of applications, in many of which only ordinal preferences, rather than the underlying cardinal utilities, are available. The metric distortion framework was established to measure the efficiency loss when only ordinal information is used. However, determining the optimal metric distortion for bipartite matching has proved to be notoriously difficult.In this talk, I will present some recent progress on metric distortion of matching under stochastic models. Average-case distortion on a uniform line: We show that when agents and items are drawn i.i.d. from a uniform distribution on a one-dimensional segment, the simple matching rule of random serial dictatorship achieves exponentially better distortion than the best-known distortion in the worst-case model. Semi-random model: When points in an arbitrary metric space are randomly partitioned into two sides, we propose a new simple matching rule with distortion better than the best-known distortion in the worst-case model.Related Publication(s): Ziyi Cai, Qing Chen, Kangning Wang, and Peng Zhang. Metric Distortion of Matching: A Stochastic Perspective. In submission. Location: CoRE 305 Committee: Assistant Professor Peng Zhang Assistant Professor Kangning Wang Assistant Professor Xintong Wang Professor Eddy Z. Zhang |
| Start Date: 10 Dec 2025; Start Time: 08:00AM - 09:30AM Title: Learning, Directing, and Validating Synthetic Crowds: Parametric Inference, Design Optimization, and Heterogeneous Reinforcement Learning with Variance-Based Evaluation Bio: Speaker: Abstract: Modern society increasingly depends on accurate, controllable simulations of human crowds for safety-critical planning, architectural design, and immersive media. Yet existing crowd simulators face fundamental challenges: they are difficult to calibrate from real-world trajectory data, too computationally expensive for large-scale design exploration, often lack learned control policies that are both general and richly heterogeneous, and provide no principled metrics to determine which parameters meaningfully drive emergent behaviors.This thesis presents a unified, learning-based framework that addresses these interconnected challenges through four parametric components. First, we introduce a trajectory-to-configuration inference framework that regresses latent motion patterns to simulator parameters, enabling interpretable calibration of classical models such as Social Force and ORCA. Second, we develop neural surrogates trained on simulated egress data that predict crowd performance metrics orders of magnitude faster than full simulation, enabling crowd-aware optimization of complex architectural layouts. Third, we propose a parametric multi-agent reinforcement learning method (HOP-RL) that yields heterogeneous, directable agent behaviors with reciprocal predictive collision avoidance, generalizing across scenarios without retraining. Fourth, we introduce variance decomposition analysis that separates parameter effects from simulation noise, providing sample-size-invariant measures of parameter importance and revealing how population structure reshapes parameter sensitivity.Across synthetic validations and real-world case studies, including evacuation analyses of complex built environments, we demonstrate that this integrated approach improves calibration fidelity, optimization efficiency, and behavioral realism while substantially reducing the effective parameter space that practitioners must manage. The result is a principled, end-to-end methodology for learning, directing, and validating synthetic crowds that is simultaneously data-informed, design-oriented, and evaluation-driven, enabling a coherent workflow from observed data to informed design decisions. Location: CoRE 301 Committee: Professor Mubbasir Kapadia Professor Vladimir Pavlovic Professor Kostas Bekris Professor Yutong Zhao (External) |
| Start Date: 10 Dec 2025; Start Time: 12:00PM - 02:00PM Title: Generation and Optimization of Physics-Based Graphics Simulations Bio: Speaker: Abstract: Simulations of physical phenomena are used in many fields, from the study of liquids and solids inmaterial science, to the exploration of diffusion phenomena, to creating animations and graphics, andmachine learning models. Such simulations are compute-intensive. Their usefulness is often limited bythe speed and accuracy that can be achieved, by memory limitations, and by the difficulty of coding.The research work presented in this dissertation addresses these limitations and offers improvements,touching on all three aspects of physical simulation: a direct use case, the speed of the simulation, andthe ease of writing the simulation.The work started with a direct use case for physics simulation: a detailed simulation of the spread ofCOVID-19 within moving crowds. It integrated an ecological model of diffusion and immunity tosimulate the disease spread, with a force model to calculate the crowd movement. The interactivesimulation was the first to model disease spread within the field of computer graphics.The model was insightful, however the simulations were slow and computationally expensive,particularly in 3D. This obstacle led to a second study, which investigated speeding up 3D simulationsvia memory access optimization. That study focused on the simulation of elastic solids. Usually,methods of optimizing simulations for speed make solver-specific or scenario-specific optimizations. Incontrast, the technique presented in this dissertation’s second study is a general acceleration methodthat can be used to speed up any simulation of elastic solids. This is accomplished by rearranging thedata of the elastic solid meshes to make the ordering of the vertices and faces cache-efficient. Therearranged data can then be used in any solver, resulting in a faster simulation run.The work done to optimize the simulations brought to the fore the complications inherent inprogramming efficient physical simulations. The programmer is required to have intricate knowledgeof the internal workings of the computer. The simulation code must be written in a low-level language(usually C++). The necessity of low-level programming results in bloated and repetitive code, withplenty of room for bugs to hide. The final part of this dissertation describes a solution to this hurdle forgrid-based fluid simulations: a domain-specific language called Spade. In Spade, users specify theparameters of a simulation in a high-level specification file, which is then automatically compiled intooptimized C++ code. The resulting simulations are based on the sparse grid representation, which hasrecently emerged as an indispensable component of fluid simulation pipelines due to its impact onmemory savings and performance. In Spade, the complexities of maximizing computational efficiency(through leveraging cache coherency and parallelization) are handled behind the scenes, allowing thedeveloper to focus on defining the grid, data, and the core algorithm. Location: CBIM 22 Committee: Professor Mridul Aanjaneya Professor Peng Zhang Professor Richard Martin |
| Start Date: 12 Dec 2025; Start Time: 03:30PM - 05:00PM Title: Generative Models for Applications in Accessible Video Communication Bio: Speaker: Abstract: The deaf community shares videos online to express rich linguistic content conveyed through both hand movements and facial expressions. However, this reliance on visible articulators makes it difficult to protect signer identity without compromising meaning. At the same time, advances in generative models within computer vision create new opportunities for accessible, privacy-preserving video communication. Motivated by the intersection of need and technological capability, this work applies modern generative modeling techniques to the task of sign language video anonymization, an emerging research area aimed at enabling secure and meaningful communication for Deaf users. The research begins with a human-centered study that identifies user requirements and constraints for sharing sign language content securely. Guided by these findings, the second contribution introduces a motion-based image animation framework that retargets identity using optical-flow–driven synthesis while maintaining the linguistic fidelity of facial and manual movements. To enhance practical applicability, the final contribution presents a diffusion-based anonymization method that incorporates text guidance control for robust performance in real-world video. Additional discussions explore how other generative models can further support accessibility. Together, these contributions form a coherent progression toward generative technologies that advance both accessibility and privacy. Location: CoRE 305 Committee: Professor Dimitris N Metaxas (Advisor) Professor Konstantinos Michmizos Professor Hongyi Wang Professor Junzhou Huang (External) |
| Start Date: 16 Dec 2025; Start Time: 09:30AM - 10:15AM Title: Multi-Modal Memory Management System for Building Personalized LLM Agents Bio: Speaker: Abstract: Long-term conversational agents struggle to maintain deep context and personal consistency using single-axis memory systems. This research addresses this by developing LocalMem, a novel memory system that integrates multi-modal personalized history with a robust hierarchical architecture. Built on the principles of utilizing multi-modality in memory management, LocalMem organizes data not only across temporal and semantic dimensions but also incorporates information from diverse user activity streams, including text, visual, and computer-usage logs. This rich, multi-dimensional data is curated and stored in a scalable memory structure, enabling the system to provide LLMs (like Gemini or ChatGPT) with deep, localized, and contextually complete recollections of the user's past activities. By leveraging hybrid management strategies against this multimodal memory, LocalMem achieves superior long-term personalization, validating a crucial path toward building truly stateful and human-like agents. Location: CoRE 305 Committee: Prof. Lesia Semenova Prof. Ruixiang Tang Prof. Desheng Zhang Prof. Yongfeng Zhang |
| Start Date: 16 Dec 2025; Start Time: 04:00PM - 06:00PM Title: Unsupervised Adaptation for Image Classification via Sparsity-Guided Steering Vectors Bio: Speaker: Abstract: Steering vision foundation models at test time, without retraining or access to large labeled datasets, is a desirable yet challenging goal, particularly in dynamic or resource-constrained settings. This talk presents Visual Sparse Steering (VS2), a lightweight, label-free test-time method that constructs a steering vector from sparse features extracted by a Sparse Autoencoder (SAE) trained on the model’s internal activations. VS2 consistently improves the top-1 accuracy of the CLIP zero-shot baseline on the downstream image classification task across 9 diverse datasets. Since not all features learned by the SAE are equally important for classification, we further introduce VS2++, a retrieval-augmented variant that selectively amplifies relevant sparse features using pseudo-labeled neighbors retrieved from an external unlabeled corpus at inference time. With oracle positive and negative sets that approximate an upper bound, VS2++ achieves substantially larger gains over the CLIP zero-shot baseline, highlighting the potential of steering vectors when relevant feature selection is accurate. VS2 and VS2++ also yield marked per-class accuracy improvements, indicating that sparse steering disproportionately benefits visually or semantically similar classes. Finally, VS2 includes a built-in reliability diagnostic based on SAE reconstruction loss, absent in common steering-vector methods, which signals when steering may underperform and safely triggers a fallback to the baseline. Location: CoRE 301 Committee: Professor Konstantinos Michmizos Professor Hao Wang Professor Badri Nath Professor Dimitris Metaxas |
| Start Date: 19 Dec 2025; Start Time: 09:00AM - 10:00AM Title: LLM Aided Digital Storytelling Bio: Speaker: Abstract: Digital storytelling, essential in entertainment, education, and marketing, faces challenges in generation efficiency. The StoryAgent framework utilizes Large Language Models and generative tools to automate and refine digital storytelling. Employing a top-down story drafting and bottom-up asset generation approach, StoryAgent tackles key issues such as manual intervention, interactive scene orchestration, and narrative consistency. This framework enables efficient production of interactive and consistent digital storytelling across multiple modalities, democratizing content creation and enhancing engagement.Based publication: https://dl.acm.org/doi/10.1145/3677388.3696321 Location: CBIM 22 Committee: Professor Mubbasir Kapadia Professor Yongfeng Zhang Professor Peng Zhang Professor Minesh Patel |
| Start Date: 19 Dec 2025; Start Time: 01:15PM - 02:30PM Title: Detecting Deep Bugs with Local Unrollings for Bounded Translation Validation Bio: Speaker: Abstract: Checking two programs for semantic equivalence is an important problem with applications to detecting compiler bugs, superoptimization, and code refactoring. Translation validation (TV) is a heuristic-guided approach for checking program equivalence and is used to prove the correctness of each compiler invocation.Bounded TV is a variant of TV that reasons about the bounded semantics of programs. A bounded TV works by unrolling all program loops up to a fixed global bound, encoding the semantics of both programs into SMT formulas, and using an SMT solver to check for a semantic refinement. Such tools have been used extensively for detecting miscompilation bugs in compiler infrastructures such as LLVM.Although bounded TV is fully automated, it fails to scale for programs with deep bugs that require large unroll factors, as a global unroll bound can lead to a blow-up in the state space of the bounded program. In this talk we introduce local unroll bound annotations which annotate loops with a minimum unroll bound attribute that a bounded TV can use. Local unroll bound annotations target specific regions of the program to expose deep bugs that are otherwise intractable with a global unroll bound. This talk will also explore heuristics for automatically annotating programs with local unroll bound annotations. We implement our approach with the MLIR compiler infrastructure, and also enable compatibility with existing bounded TV infrastructures. Our results indicate a substantial improvement in the scalability of deep bug detection. Location: CoRE 301 Committee: Professor Richard Martin Professor Santosh Nagarakatte Professor He Zhu Professor Roie Levin |
| Start Date: 19 Dec 2025; Start Time: 03:30PM - 05:00PM Title: Adaptive AI Under Distribution Shifts: Methods for Vision-Language Models and Medical Imaging Bio: Speaker: Abstract: This dissertation addresses the fundamental challenge of distribution shifts, which cause machine learning models to degrade when deployed in environments that differ from their training data. It develops new test-time adaptation techniques that allow models to adjust on the fly without retraining or access to source data, thereby improving performance and reliability across diverse settings. In this talk, I will first outline our latest efforts to advance adaptation strategies for general-purpose vision systems and address the limitations of existing work for this task. Then, I will present our work on improving model reliability in medical image analysis, and I will discuss the challenges associated with this task, particularly when models must operate across heterogeneous imaging conditions and clinical environments. Together, these contributions move toward adaptable AI systems capable of maintaining strong performance under real-world and unpredictable distribution shifts. Location: CBIM 22 Committee: Professor Dimitris N. Metaxas Professor Hongyi Wang Professor Konstantinos Michmizos |
| Start Date: 22 Dec 2025; Start Time: 10:00AM - 12:00PM Title: Multi-LLM Collaboration for Automatic Data and Representation Engineering Bio: Speaker: Abstract: Large Language Models (LLMs) are increasingly deployed as general-purpose inference engines. Prior research has shown that LLM behavior can be highly sensitive to how task inputs are presented. Motivated by this, our work studies a paradigm that collaboratively leverages multiple LLMs as automatic, task-specific data and representation engineers to transform raw task inputs (or outputs) to satisfy downstream objectives and constraints, either as a pre-processing step prior to inference or as training signals for fine-tuning. Specifically, we study this paradigm in three settings: (i) privacy-preserving text obfuscation that reduces sensitive exposure while largely maintaining task utility; (ii) data optimization via automated content engineering and structural reformulation to systematically improve downstream performance beyond prompt-only tuning; and (iii) generative retrieval with learned indexing, where we learn hierarchical semantic indices for items and perform retrieval via constrained decoding to guarantee validity, enabling an end-to-end, cold-start-friendly framework that avoids multi-stage retrieval pipelines.Publications Included:EmojiPrompt: Generative Prompt Obfuscation for Privacy-Preserving Communication with Cloud-based LLMs (https://aclanthology.org/2025.naacl-long.614/)ADO: Automatic Data Optimization for Inputs in LLM Prompts (https://aclanthology.org/2025.findings-acl.1340/)Multi-layer Generative Item Retrieval for Natural Language Queries (preprint attached) Location: Remote Via Zoom Committee: Professor Yongfeng Zhang Professor Ruixiang Tang Professor Dong Deng Professor Mario Szegedy |
| Start Date: 12 Jan 2026; Start Time: 10:30AM - 12:00PM Title: Towards Realistic Environment Representation for Human Simulation Bio: Speaker: Abstract: Realistic human simulation in virtual environments is foundational for games, training, and design evaluation, yet current navigation pipelines often trade off plausibility, speed, and scalability. We introduce a unified environment representation that targets the mid-level of route choice: a novel method that first constructs an embedded navigation graph to capture accessibility-aware topology and geometric cues, then leverages this graph to synthesize a navigable surface mesh for robust locomotion. Building on this representation, we propose a flexible route-choice framework that composes factors such as topology-driven accessibility and energy cost (e.g., slope) with visibility analysis, and supports continuous blending between competing objectives (e.g., maximizing distance versus maximizing view). Finally, to support dense crowds, we present a GPU-oriented approach that decouples utilization from environment scale, addressing performance bottlenecks common in resolution- or scene-size-dependent methods. Together, these contributions enable more realistic, controllable, and scalable human navigation in complex virtual environments. Location: Remote Via Zoom Committee: Associate Professor Mubbasir Kapadia Professor Vladimir Pavlovic Professor Jie Gao Associate Professor Sudarsun Kannan |
| Start Date: 22 Jan 2026; Start Time: 10:30AM - 12:00PM Title: Limits of Memory-Efficient Learning: New Information-Theoretic Tools Bio: Sumegha Garg is an Assistant Professor in the Department of Computer Science at Rutgers University. Prior to joining Rutgers, she was a postdoctoral fellow in the CS Department at Stanford University and a Rabin Postdoctoral Fellow in the Theory of Computation group at Harvard University. She completed her Ph.D. in Computer Science from Princeton University, where she was advised by Mark Braverman. Her research interests span complexity theory, information theory, and learning theory, with a particular emphasis on memory lower bounds and the theory of responsible machine learning. Speaker: Abstract: In this talk, I will discuss my work toward developing a foundational theory of memory-constrained machine learning (ML). While classical learning theory has extensively studied the data and time requirements of ML tasks, our understanding of their memory requirements remains limited. At the same time, the growing memory demands of large-scale ML systems, including large language models, make this question compelling both theoretically and practically. I will begin with an overview of my recent results quantifying memory requirements for a range of ML tasks, such as classification, mean estimation, and outlier detection. I will then introduce a new information-complexity–based framework for proving memory lower bounds in these settings. In the last part of the talk, I will briefly present my other recent results in complexity theory, particularly in memory-constrained property testing and coding theory. Location: CoRE 301 Committee: |
| Start Date: 11 Feb 2026; Start Time: 10:00AM - 12:00PM Title: Robot Data is Not Enough Data Bio: Lerrel Pinto is the co-founder of Assured Robot Intelligence (ARI) and an Assistant Professor of Computer Science at NYU. His research aims to get robots to generalize and adapt in the messy world we live in. To this end his work focuses broadly on robot learning and decision making, with an emphasis on large-scale learning (both data and models), representation learning for sensory data, developing algorithms to model actions and behavior, reinforcement learning for adapting to new scenarios, and building open-sourced affordable robots. This work has received best paper awards at ICRA 2016, RSS 2023, and ICRA 2024. Lerrel has received the Samsung AI researcher of the year award, Sloan Fellowship, Packard Fellowship, CIFAR Fellowship, TR35 innovator under 35, IEEE RAS Early Career and NSF CAREER awards. Several of his works have been featured in popular media such as The Wall Street Journal, TechCrunch, MIT Tech Review, Wired, and BuzzFeed among others. His recent work can be found on www.lerrelpinto.com. Speaker: Abstract: The past decade of robot learning has been fueled by piles of human-teleoperated robot data. But this strategy is hitting a wall. Unlike computer vision and natural language processing, fields supercharged by mountains of passive, internet-scale human-labeled data, robotics faces a harsher reality. Robot data is expensive. It is slow. It is narrow. And most critically, we don’t even know which demonstrations or labels truly matter for embodied intelligence. Chasing more of the same is a dead end.In this talk, I will argue that robot data alone will never deliver the leap we need. We must demand more. Robots should learn directly from humans. They should feel the world through touch, rather than staring at pixels alone. And they must go beyond purely reactive modes and instead reason, plan, and act with foresight. If we are serious about building intelligent machines, we must move beyond the fixation on “just more data” and instead embrace the hard, messy, human-centered problems that will define the next era of robotics. Location: Edward J. Bloustein School of Planning and Public Policy 33 Livingston Avenue New Brunswick, NJ 08901 Room 261 Committee: |
| Start Date: 13 Feb 2026; Start Time: 12:00PM - 02:00PM Title: Beyond the Classroom: Careers in Computer Science Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 18 Feb 2026; Start Time: 04:00AM - Title: Bounding Distributional Shifts in World Modeling through Novelty Detection Bio: Speaker: Abstract: Recent work on visual world models shows significant promise in latent state dynamics obtained from pre-trained image backbones. However, most of the current approaches are sensitive to training quality, requiring near-complete coverage of the action and state space during training to prevent divergence during inference. To make a model-based planning algorithm more robust to the quality of the learned world model, we propose in this work to use a variational autoencoder as a novelty detector to ensure that proposed action trajectories during planning do not cause the learned model to deviate from the training data distribution. To evaluate the effectiveness of this approach, a series of experiments in challenging simulated robot environments was carried out, with the proposed method incorporated into a model-predictive control policy loop extending the DINO-WM architecture. The results clearly show that the proposed method improves over state-of-the-art solutions in terms of data efficiency. Location: Online Only Committee: Associate Professor Abdeslam Boularias Professor Kostas Bekris Associate Professor Mridul Aanjaneya Assistant Professor He Zhu |
| Start Date: 20 Feb 2026; Start Time: 10:30AM - 12:00PM Title: Tales from eBPF in Production and Research Ramblings Bio: Theophilus A. Benson is a professor in the Electrical and Computer Engineering Department at Carnegie Mellon University and Carnegie Mellon University-Africa, specializing in models, algorithms, and frameworks to improve computer network performance and availability. Speaker: Abstract: Meta is one of the largest users of eBPF, with recent studies showing that individual servers host as many as 250 programs. In this talk, I will provide a brief characterization of BPF use at Meta and discuss an effort I led to develop a management framework to address the emerging complexity by introducing new abstractions. Then, discuss a research effort from my group that introduces frameworks to address a significant pain developers face when trying to ensure optimality. Specifically, I will discuss a framework that performs source-to-source transpilation to optimize eBPF code bases. Finally, I will conclude with a broader discussion of my group's attempt to tackle challenges using eBPF for a specific use case, i.e, observability. Location: CoRE 301 Committee: |
| Start Date: 25 Feb 2026; Start Time: 02:15PM - 03:30PM Title: Reliable Frontier Expansion for Goal-Conditioned Reinforcement Learning Bio: Speaker: Abstract: Efficient exploration in sparse-reward, long-horizon environments remains a core challenge for goal-conditioned reinforcement learning (GCRL). Frontier-based exploration is a common strategy, but frontier states are often difficult for a partially trained policy to reach, limiting its effectiveness. We introduce CE2 (Cluster Edge Exploration), which prioritizes exploratory goals on the boundary of clusters of states reachable under the current policy, balancing novelty with reachability to expand explored regions more reliably. To further reduce the cost of frontier discovery and improve generalization, we propose MUN (World Models for Unconstrained Goal Navigation), which identifies key subgoal states from replay and learns bidirectional transitions between arbitrary pairs, enabling broader coverage of the state space. While CE2 and MUN are fully unsupervised, many robotic environments contain large, irrelevant regions that hinder efficient learning. To address this, we introduce Cago (Capability-Aware Goal Sampling), which leverages demonstration trajectories to build a dynamic, capability-aligned curriculum that guides exploration toward task-relevant goals. Together, CE2, MUN, and Cago form a unified framework for scalable exploration in complex, real-world GCRL settings. Location: CoRE 305 Committee: Assistant Professor He Zhu Associate Professor Abdeslam Boularias Assistant Professor Xintong Wang Assistant Professor Sharon Levy |
| Start Date: 25 Feb 2026; Start Time: 05:00PM - 07:00PM Title: Beyond the Classroom: Careers in Computer Science Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 26 Feb 2026; Start Time: 10:30AM - 12:00PM Title: Security: A Next Frontier in AI Coding Bio: Jingxuan He is a Postdoctoral Researcher at UC Berkeley, working with Prof. Dawn Song. He earned his PhD at ETH Zurich, where he was advised by Prof. Martin Vechev. His research lies at the intersection of security and AI, with a focus on quantifying AI’s impact on cybersecurity and developing proactive defenses to mitigate associated risks. His work has been recognized with an ACM CCS Distinguished Paper Award and an ETH Medal for Outstanding Doctoral Thesis, and is adopted by leading AI labs like Anthropic. For more information, please visit https://jxhe.info/. Speaker: Abstract: AI is reshaping software development, yet this rapid adoption risks introducing a new generation of security debt. In this talk, I will present my research program aimed at transforming AI from a source of vulnerabilities to a security enabler. I will begin by introducing a systematic framework for quantifying AI-induced cybersecurity risks through two benchmarks: CyberGym, which evaluates AI agents’ offensive capabilities in vulnerability reproduction and discovery, and BaxBench, which measures LLMs’ propensity to introduce security flaws when generating code. Building on these findings, I will present a secure-by-design approach for AI-generated code. This includes security-centric fine-tuning that embeds secure coding practices directly into models, as well as a decoding-time constraining mechanism based on type systems to enforce safety guarantees. Finally, I will conclude by discussing my future research on building broader security and trust in AI-driven software ecosystems. Location: CoRE 301 Committee: |
| Start Date: 27 Feb 2026; Start Time: 02:00PM - 03:00PM Title: Computer Science Department Tour - Prospective Students Bio: Speaker: Abstract: Location: CoRE 301 Committee: |
| Start Date: 03 Mar 2026; Start Time: 10:30AM - 12:00PM Title: Rethinking AI Systems Through Efficient Model Communication Bio: Yuhan Liu is a fifth-year PhD student at the University of Chicago, co-advised by Junchen Jiang and Shan Lu. Her research interest is in building efficient large-scale system and networking support for ML model inference. Her works appeared in top computer system/networking conferences, such as OSDI, SIGCOMM, NSDI. She received MIT EECS rising star, EuroSys best paper award, and UChicago’s Neubauer PhD fellowship for her research. She also leads two open-source projects that build large-scale KV caching layer for efficient LLM inference, and are used in over 30 companies in production, including Google Cloud, Amazon AWS, NVIDIA, IBM etc. Speaker: Abstract: For decades, AI models interacted with humans directly through human-readable inputs and outputs (e.g., texts, images). Today, they are used much more ubiquitously and often interact through complex software systems, interacting with other models or software rather than directly with humans. This paradigm shift raises a natural question: can models interact with other models and software using model-native languages?In this talk, I will present my work on facilitating model-native interactions among models and between models and software. To enable more efficient and practical model interactions using model-native states (i.e., KV cache) in LLM systems, my work CacheGen is the first system to share KV cache across different user queries by compressing it into compact bitstreams, and my work DroidSpeak is the first system to share KV cache across different models. My research made real-world impacts via the open-source project, LMCache, widely used in production by top-tier AI companies. Together, these works make LLM inference 5–10x faster than state-of-the-art inference engines. To enable more accurate model-to-software communication, my work ChameleonAPI encodes software code structure into model-native loss functions, allowing models to be retrained for up to 43% higher application-level accuracy in vision applications. Location: CoRE 301 Committee: |
| Start Date: 03 Mar 2026; Start Time: 02:00PM - Title: From Biological to Artificial Learning: Computational Frameworks for Interpreting Representational Change Bio: Speaker: Abstract: Learning is a defining characteristic of intelligent systems, yet our understanding of how it unfolds differs across biological and artificial domains. Neuroscience has revealed that learning is accompanied by continual reorganization of neural representations. Conversely, machine learning has largely focused on optimizing performance under static assumptions. This disconnect limits our ability to interpret learning as a dynamic process that links internal representations to behavior. My research seeks to address this by developing computational frameworks that treat representational change as a central object of study. This work emphasizes the interpretability of learning dynamics, alignment with biological principles, and the development of methods that reveal how representations reorganize over time. By examining learning in both neural and artificial systems, I identify shared structural signatures and demonstrate how computational models can serve as probes of plasticity. This perspective positions learning as a process that can be interpreted, compared, and tracked across systems, tasks, and timescales, contributing to an understanding of how stable behavior emerges from dynamic representations, with implications for machine learning, neuroscience, and the design of adaptive intelligent systems. Location: CoRE 301 Committee: Professor Konstantinos Michmizos Professor Dimitris Metaxas Professor Casimir Kulikowski Professor James Abello |
| Start Date: 04 Mar 2026; Start Time: 10:30AM - 12:00PM Title: Foundations of Generative AI: Data Dependence in Diffusion Models Bio: Chris Scarvelis is a final-year PhD student in the Geometric Data Processing Group at MIT, where he works on the theoretical foundations of generative AI. His research focuses on understanding why modern generative models produce novel content rather than memorizing their training data, and how a model’s behavior is shaped by its training data. His work has been supported by an MIT–Google Future Research Cohort Fellowship, an Exponent Fellowship, a Siebel Scholarship, and an NSERC PGS-D Scholarship. Beyond his core research in generative modeling, Chris has worked on graph neural networks at Twitter, 3D deep learning at Backflip AI, and model merging methods for large language models at the MIT-IBM Watson AI Lab. Speaker: Abstract: Diffusion models are fundamental tools in generative AI, powering systems that synthesize data ranging from images to video to molecules. However, their empirical success has been accompanied by striking theoretical paradoxes. For example, training a diffusion model to optimality yields a model that can only generate training data, and diffusion models succeed in practice precisely because they fail to reach this trivial optimum. This gap between theory and practice highlights how little we know about the relationship between diffusion models’ real-world behavior and their training data.This talk will describe my recent efforts toward closing this gap. In the first part of the talk, I will explore how neural networks’ smoothness bias enables neural diffusion models to generate new samples, and how this perspective leads to a training-free generative model obtained by simulating this bias directly. I will then introduce a regularizer that captures diffusion models’ bias toward learning locally low-rank functions and show how to exploit the compositional structure of neural networks to implement this regularizer efficiently. Zooming in on the impact of the training data on a particular trained model’s behavior, I will finally describe a framework for differentiating diffusion models with respect to their target distributions. This sensitivity analysis reveals how a diffusion model’s score function and generated samples respond to infinitesimal perturbations of its training set.Together, these results build toward a predictive theory of generative AI that explains the real-world behavior of generative models via neural architectures, training procedures, and the geometry of their training data. Beyond theory, this framework has practical implications for data attribution, targeted data acquisition, and the design of more interpretable and controllable generative models. Location: CoRE 301 Committee: |
| Start Date: 05 Mar 2026; Start Time: 10:30AM - 12:00PM Title: Architecting Physical Intelligence: Cross-Stack Co-Design from Systems to Silicon Bio: Zishen Wan is a postdoctoral fellow at Harvard University, working with Prof. Vijay Janapa Reddi. He received his Ph.D. from Georgia Tech, advised by Profs. Arijit Raychowdhury and Tushar Krishna. His research focuses on computer architecture, with an emphasis on cross-stack co-design of systems, architectures, and silicon for physical intelligence. His work appears in venues including ASPLOS, MICRO, HPCA, JSSC, ISSCC, and DAC, and has been recognized with Best Paper Awards at DAC, CAL, and SRC JUMP2.0, First Place Awards in DAC PhD Forum and ACM Student Research Competition, and honorable mention in IEEE Micro Top Picks. He is a recipient of Qualcomm, Baidu, and CRNCH PhD Fellowships, and was named as ML and Systems Rising Star and Cyber-Physical Systems Rising Star. His research has been featured in MIT Technology Review and Fortune, and adopted by industry partners including Intel, IBM, and Google. For more information, please visit https://zishenwan.github.io/. Speaker: Abstract: Physical intelligence – where embodied agents perceive, reason, plan, and act in the physical world – is emerging as a new computing frontier spanning robotics, autonomous systems, and spatial AI. However, today’s physical intelligence systems remain constrained by high latency, energy cost, and fragile reliability, due to fundamental mismatch between their compositional nature and existing computing architectures. The core challenge extends beyond algorithms, to how we architect computing systems and silicon that natively support intelligence that reasons and adapts under real-world constraints.In this talk, I will present a principled cross-stack system-architecture-silicon co-design approach to building the computational foundations for physical intelligence. (1) First, I will introduce CogSys, a flexible hardware architecture for efficient neuro-symbolic cognition. CogSys features unified kernel abstractions and reconfigurable dataflows, delivering notable efficiency improvements across neuro-symbolic models and validated on FPGAs. (2) Next, I will showcase the first programmable SoC tapeout for neuro-symbolic cognition, demonstrating how tightly integrated memory-centric computing, heterogeneous architectures, end-to-end compilation flow, and adaptive power management enable efficient cognition in silicon. (3) Building on this foundation, I will present ReCA, an integrated hardware architecture that bridges high-level cognition and low-level autonomy under stringent power and latency constraints by leveraging spatial-aware runtimes, heterogeneous fabrics, and hybrid memory hierarchies. (4) Finally, I will highlight our agile SoC design flows that translate evolving cognition and autonomy workloads into efficient silicon implementations.By bridging computer architecture, system software, and silicon validation, my research establishes adaptive, accelerator-rich computing substrates for physical intelligence. This work advances a vision in which AI and hardware are co-designed, co-reason, and co-adapt, architecting future computing systems as active enablers of intelligence in the physical world. Location: CoRE 301 Committee: |
| Start Date: 05 Mar 2026; Start Time: 12:00PM - 01:30PM Title: From Physical Modeling to Interactive Systems for Physics-Based Graphics Simulation Bio: Speaker: Abstract: Physics-based simulation has produced a wide range of vivid and realistic results in computer graphics, supporting a range of simulation goals across different application domains. Achieving such realism often requires carefully designed physical models that capture coupled physical processes and fine-scale detail. This work begins with a Lagrangian particle-based formulation for simulating deformation, fracture, and diffusion in thin membrane-like structures, such as aluminum foil, rubbery films, and seaweed flakes. The proposed deformation-diffusion coupling framework generates detailed and heterogeneous fracture growth for both in-plane and out-of-plane motions, producing geometrically rich fracture patterns induced by diffusion. In addition to material-scale phenomena, physics-based simulation is also applied to systems composed of large numbers of interacting discrete entities. The second contribution introduces an agent-to-agent contagion-immunity formulation for simulating infectious disease spread within moving crowds, where crowd motion is computed offline and disease propagation is evaluated interactively, allowing real-time exploration of infection scenarios and health interventions such as immunity and vaccination. While this approach enables interactivity at the application level, the simulation workflow still follows a staged design, with computation, visualization, and interaction handled separately. These limitations motivate the final contribution, Weber, an interactive, end-to-end, web-based framework for physics-based simulation that integrates computation, visualization, and interaction from the outset. Implemented in Rust and compiled to WebAssembly, Weber features a lightweight domain-specific language that allows users to define, modify, and experiment with simulation models and parameters directly in the browser. To address the performance constraints of web environments, Weber incorporates multiple levels of simulation-specific optimizations, enabling interactive execution while preserving numerical robustness. Experimental results demonstrate that Weber achieves performance modestly slower than native C++ implementations for grid-based simulations, while significantly improving accessibility, portability, and ease of experimentation. Together, these contributions demonstrate a progression from offline simulation toward integrated, accessible, and interactive physics-based modeling. Location: Zoom Committee: Professor Mridul Aanjaneya Professor He Zhu Professor Kostas Bekris |
| Start Date: 09 Mar 2026; Start Time: 10:30AM - 12:00PM Title: Decentralization at Scale: Rethinking Design of Distributed Systems Bio: Suyash Gupta is an Assistant Professor at the Department of Computer Science, University of Oregon. Prior to joining UO, he received his Ph.D. degree from University of California, Davis and did postdoctoral research as University of California, Berkeley. He also holds two Master of Science degrees: one from Purdue University and other from Indian Institute of Technology Madras. His current research focuses on designing efficient, fault tolerant distributed consensus and communication, and federated learning algorithms. He has also co-authored a book on fault-tolerant distributed transaction processing at Morgan & Claypool. He has been awarded the Best Graduate Researcher Award for 2021 by UC Davis, Best Paper Award at EuroSys'23, and Distinguished Reviewer Award by SIGMOD'25. He also led the design of Apache ResilientDB, which is incubating under Apache Software Foundation (ASF) and is employed by researchers at multiple academic and industrial organizations. In his free time, Suyash likes to code, and his team won Best Hacker Award at BostonHacks, HackIllinois, and HackPrinceton, among others. Speaker: Abstract: Nearly half of the world’s population lives in regions where access to open and independent financial systems is limited or heavily regulated; individuals face restrictions on using traditional financial services or freely transferring funds. Even in stable economies, a high concentration of control over public resources creates systemic risk. A common approach to mitigating such centralized control is to adopt decentralized technologies: decentralized applications operate without the authority of any single organization by allowing multiple parties to share responsibility for data management. However, enabling multiple parties to operate together introduces the possibility of arbitrary failures. A well-known strategy to protect a consensus protocol—and, by extension, the underlying decentralized system—from arbitrary failures is to make it Byzantine Fault Tolerant (BFT). Existing BFT protocols, however, deliver substantially lower throughput and higher latency than industrial-grade protocols designed only for simpler crash failures. Furthermore, most BFT protocols depend on unrealistic trust assumptions, which has hindered their widespread adoption.My research focuses on making BFT consensus protocols more practical, which in turn can accelerate the adoption of decentralized systems beyond cryptocurrencies. Specifically, my work comprises three main thrusts: (1) optimizations for efficient consensus, (2) primitives for high-throughput communication among decentralized systems, and (3) applications to large-scale data management. In this talk, I will describe these thrusts and illustrate their impact on decentralized system design. I will also highlight efficient communication primitives that enable multiple groups of parties to communicate effectively in wide-area network (WAN) environments. Location: CoRE 301 Committee: |
| Start Date: 10 Mar 2026; Start Time: 04:00PM - 05:30PM Title: Beyond the Classroom: Careers in Computer Science Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 11 Mar 2026; Start Time: 03:30PM - 04:30PM Title: Memory-Centric Agent Systems for Long-Horizon and Multimodal Reasoning Bio: Speaker: Abstract: Large language models (LLMs) excel at short-context tasks but struggle with long-horizon reasoning that requires structured, persistent memory. Traditional retrieval-augmented generation (RAG) relies on flat similarity search, limiting its effectiveness in complex, multi-step environments. To enable structured memory access, we introduce ReaGAN, a node-centric agentic framework that utilizes a dual-retrieval mechanism over graph memory to capture long-range dependencies without the training costs of fully supervised GNNs. For tasks requiring iterative memory coordination, we propose DeepSieve, an autonomous multi-agent framework that recursively decomposes queries and routes sub-questions across heterogeneous sources via an information sieving mechanism. To ensure memory fidelity over extended periods, we introduce TwinVoice, a longitudinal digital twin benchmark spanning ten years, and propose the Individual Turing Test (ITT) to evaluate lifelong memory and identity consistency. While these text-based frameworks establish a robust memory foundation, real-world tasks increasingly demand sensory perception. To bridge this gap, we are developing a vision-centered multimodal memory benchmark targeting compositional inference and cross-session reasoning. Together, ReaGAN, DeepSieve, TwinVoice, and our ongoing multimodal extensions form a unified memory-centric system stack for scalable, long-horizon reasoning in next-generation AI agents. Location: CBIM 22 & Zoom Committee: Professor Dimitris Metaxas Professor Ruixiang Tang Professor Hongyi Wang Professor Karthik C. Srikanta |
| Start Date: 12 Mar 2026; Start Time: 10:30AM - 12:00PM Title: Bridging the Hardware-Software Gap: Practical Systems for a Heterogeneous World Bio: Rohan Kadekodi is a postdoc at the University of Washington, advised by Prof. Baris Kasikci. His research focuses on building practical systems for heterogeneous hardware to meet the demands of modern storage, memory, and AI workloads. He received his PhD from the University of Texas at Austin, where he was advised by Prof. Vijay Chidambaram. Speaker: Abstract: Modern applications demand high bandwidth, low latency, and large capacity simultaneously. In response, the hardware landscape has diversified to span GPUs, DRAM, CXL-attached memory, storage-class memory (SCM), and SSDs. However, software systems have remained monolithic, built on decades-old assumptions that target a single hardware technology. My research bridges this gap through a principled methodology: diagnose bottlenecks, decompose monolithic architectures, and accelerate only what is slow, yielding systems that are both high-performing and practical by preserving the stability of mature software while selectively exploiting new hardware where it matters most. In this talk, I will focus on two key aspects of this agenda: storage systems for new technologies and AI systems that leverage DRAM to alleviate GPU memory pressure. For storage, I will show how decomposing file system responsibilities enables near-device-speed performance on modern storage devices without sacrificing reliability. For AI systems, I will show how decomposing retrieval and inference pipelines enables efficient offloading of data and model weights from GPU to DRAM without sacrificing latency or accuracy. I will conclude with future directions on GPU resource management, where exponentially growing AI infrastructure spending meets strikingly low GPU cluster utilization, presenting opportunities for heterogeneity-aware systems from hyperscalers to the edge. Location: CoRE 301 Committee: |
| Start Date: 12 Mar 2026; Start Time: 02:00PM - 03:30PM Title: Deep Learning Methods for Accelerated MRI Reconstruction: Optimization-Inspired Unrolled Networks and Beyond Bio: Speaker: Abstract: Magnetic Resonance Imaging (MRI) provides superior soft-tissue contrast without ionizing radiation, but long acquisition times limit clinical efficiency and increase motion artifacts. Accelerated MRI reconstructs images from undersampled k-space measurements, forming an ill-posed inverse problem that requires strong prior knowledge. Deep unrolled networks have become apowerful framework by combining data consistency with learned regularization; however, existing models require separate networks for different acquisition settings, rely on slow and poorly scalable gradient descent schemes, and often become unstable when scaled. This dissertation develops a unified and scalable framework for accelerated MRI reconstruction. We first establish a learned half-quadratic splitting formulation that provides a principled foundation for deep unrolling and enables prompt-based conditioning, allowing a single model to generalize across diverse acquisition protocols. We then introduce adaptive gradient descent with pixel-wise learning rates and momentum acceleration, achieving up to 9× faster convergence while reducing GPU memory usage by 55%. To support large-scale modeling, we scale the architecture to 245 million parameters and propose a cascade-wise spectral-norm optimization strategy for stable training. Experiments on cardiac, brain, and knee MRI benchmarks demonstrate state-of-the-art reconstruction quality across datasets and sampling patterns. Beyond supervised unrolled models, we further explore diffusion-based generative priors for plug-and-play reconstruction and implicit neural representations for fully self-supervised reconstruction. Together, these contributions improve scalability, generalization, and stability, revealing a fundamental tradeoff between reconstruction quality and supervision. Location: CBIM 22 Committee: Professor Dimitris N. Metaxas Professor Hongyi Wang Professor Konstantinos Michmizos Professor Junzhou Huang (University of Texas at Arlington) |
| Start Date: 19 Mar 2026; Start Time: 10:30AM - 12:00PM Title: From Scaling up to Scaling out: Reality World Simulators for Physical AI Bio: Dr. Wayne Wu is a postdoctoral researcher at UCLA Computer Science, working closely with Bolei Zhou, and collaborating with Trevor Darrell (UC Berkeley EECS) and Jiaqi Ma (UCLA CEE). He received his Ph.D. in Computer Science and Technology from Tsinghua University in June 2022 and was previously a visiting Ph.D. student at Nanyang Technological University. He also spent seven years in industry, where he led the research and development of products that reached more than 10 million end users worldwide. His research lies at the intersection of computer vision, robotics, and computer graphics. He focuses on developing infrastructure and methods to scale physical AI, enabling robots to work reliably and safely in the open world. He has published over 50 papers at top-tier venues including CVPR, ICCV, ICLR, NeurIPS, and ICRA, with over 9,500 citations and 10,000 GitHub stars. His work has received a CVPR Best Paper Candidate and multiple Oral, Spotlight, and Highlight presentations. He was also honored with the 2025 UCLA Chancellor’s Award for Postdoctoral Research, recognizing the best postdocs at UCLA, and he was the only awardee from the School of Engineering. He serves as an Area Chair at CVPR 2026. Page: https://wywu.github.io/ Speaker: Abstract: Recent progress in large language and vision models demonstrates how far we can go by scaling with vast internet-scale data. In contrast, physical AI, agents that perceive and act in the real world, still lags far behind.Today, both academia and industry primarily pursue generalizable physical AI by scaling up: collecting large-scale action–video datasets or training world models that enable interaction through learned environments. However, this paradigm is inherently inefficient and will soon reach a data ceiling.In this talk, I argue for a shift from scaling up to scaling out. I introduce reality world simulators, a new paradigm that converts real-world videos into diverse, interactive simulation environments. Instead of relying on more data collection, this approach expands data through structured reconstruction and recomposition, enabling both higher data efficiency and physically grounded interaction.I will present a three-pronged approach: 1) Scaling out via Digital Twins: reconstructing controllable, interactive environments from monocular videos to support diverse agent exploration. 2) Scaling out via Digital Cousins: disentangling scene structure into compositional elements to generate large-scale variations of real-world environments. 3) Scaling out via Embodied Humans: incorporating realistic human dynamics to improve safety and social compliance in robot learning. Finally, I will outline a roadmap toward building generalizable and safe physical AI systems for open-world deployment. Location: CoRE 301 Committee: |
| Start Date: 21 Mar 2026; Start Time: 01:30PM - 03:00PM Title: Learning to Adapt: From Structured Domain Adaptation to Efficient Inference for Large Language Models Bio: Speaker: Abstract: Adaptation is a fundamental challenge in machine learning: models must remain effective under domain shift, changing task requirements, and limited computational budgets. We study adaptation as a unifying principle across two settings: structured domain adaptation and efficient inference for large language models. First, it develops methods that move beyond uniform domain alignment by modeling relationships among domains through graphs, taxonomies or latent domain indices, enabling more effective transfer under complex distribution shifts. Second, it extends adaptation to inference in large language models, introducing methods that compress or implicitly encode task context to reduce inference cost while maintaining strong performance. Together, these contributions present a broader view of adaptation, showing how machine learning systems can be designed to operate effectively under structural, informational, and computational constraints. Location: CoRE 301 Committee: Assistant Professor Hao Wang Professor Chengzhi Mao Professor Ruixiang Tang Prof. Yuzhe Yang (external) |
| Start Date: 23 Mar 2026; Start Time: 10:30AM - 12:00PM Title: From Mechanics to Learning: Building Robust Contact-Rich Manipulation Skills Bio: Yifan Hou is a Postdoctoral scholar at the Department of Electrical Engineering, Stanford University. Before that, he was an Applied Scientist at Amazon Robotics. He received his Ph.D. in Robotics from CMU in 2021, received B.E. in Automation from Tsinghua University in 2015. His research focuses on robust and dexterous manipulation in contact-rich environments. His work has won the Outstanding Paper Finalist award at CoRL 2025, and two best paper awards at workshops in ICRA 2025 and CoRL 2025. Speaker: Abstract: Robotic manipulation in the real world demands robustness to uncertainty and variability. Purely learning-based approaches often struggle in contact-rich settings due to data inefficiency and limited physical understanding, while classical planning and control methods lack scalability for unstructured systems. In this talk, I will present a series of work that explore the connection between the two complimentary perspectives. I will introduce principled analyses that characterize manipulation failures and quantify robustness through variable compliance, leading to efficient control and planning algorithms. Building on these insights, I will show how force and compliance can be incorporated into learning from demonstrations and corrections, enabling robots to acquire robust manipulation skills from limited human data. I will conclude by discussing future challenges for deploying robust manipulation in diverse real-world environments. Location: CoRE 301 Committee: |
| Start Date: 24 Mar 2026; Start Time: 09:00AM - 11:00AM Title: Context-Aware Decision Making for Disruption-Resilient Urban Logistics: From Execution Uncertainty to Infrastructure Failure Bio: Speaker: Abstract: Modern urban logistics systems operate at unprecedented scale, processing millions of parcels and coordinating thousands of vehicles daily. While recent advances in optimization and machine learning have substantially improved system performance under nominal conditions, real-world operations are inevitably subject to disruptions that violate the assumptions of conventional approaches. These disruptions manifest across multiple levels of the logistics pipeline, from physical execution variability in sorting centers to information errors in address data and infrastructure failures in transportation networks, and their effects frequently cascade beyond the point of origin. Yet, existing methods largely address each level in isolation and assume a benign operating environment, leaving logistics systems fragile in precisely the scenarios that matter most.This dissertation develops a unified methodological perspective for building disruption-resilient urban logistics systems. The central thesis is that resilience to real-world disruptions requires *context-aware* decision making: systems must continuously incorporate external feedback, whether from execution outcomes, user-generated data, or environmental signals, and adapt their decisions accordingly. We instantiate this perspective through three complementary systems targeting distinct disruption types along the logistics pipeline.First, we address execution-level disruption in large-scale parcel sorting centers, where the heterogeneous and time-varying capabilities of human workers and robotic arms render traditional bin packing formulations inadequate. We propose SHARP, a scalable human-robot collaborative packing framework that models execution feasibility as the interaction between task-induced difficulty and handler-side adaptability. By decoupling execution context modeling from allocation decisions, SHARP achieves real-time deployability under stringent latency constraints. Extensive evaluation with JD Logistics demonstrates that SHARP reduces human workload by 12.6%, improves space utilization by 26.5%, and achieves a 29.6% reduction in operational cost during multi-week online deployment.Second, we address information-level disruption caused by abnormal addresses in nationwide logistics. Erroneous addresses, including missing administrative regions, nested locations, and misspellings, lead to approximately 25,000 daily re-routing events and over $2 million in annual losses at JD Logistics alone. We propose AddrLLM, a retrieval-augmented large language model framework for address rewriting that integrates multi-instruction supervised fine-tuning, address-centric retrieval augmented generation, and a bias-free objective alignment module grounded in the operational LBS system. AddrLLM reduces parcel re-routing caused by abnormal addresses by over 40% across multiple provinces in four-month online deployment.Third, we address infrastructure-level disruption in urban transportation networks, where traffic signal malfunctions degrade intersection capacity and trigger cascading congestion that directly impairs logistics delivery efficiency. We propose MalLight, an influence-aware coordinated traffic signal control framework that enables functioning intersections to collaboratively mitigate the impact of malfunctioning signals. Through graph diffusion-based state and reward aggregation, MalLight captures spatially and temporally varying inter-intersection influences, allowing each agent to balance local optimization with system-wide resilience. Experiments on real-world datasets demonstrate that MalLight alleviates throughput reduction by up to 48.6% compared to state-of-the-art methods.Across all three systems, we identify a common design principle: effective disruption resilience arises not from building more complex monolithic models, but from structuring learning architectures around domain-specific context, explicitly encoding what feedback matters, how it propagates, and how decisions should adapt in response. All systems are validated on large-scale real-world datasets and industrial deployments, demonstrating that context-aware decision making offers a practical and principled path toward robust urban logistics at scale. Location: CoRE 305 Committee: Associate Professor Desheng Zhang Associate Professor Yongfeng Zhang Associate Professor Deng Dong Professor Guang Wang (external) |
| Start Date: 24 Mar 2026; Start Time: 11:30AM - 01:30PM Title: Exploring Truthfulness and Applications of Prediction Models: Dynamics between Learners and Experts Bio: Speaker: Abstract: Prediction models are prevalent in the real world; the ability to accurately predict future events has historically been a valuable one. New tools and methods to better understand prediction models are being developed every day. Our primary aim is to gain a deeper understanding between interactions of prediction models, those who construct them, and the experts or forecasters they query. Our goals of this dissertation are three-fold. First, we provide guidelines that experts, that is, predictors of events, can use to be better informed about prediction modeling, such as whether or not to be truthful. We explore experts algorithms, where a learner seeks to accurately predict future events given predictions from experts. For settings where the number of experts grows to infinity, we give results about different settings where experts ought to be truthful. We then also relax the assumption that experts must be fully rational and explore whether experts should be truthful in this setting as well. Second, we provide new tools that organizers and architects of prediction models can use to build better models. To do so, we explore forecasting competitions, where forecasters of events compete to predict future events. Following the important results of Witkowski et al., we generalize their setting where events are independent, to settings where events may be correlated. We show that given their forecasting competition mechanism, called Event-Lotteries Forecasting Competition (ELF), forecasters ought to be truthful in settings of two arbitrarily correlated events. We also show that for any three correlated events, ELF does not properly incentivize forecasters to be truthful. We then introduce a new mechanism called Bayesian Network ELF (BNELF) that properly incentivizes forecasters to be truthful for any setting of arbitrarily correlated events and chooses the most accurate forecaster given enough events. Notably, BNELF relaxes the assumption that forecasters cannot update their beliefs of future events, making our mechanism even more applicable to real world settings. Therefore, we build on previous results in the literature and bridge the gap between prediction models and correlated events. Finally, we provide an application of prediction models in the field of hepatopathology, the study of the diagnosis of diseases in the liver. We introduce a stable and foundational diagnosis prediction model using Voronoi diagrams for clinical care in hepatopathology, providing an environment where doctors can gain new insight on the spread of diseases in the liver, as well as a systematic way to convey these new findings to other medical experts. We strengthen our results with empirical and histological data. Location: DIMACS Seminar Room, CoRE 431 Committee: Professor David Pennock (Advisor/Chair) Professor Lirong Xia Assistant Professor Xintong Wang Professor Rafael (Raf) M. Frongillo (external) |
| Start Date: 25 Mar 2026; Start Time: 05:00PM - 07:00PM Title: Beyond the Classroom: Careers in Computer Science Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 26 Mar 2026; Start Time: 10:30AM - 12:00PM Title: Beyond One Size Fits All: Rethinking LLM Alignment Bio: Cristina Garbacea is a Postdoctoral Scholar at the Data Science Institute at the University of Chicago. She received her PhD in Computer Science and Engineering from the University of Michigan and an MSc in Artificial Intelligence from the University of Amsterdam. Her research focuses on the alignment, personalization, and interpretability of large language models (LLMs), with the goal of building AI systems that are reliable, transparent, and aligned with diverse human preferences. Her work spans preference learning, human-centered evaluation, and the safety of generative AI systems. Cristina's research has been published at leading venues including NeurIPS, ICML, ACL, EMNLP, and TMLR. She has completed five research internships at Microsoft Research and Google DeepMind. Her long-term research vision is to develop trustworthy AI systems that are capable, transparent, and safe for real-world deployment. Speaker: Abstract: Large language models are typically aligned to an “average user,” but no such user exists. People differ in values, expertise, and intent; as a result, a response that is helpful for one person can be unsafe or ineffective for another. This creates a crisis of the average: population-level alignment methods and aggregate benchmarks can systematically obscure failures that matter for real users.In this talk, I propose adaptive alignment: scalable, personalized, and evaluable alignment for LLM assistants. I first present BoNBoN, which shows that Best-of-N selection is optimal under a win-rate–KL tradeoff, and trains models to match Best-of-N alignment at 1x inference cost while preserving off-target capabilities. I then introduce HyPerAlign, which personalizes model behavior by inferring interpretable, human-editable hypotheses about a user and conditioning generation on them, enabling transferable personalization without fine-tuning. Finally, I propose Personalized Benchmarking, demonstrating that individualized evaluation can fundamentally change our conclusions about alignment, and I introduce protocols that directly measure preference heterogeneity. Together, these works lay foundations for AI systems that adapt to individual users, respond to changing contexts, and provide transparent reasoning - moving us toward alignment methods that better serve diverse human needs. Location: CoRE 301 Committee: |
| Start Date: 26 Mar 2026; Start Time: 04:00PM - 06:00PM Title: Conceptual Explanations for Vision and Language Foundation Models Bio: Speaker: Abstract: Vision and language foundation models, including Vision Transformers, Pretrained Language Models, and Multimodal Large Language Models, have achieved remarkable success across a wide range of applications. However, their increasing complexity raises critical challenges in interpretability, trustworthiness, and evaluation, particularly in understanding model behavior at a conceptual level across modalities. In this dissertation, we develop a unified framework for conceptual explanations and evaluation of foundation models, bridging the gap between model representations and human-understandable concepts. We first propose variational Bayesian approaches for post-hoc conceptual explanation, including Probabilistic Concept Explainers for vision models and Variational Language Concepts for language models, which model latent distributions of visual and textual embeddings to produce multi-level conceptual explanations at the dataset, instance, and token or patch levels, satisfying key desiderata such as faithfulness, stability, and sparsity. Empirical results demonstrate that these methods consistently outperform existing approaches in generating trustworthy explanations. Beyond interpretability, we further introduce the MultiModal Needle-in-a-Haystack benchmark to systematically evaluate the long-context reasoning and retrieval capabilities of multimodal foundation models, revealing significant limitations such as performance degradation in long-context settings and hallucination behaviors. Overall, this dissertation advances the understanding of foundation models through conceptual representations and rigorous evaluation, contributing toward the development of trustworthy multimodal AI systems. Location: CoRE 301 Committee: Assistant Professor Hao Wang (Advisor/Chair) Assistant Professor Lesia Semenova Assistant Professor Chengzhi Mao Hao Ding (External Member) |
| Start Date: 27 Mar 2026; Start Time: 02:00PM - 03:00PM Title: Computer Science Department Tour - Prospective Students Bio: Speaker: Abstract: Location: CoRE 301 Committee: |
| Start Date: 02 Apr 2026; Start Time: 10:30AM - 12:00PM Title: AI for Human Agency and Pluralism in Society: from Data to Systems Bio: Jad Kabbara is a research scientist at MIT working with Prof. Deb Roy at the Media Lab and the Center for Constructive Communication (CCC). His broad interests are in Artificial Intelligence and Natural Language Processing and their applications to address important societal problems. He received his PhD in Computer Science in 2022 from McGill University & Mila where he was advised by Prof. Jackie Cheung. His work has appeared in top NLP, HCI and ML venues and has received recognitions including Tech Review Innovators Under 35 (MENA), MIT Generative AI Impact Award (2023), COLING 2022 Best Short Paper Award, ACL 2018 Best Paper Award. Jad’s work has been covered by the Washington Post, New York Times, Science, ACM News, IEEE Spectrum, Yahoo Finance, The Globe & Mail, Vox, VentureBeat, and MIT Technology Review. Prior to his PhD, he received a Masters from McGill University in 2014 and a Bachelors from the American University of Beirut in 2011. Speaker: Abstract: The unprecedented advances in AI have sparked widespread debate. Many worry that LLMs will replace their jobs, while others wonder when LLMs will fully automate certain tasks. My research shifts the focus and asks: How can we design AI systems that assist human processes¬, not replace them, in an intentional way that preserves human agency and promotes a pluralistic and equitable society? The talk will cover three core dimensions for answering this question. First, in contrast to many applications of modern AI that reinforce or amplify societal inequities, I have designed human-AI systems that help level the playing field in critical domains such as law. Second, I will describe our work investigating social and political biases in LLMs and discuss methods to mitigate LLM bias. Third, I will present work examining the data foundations of modern AI, analyzing how data practices and representation shape system behavior and downstream societal impacts, and presenting tools and policy guidelines that empower different stakeholders and address some of these challenges. Across these areas, my work combines techniques from NLP, HCI, and responsible AI to build systems that are both technically robust and socially grounded. Finally, I will outline my research vision regarding future opportunities and challenges in developing AI systems that augment human capabilities while promoting fairness and accountability. Location: CoRE 301 Committee: |
| Start Date: 02 Apr 2026; Start Time: 02:00PM - 03:00PM Title: Dense Reward Shaping for Reinforcement Learning with LTL Specifications Bio: Speaker: Abstract: Reinforcement learning (RL) with Linear Temporal Logic (LTL) specifications enables rich, long-horizon objectives, but rewards are typically sparse, often provided only upon satisfaction of the specification, which leads to weak learning signals, poor exploration, and severe long-horizon credit assignment challenges. In this talk, we present an LTL-aware reward design framework that exploits the structure of limit-deterministic Büchi automata (LDBAs) to provide dense, informative feedback for policy learning. By encoding incremental progress toward temporal logic satisfaction into the reward, our approach transforms rare acceptance signals into structured supervision that guides exploration and supports long-horizon optimization. The framework employs a staged reward formulation that autonomously balances exploration and exploitation, steering agents toward accepting behaviors without additional supervision. Experiments on high-dimensional continuous-control tasks demonstrate improved learning efficiency, robustness, and satisfaction of LTL specifications compared to prior reward designs.Based publications: https://openreview.net/pdf?id=MKLcSmSTQI Location: CoRE 301 Committee: Professor He Zhu Professor Emily First Professor Yongfeng Zhang Professor Sumegha Garg |
| Start Date: 03 Apr 2026; Start Time: 10:30AM - 12:00PM Title: Exploring Away from Home: Robust Perception for Planetary Exploration and Scientific Discovery Bio: Georgios Georgakis is a Robotics Technologist at the Jet Propulsion Laboratory, California Institute of Technology. His research interests lie at the intersection of computer vision and machine learning that have applications in robotic perception and planetary exploration. Georgios has won competitive NASA research funding awards, and his work has impacted several space applications including developing technology for a future Mars helicopter, improving Lunar Terrain Relative Navigation, and creating a learning-based photometric calibration approach for Juno's camera. He is the recipient of multiple Team and Technology Spotlight awards from the Autonomous Systems and Robotics section at JPL. Before joining JPL, he was a Postdoctoral Researcher and Part-time Lecturer in the GRASP lab at the University of Pennsylvania, where he worked on vision-based navigation for indoor semantic tasks. He earned his PhD in Computer Science from George Mason University in 2020. Speaker: Abstract: Planetary exploration demands autonomous systems that can operate reliably in extreme, uncertain, and data-scarce environments far from Earth. Communication delays, limited onboard resources, and heterogeneous sensing make traditional, heavily supervised robotic pipelines difficult to scale to future missions. This talk presents methods for robust, adaptive perception designed to enable the next generation of autonomous planetary exploration and scientific discovery. I will highlight three key contributions. First, a learning-based localization method that enables long-range navigation by bridging large gaps between orbital and onboard imagery. Second, an adaptive depth perception technique that provides dense geometric understanding for safer ground mobility. Third, a data-driven approach to cross-sensor calibration that transforms uncalibrated spacecraft imagery into scientifically meaningful data products, enabling new atmospheric analyses of Jupiter. Together, these efforts advance a shift from carefully scripted robotic operations toward adaptive autonomy, reducing reliance on privileged data and large operations teams while expanding the scientific return of planetary missions. Location: CoRE 301 Committee: |
| Start Date: 07 Apr 2026; Start Time: 10:30AM - 12:00PM Title: Verifiable PDE Reasoning and Modeling with Neurosymbolics Bio: Dr. Wuyang Chen is a tenure-track Assistant Professor in Computing Science at Simon Fraser University. He is also a visiting research scientist at Microsoft. Previously, he was a postdoctoral researcher in Statistics at the University of California, Berkeley, advised by Professor Michael Mahoney. He obtained his Ph.D. in Electrical and Computer Engineering from the University of Texas at Austin in 2023, advised by Professor Atlas Wang. Dr. Chen's research focuses on integrating AI methods with physical knowledge, scientific machine learning, and theoretical understanding of deep networks. Dr. Chen has published papers at CVPR, ECCV, ICLR, ICML, NeurIPS, and other top conferences. Dr. Chen's research has been recognized by the US NSF newsletter, two Doctoral Dissertation Awards from INNS and iSchools, AAAI New Faculty Highlights, and NVIDIA Academic Grant Award. Dr. Chen also hosted and co-organized many conference workshops at NeurIPS, ICLR, CVPR. Speaker: Abstract: Recent progress in Large Language Models (LLMs) has transformed text and code generation, yet models still falter on Partial Differential Equations (PDEs) where correctness, constraints, and physical consequences are critical. This talk explores how formal LLM reasoning can advance symbolic PDE modeling. First, our PDE-Controller formalizes informal PDEs, synthesizes solver-ready code, and plans subgoals to tackle nonconvex control via interactions with external solvers. Second, our Lean Finder accelerates PDE formalization via a semantics-aware search engine for Lean/Mathlib that retrieves relevant theorems, outperforming GPT models and gaining significant traction in the AI-for-math community. Through these efforts, we aim to design a semantics-first LLM that autoformalizes informal PDE problems into machine-checked specifications and synthesizes solver-ready code. This closes the loop between formal analysis and LLM reasoning, ultimately surpassing human heuristics across diverse PDEs. Location: CoRE 301 Committee: |
| Start Date: 08 Apr 2026; Start Time: 10:00AM - 11:30AM Title: Evolving Multimodal Agents through Knowledge-grounded Simulation and Synthesis Bio: Speaker: Abstract: The development of Large Language Models (LLMs) catalyzes the long-term planning and decision-making capabilities of multimodal agents in sophisticated digital environments (such as computer-use scenarios). Multimodal agents navigate digital environments by perceiving visual states and executing discrete actions to fulfill complex user intents. However, the advancement of these agents is severely bottlenecked by a reliance on human-in-the-loop supervision, which is expensive, difficult to scale, and insufficient for capturing the diversity of real-world interaction traces. We argue that multimodal agents are able to continuously evolve with limited human supervision if their knowledge can be well grounded to specific environments. Specifically, we first propose the Retrieval-augmented World Model (R-WoM), a framework that enables evolution by grounding agent simulations in external, environment-specific tutorials. This grounding stabilizes long-horizon reasoning and reduces compounding hallucinations, yielding relative performance improvements of up to 23.4% on realistic benchmarks (OSWorld and WebArena). Second, we introduce GUIDED, which drives the evolution of multimodal perception and action through the self-distillation of internal coding knowledge into executable trajectories and programmatic verifiers. By utilizing Group Relative Policy Optimization (GRPO) on these synthesized signals, the agent bootstraps its own performance, surpassing human-annotated baselines in efficiency and success of task completion. Moving forward, we plan to extend this evolving paradigm by continuously refining the world models of multimodal agents to help them adapt to more unseen environments. Location: CBIM 22 Committee: Professor Dimitris N. Metaxas Professor Hongyi Wang Professor Ruixiang Tang Professor Mingsum Kim |
| Start Date: 08 Apr 2026; Start Time: 05:00PM - 07:00PM Title: Beyond the Classroom: Careers in Computer Science Series Bio: Speaker: Abstract: Location: Busch Student Center -BSC Center Hall Committee: |
| Start Date: 09 Apr 2026; Start Time: 10:30AM - 12:00PM Title: Your host is a distributed system! Bio: Midhul Vuppapalapati is a PhD student in the Computer Science department at Cornell University, advised by Prof. Rachit Agarwal. He is a recipient of the Cornell University Fellowship, a SIGCOMM Best Student Paper award, SIGMETRICS Best-of-the-rest recognition, and two Cornell CS Outstanding TA awards Speaker: Abstract: The host architecture has evolved rapidly over the past decade: modern hosts have multiple heterogeneous compute devices, multiple memory devices potentially distributed across multiple cache coherence domains, and multiple network devices, all interconnected by an intra-host network. Such a host architecture alters the many fundamental assumptions that once guided the design and optimization of modern operating systems. As a result, applications running atop modern hosts suffer from suboptimal properties and performance, while host resources remain extremely underutilized.In this talk, I’ll outline my research vision: next-generation operating and distributed systems should be designed with the perspective of each individual host being a distributed system itself! Using examples from my thesis research, I will discuss the many ways in which “host as a distributed system” perspective compels us to revisit classical problems in operating systems, as well as enables green field explorations across the systems stack. Location: CoRE 301 Committee: |
| Start Date: 13 Apr 2026; Start Time: 10:30AM - 12:00PM Title: From raw reads to biological discovery: optimizing across the genomics pipeline Bio: Tavor Baharav is a postdoctoral fellow at the Eric and Wendy Schmidt Center at the Broad Institute, working with Rafael Irizarry. His research co-designs the machine learning pipeline for computational genomics, jointly optimizing upstream processing stages with downstream inference. He is broadly interested in high-dimensional statistics, adaptive algorithms, statistical machine learning, and their application to problems in computational genomics. Speaker: Abstract: Modern genomics produces massive, high-dimensional datasets, yet extracting reliable biological insight remains challenging. A central challenge, and opportunity, in genomics is that we do not directly receive a sanitized data matrix: our machine learning pipeline starts upstream, with what samples we choose to collect, how we measure them, and how we transform raw signals before inference. These stages are typically handled in isolation, quietly introducing bias and discarding information, and limiting discovery potential. My research addresses this by treating genomics as an end-to-end system, developing rigorous machine learning methods that account for, and leverage, these upstream choices.In this talk, I will illustrate this approach through my work on reference-free genomic analysis. Alignment of reads to a reference genome, though ubiquitous, fundamentally limits discovery of novel biology that deviates from the reference. To overcome this, we developed SPLASH, a statistical tool that compares raw sequencing reads directly across conditions. SPLASH rediscovers strain-defining mutations in SARS-CoV-2 and identifies previously unannotated tissue-specific transcripts in octopus, enabling discovery without any reference or annotation. Bypassing alignment reshapes the statistical problem: to identify genomic features of interest, we developed a new statistical test for contingency tables. Aggregating information across the resulting data matrices raised broader methodological and theoretical questions about data integration, leading us to develop a random matrix theory framework for detecting shared structure across datasets. Together, these results show how rethinking upstream pipeline choices can simultaneously improve biological discovery and yield generalizable statistical insights. Location: CoRE 301 Committee: |
| Start Date: 15 Apr 2026; Start Time: 12:30PM - 02:00PM Title: Multi-Agent Autonomy: From Interaction-Aware Navigation to Coordinated Mobile Manipulation Bio: Javier Alonso-Mora is a Full Professor in the Cognitive Robotics Department at Delft University of Technology, where he leads the Autonomous Multi-Robot Systems Lab and serves as Director of the TU Delft Robotics Institute. He received his Ph.D. from ETH Zurich, in partnership with Disney Research Zurich, and was a Postdoctoral Associate at the Massachusetts Institute of Technology. He co‑chairs the IEEE Robotics and Automation Society Technical Committee on Multi‑Robot Systems and serves as an Associate Editor for IEEE Transactions on Robotics and Autonomous Robots, as well as for several major robotics conferences. His research focuses on autonomous mobile robotics, with an emphasis on navigation, motion planning, learning, and control. Key application areas include mobile manipulation, autonomous vehicles, aerial robotics, on‑demand mobility, and multi‑robot coordination in dynamic, human‑shared environments. His awards include an NWO VENI Talent Scheme Award (2017), the ICRA Best Paper Award on Multi‑Robot Systems (2019), an ERC Starting Grant (2021), the IEEE Transactions on Automation Science and Engineering Best Paper Award (2024), and the IEEE Intelligent Transportation Systems Young Researcher Award (2025). His research on ride‑pooling has led to the creation of a commercial spin‑off company, The Routing Company. Speaker: Abstract: In the pursuit of scalable, socially aware, and safety-critical autonomous systems, our recent research has focused on integrating learning, planning, and control across aerial, ground, and maritime robotic platforms. Central to this effort is the fusion of model-based and data-driven approaches, enabling robust decision-making in dynamic and uncertain environments, seamless multi-robot coordination, and the ability to learn from human demonstrations. This talk will highlight recent advances in three key areas: 1) interaction-aware navigation among other robots and humans, using sampling-based model predictive control, socially compliant behavior learning, and semantic mapping; 2) real-time task and motion planning for teams of mobile manipulators through expert demonstrations, physically grounded plans, and whole-body control; and 3) decentralized 6-DoF manipulation of cable-suspended loads by a team of drones using multi-agent reinforcement learning. These contributions advance the frontier of scalable autonomy in dynamic, multi-agent environments across diverse robotic platforms. Location: Room 403, 4th floor, 1 Spring Street, Downtown New Brunswick Committee: |
| Start Date: 20 Apr 2026; Start Time: 10:30AM - 12:00PM Title: Evaluating LLMs When They Do Not Know the Answer: Statistical Evaluation of Mathematical Reasoning via Comparative Signals Bio: Linjun Zhang is an Associate Professor in the Department of Statistics, at Rutgers University. He obtained his Ph.D. in Statistics at the Wharton School, the University of Pennsylvania in 2019, and received J. Parker Bursk Memorial Prize and Donald S. Murray Prize for excellence in research and teaching, respectively upon graduation. He also received the NSF CAREER Award, Rutgers Presidential Teaching Award in 2024, and the Warren I. Susman Award for Excellence in Teaching in 2025. His current research interests include statistical foundations of large language models, algorithmic fairness, privacy-preserving data analysis, and deep learning theory. Speaker: Abstract: Evaluating mathematical reasoning in LLMs is constrained by limited benchmark sizes and inherent model stochasticity, yielding high-variance accuracy estimates and unstable rankings across platforms. On difficult problems, an LLM may fail to produce a correct final answer, yet still provide reliable pairwise comparison signals indicating which of two candidate solutions is better. We leverage this observation to design a statistically efficient evaluation framework that combines standard labeled outcomes with pairwise comparison signals obtained by having models judge auxiliary reasoning chains. Treating these comparison signals as control variates, we develop a semiparametric estimator based on the efficient influence function (EIF) for the setting where auxiliary reasoning chains are observed. This yields a one-step estimator that achieves the semiparametric efficiency bound, guarantees strict variance reduction over naive sample averaging, and admits asymptotic normality for principled uncertainty quantification. Across simulations, our one-step estimator substantially improves ranking accuracy, with gains increasing as model output noise grows. Experiments on GPQA Diamond, AIME 2025, and GSM8K further demonstrate more precise performance estimation and more reliable model rankings, especially in small-sample regimes where conventional evaluation is pretty unstable. If time permits, we will also talk about how the AI can transform the statistics community in general. Location: CoRE 301 Committee: |
| Start Date: 21 Apr 2026; Start Time: 10:30AM - 12:00PM Title: Toward Trustworthy Agentic Systems for Smart Spaces Bio: Jorge Ortiz is an Associate Professor at SOE Rutgers and Director, Sensing & Reasoning Lab Speaker: Abstract: Sensor-rich environments are moving from passive monitoring toward software systems that observe, reason, and act. This shift raises a core systems question about how to build AI-enabled infrastructure that remains trustworthy when it combines sensitive sensing, multi-step coordination, and real-world actuation. This talk outlines a research agenda organized around three problems, namely privacy-conscious data access for sensing-rich applications, durable coordination for long-running agentic computation, and verification-based debugging of coordination failures in multi-agent systems. I will describe ongoing efforts in each of these areas and argue that, taken together, they point toward a broader systems foundation for trustworthy smart spaces. Location: CoRE 301 Committee: |
| Start Date: 22 Apr 2026; Start Time: 02:00PM - 04:00PM Title: Failure Forecasting for Robust Sim-to-Real Rhythmic Insertion Policies Bio: Speaker: Abstract: Many real-world manipulation tasks require a robot to perform the same precise insertion motion repeatedly, where even small errors can accumulate and cause failure over time. In this qualifying exam, I will present our work on rhythmic insertion, using the task of screwing a nut onto a bolt with a wrench as a representative example. The talk will describe a sim-to-real framework that combines a reinforcement learning insertion policy with a learned failure-forecasting module that predicts unsuccessful executions and triggers recovery. By using an object-centric pose representation, the method improves transfer from simulation to the real world and achieves robust performance in repeated long-horizon insertion tasks. More information can be found in the website: https://jaysparrow.github.io/rit/ Publications: “Failure Forecasting Boosts Robustness of Sim2Real Rhythmic Insertion Policies”, IROS 2025 Location: Room 403, 4th floor, 1 Spring Street, Downtown New Brunswick Committee: Professor Abdeslam Boularias Professor Kostas Bekris Assistant Professor He Zhu Assistant Professor Yipeng Huang |
| Start Date: 22 Apr 2026; Start Time: 03:00PM - 05:00PM Title: Correctly Rounded Dot Products under Round-to-Odd Bio: Speaker: Abstract: This talk presents a new algorithm for computing correctly rounded dot products using floating-point (FP) operations under the non-standard round-to-odd (RO) rounding mode. Our algorithm applies error-free transformations (EFT) for FP multiplication to reduce dot products to summations over vectors. By leveraging EFTs for FP addition tailored to RO, we implement an algorithm that produces faithfully rounded sums over vectors under RO. Using the faithfully rounded sum and the properties of RO, our dot product algorithm subsequently determines the correctly rounded result. Through correct rounding, our algorithm enables accurate dot products and matrix multiplications for future hardware that natively supports RO. Location: CoRE 305 Committee: Professor Santosh Nagarakatte Professor Rich Martin Associate Professor Mridul Aanjaneya Assistant Professor Arpita Biswas |
| Start Date: 27 Apr 2026; Start Time: 02:00PM - 04:00PM Title: Reliable Retrieval and Reasoning for Knowledge-Intensive AI Bio: Speaker: Abstract: As AI systems are increasingly deployed in knowledge-intensive settings, their performance depends not only on generation ability, but also on whether they can reliably identify, retrieve, and reason over the right external resources. In practice, however, reliable retrieval and reasoning remain difficult because current systems still lack robust mechanisms for acquiring, organizing, and validating external knowledge under complex real-world conditions. We first study tool recommendation, where the goal is to provide AI systems with a task-aligned and sufficient set of external tools rather than an overinclusive ranked list, improving both efficiency and downstream problem solving. We then turn to reasoning over knowledge graphs, proposing a framework that goes beyond rigid traversal over explicit edges and instead supports more robust reasoning in noisy and sparse graphs through guided navigation and similarity-enhanced expansion. These challenges further motivate a closer examination of the retrieval process itself. In response, we investigate semantic shift and show that retrieval degradation is driven not simply by text length, but by the structured evolution and dispersion of meaning within text, which weakens the discriminative quality of pooled embeddings. Building on this view of retrieval reliability, we extend the discussion to multimodal retrieval-augmented generation through CLIMB, a train-free framework that organizes complementary evidence and performs confidence-guided refinement so that answer updates are accepted only when grounded reliability improves. Taken together, these works advance a broader vision of reliable knowledge-intensive AI. Location: CBIM 22 Committee: Professor Dimitris N. Metaxas Professor Konstantinos Michmizos Professor Hongyi Wang Professor Peng Zhang |
| Start Date: 29 Apr 2026; Start Time: 02:00PM - 04:00PM Title: Algorithmic Foundations for Efficient and Optimal Motion and Task-Motion Planning Bio: Speaker: Abstract: We present two approaches that address distinct but complementary challenges in motion planning. First, we introduce the Rotation-Stacked Visibility Graph (RVG), which generalizes reduced visibility graph methods to include discretized robot orientations. By stacking layered roadmaps, RVG achieves resolution-completeness and asymptotic optimality for SE(2) planning, significantly outperforming sampling-based alternatives in both computational efficiency and path quality. Second, we present the Synchronous Dual-Arm Rearrangement (SDAR) planner, which tackles the complexity of tightly coupled multi-robot coordination. By integrating dependency-driven task logic with a synchronous motion planner, SDAR effectively navigates the interleaved constraints of tabletop rearrangement. Collectively, these methods demonstrate a robust path toward achieving real-time, optimal performance in complex, constrained environments. Location: Room 403, 4th floor, 1 Spring Street, Downtown New Brunswick Committee: Associate Professor Jingjin Yu Professor Abdeslam Boularias Professor Kostas Bekris Professor Lirong Xia |
| Start Date: 30 Apr 2026; Start Time: 11:00AM - 12:30PM Title: LLM Enhanced Open-Vocabulary Object Detection without Human Curated Data Generation Bio: Speaker: Abstract: Large foundation models trained on large-scale vision-language data can boost Open-Vocabulary Object Detection (OVD) via synthetic training data, yet hand-crafted pipelines often introduce bias and overfit to specific prompts. We sidestep this issue by directly fusing hidden states from Large Language Models (LLMs) into detectors—an avenue that remains surprisingly under-explored. This paper presents a systematic method to enhance visual grounding by utilizing decoder layers of the LLM within an MLLM. We introduce a zero-initialized cross-attention adapter to enable efficient knowledge fusion from LLMs to object detectors, a new approach called LED (LLM Enhanced Open-Vocabulary Object Detection). We find that intermediate LLM layers already encode rich spatial semantics, and adapting only the early layers yields most of the gain. With Swin-T as the vision encoder, Qwen2-0.5B + LED improves Grounding-DINO by 3.82% on OmniLabel at only 8.7% additional GFLOPs, and a larger vision backbone increases the improvement to 6.22%. Extensive ablations on adapter variants, LLM scales, and fusion depths further corroborate our design. Location: CoRE 305 Committee: Professor Dimitris Metaxas Professor Desheng Zhang Assistant Professor Hongyi Wang Assistant Professor Ruixiang Tang |
| Start Date: 05 May 2026; Start Time: 11:00AM - 12:00PM Title: Advances in Watermarking Large Language Models Bio: Yu-Xiang Wang is an Associate Professor at the Halıcıoğlu Data Science Institute of UC San Diego, also affiliated with the Department of Computer Science and Engineering. He runs the Scalable Statistical Machine Learning lab. He is also an Amazon Scholar. Yu-Xiang received his PhD in Statistics and Machine Learning in 2017 from Carnegie Mellon University (CMU). His research interests include the foundation of machine learning, large-scale optimization, reinforcement learning, online learning, deep learning and differential privacy. His work had been supported by an NSF CAREER Award, Research Award from Amazon, Google, Adobe, OpenAI, JPMorgan-Chase, and he had won paper awards from KDD'15, WSDM'16, AISTATS'19 and COLT'21. Speaker: Abstract: As the digital age progresses, artificial intelligence (AI), especially in the form of large language models (LLMs) like ChatGPT, has become increasingly influential in our daily lives. These technologies have the power to write essays, generate news articles, and even create realistic conversations. However, this capability also presents risks such as the generation of misinformation, academic dishonesty, and cybersecurity threats. A promising recent approach to mitigate these challenges is “watermarking”. By injecting subtle statistical signals to the LLM-generation process, one can reliably detect AI-generated texts or event attribute downstream models trained using these texts. In this talk, I will talk about existing attempts in formalizing this problem and discuss two recent work of ours: (1) Unigram (Green-Red) watermark https://arxiv.org/abs/2306.17439; (2) Permute-and-Flip watermark https://arxiv.org/abs/2402.05864. If time permits, I may talk about recent work on watermarking models and watermarking training data. Location: CoRE 301 Committee: |
| Start Date: 05 May 2026; Start Time: 11:20AM - 12:20PM Title: Trustworthy AI for Structured Reasoning: Conformal Guarantees in Knowledge Graph Question Answering Bio: Speaker: Abstract: Knowledge Graph Question Answering (KGQA) enables interpretable multi-hop reasoning, yet existing approaches lack rigorous reliability guarantees for their predictions. This research investigates trustworthy reasoning over knowledge graphs by integrating conformal prediction, a distribution-free framework for uncertainty quantification, into structured graph reasoning. The core challenge lies in reconciling conformal prediction's exchangeability assumption with the sequential dependencies inherent in multi-hop graph traversal. We propose a path-level calibration framework that restores valid statistical guarantees while maintaining compact, actionable answer sets. Our approach combines principled calibration theory with learned scoring mechanisms to bridge the gap between formal coverage guarantees and practical KGQA performance.Publications: 1. S. Lin*, C. Zhou*, et al. "Conformal Path Reasoning: Trustworthy Knowledge Graph Question Answering via Path-Level Calibration." Under Review at ICML 2026. 2. F. Nie, X. Hou, S. Lin, et al. "FactTest: Factuality Testing in Large Language Models with Statistical Guarantees." ICML 2025. 3. T. Wang, D. Gui, Y. Hu, S. Lin, et al. "MPO: Efficient Post-Processing for Mixing Diverse Preference Alignment." ICML 2025. Location: CBIM 22 Committee: Professor Dimitris N. Metaxas Assistant Professor Ruixiang Tang Assistant Professor Hongyi Wang Assistant Professor He Zhu |
| Start Date: 06 May 2026; Start Time: 03:30PM - 04:30PM Title: Towards Universal and Interactive Medical Image Segmentation Bio: Speaker: Abstract: Medical image segmentation remains a fundamental yet challenging problem due to the diversity of anatomical structures, imaging modalities, and task settings. Existing methods are often designed for a single segmentation paradigm, which limits their generalization and practical utility. We develop a series of models for universal and interactive medical image segmentation, aiming to support diverse segmentation scenarios within a unified framework. By integrating flexible prompting and interaction mechanisms, the proposed approach improves adaptability, robustness, and scalability across a wide range of medical imaging applications. Experimental results on multiple datasets demonstrate competitive performance across different segmentation settings and highlight the potential of a unified framework for general-purpose medical image segmentation. Location: CoRE 305 Committee: Professor Dimitris Metaxas Assistant Hongyi Wang Assistant Ruixiang Tang Professor Badri Nath |
| Start Date: 06 May 2026; Start Time: 05:00PM - 06:00PM Title: Understanding and Controlling Internal Representations in Large Language Models Bio: Speaker: Abstract: I focus on understanding how large language models encode, organize, and utilize internal representations for knowledge acquisition, out-of-distribution generalization, and controllable generation. The exam is based on three recent works. The first studies how representation sparsity changes under distribution shift and shows that harder or more out-of-distribution inputs tend to induce sparser hidden representations, revealing a potential mechanistic signature of model uncertainty and failure. The second investigates how concepts emerge across layers in large language models, providing evidence that different layers capture knowledge and abstraction at different depths. The third explores how sparse representation spaces derived from sparse autoencoders can be used for reliable supervised steering, enabling more precise and interpretable control of model behavior. Together, these works advance a unified view that the geometry and sparsity of internal representations play a central role in both understanding and controlling large language models. Location: CBIM 22 Committee: Professor Dimitris N. Metaxas Assistant Professor Hongyi Wang Assistant Professor Chengzhi Mao Associate Professor Sudarsun Kannan |
| Start Date: 12 May 2026; Start Time: 02:30PM - 04:30PM Title: Enhancing Consistency Models for Multi-Agent Trajectory Prediction Bio: Speaker: Abstract: Diffusion models have achieved impressive results in many prediction tasks. However, their performance is limited by iterative denoising, which leads to inference that is prohibitively slow in time-critical settings, such as autonomous driving. Many alternatives which alleviate this issue have been developed. Among these alternatives, consistency models (CMs) stand out for their capability to generate high-quality samples in a single step. Despite this, CMs are often difficult to train from scratch. We propose ECTraj, an enhanced CM-based trajectory prediction pipeline with an improved training objective. In particular, ECTraj extends the standard student-teacher consistency training scheme: the student produces standard outputs (future trajectories), while the teacher explicitly fuses its predictions with parts of the ground truth future trajectories to achieve stronger supervision. Additionally, we exploit the direct denoising capability of CMs for top-K multi-shot generation during training. Combining conditional generation with this enhanced consistency objective results in faster inference and improved prediction accuracy on the large-scale Argoverse 2 dataset. Location: CBIM 22 Committee: Professor Vladimir Pavlovic Professor Chengzhi Mao Professor Hao Wang Professor Minesh Patel |
| Start Date: 13 May 2026; Start Time: 01:00PM - 02:00PM Title: Towards Efficient Multi-LLM Collaborative Debate via Reinforcement Learning Bio: Speaker: Abstract: LLM-based Multi-Agent Debate (MAD), a test-time scaling method featuring "cognitive inflection points," is a prominent research area. By leveraging the complementary knowledge and reasoning of diverse LLMs, MAD often outperforms single models in complex tasks. However, it faces three key challenges: (1) the necessity of multi-round debate is controversial, as it frequently fails to surpass communication-free collaboration; (2) error propagation and accumulation can lead to a single agent misleading the collective; and (3) redundant communication flows increase computational overhead. The aforementioned drawbacks of MAD highlight the critical need for pruning communication flows and an effective pruning algorithm should enhance collective reasoning while reducing token overhead. However, existing research suffers from several limitations: (1) a lack of rigorous comparison against strong baselines like communication-free frameworks, often coupled with overoptimistic estimates of token efficiency ratios; (2) either proposes static optimization frameworks or relies merely on simple LLM role-based profiles for dynamic pruning, failing to fully leverage the specific details within multi-round debates; and (3) a lack of RL rewards and algorithms specifically tailored for flow pruning in multi-round scenarios. Consequently, current frameworks fail to demonstrate the inherent superiority of debate-based collaboration, and suboptimal RL pipelines leave the potential of MAD largely untapped. To address the above challenges, we propose a lightweight pruning model, trained via an innovative multi-round debate reward. Experimental results show that our framework achieves SOTA performance on reasoning benchmarks such as MATH-500, while significantly reducing token costs under fair and rigorous evaluation. Location: CBIM 22 Committee: Professor Dimitris Metaxas Assistant Professor Hongyi Wang Associate Professor Konstantinos Michmizos Assistant Professor Sharon Levy |
| Start Date: 15 May 2026; Start Time: 01:00PM - 03:00PM Title: Towards Compliant Human Trajectory Prediction: From Environment-Aware Generation to Multi-Agent Joint Alignment Bio: Speaker: Abstract: Human trajectory prediction is inherently stochastic, requiring models to generate diverse future behaviors from limited observations. While recent advances in generative modeling have significantly improved predictive accuracy, they often overlook an equally critical requirement: compliance. In practice, predicted trajectories must not only match ground truth but also adhere to environmental constraints and remain socially feasible among interacting agents. In this talk, I present two complementary approaches that address this challenge. First, I introduce TrajDiffuse, a guided conditional diffusion framework that formulates trajectory prediction as a denoising process conditioned on agent intent and scene context, enabling explicit control over generation and strong adherence to environmental constraints . Second, I present JACoP, a multi-stage framework that enforces scene-level consistency by selecting environment-compliant trajectory prototypes and jointly aligning multi-agent predictions through a Markov Random Field formulation .Together, these works demonstrate that incorporating structured guidance—both at the single-agent generation level and at the multi-agent interaction level—can significantly improve the realism and usability of trajectory predictions. I will conclude by discussing insights on balancing accuracy, diversity, and compliance, and outline future directions toward scalable, reliable prediction models for real-world applications. Location: CoRE 305 Committee: Professor Vladimir Pavlovic Associate Professor Mubbasir Kapadia Associate Professor Jingjin Yu Assistant Professor Karthik Srikanta |
| Start Date: 21 May 2026; Start Time: 02:00PM - 03:00PM Title: Towards Reliable and Interpretable Multimodal Intelligence Bio: Speaker: Abstract: Large vision-language models (LVLMs) have demonstrated strong multimodal capabilities, yet the internal mechanisms underlying their visual reasoning remain poorly understood, limiting both their reliability and controllability. Our research investigates LVLM visual reasoning from a mechanistic interpretability perspective, with a focus on two representative phenomena: object hallucination and visual counting. First, we show that object hallucinations are closely related to a small subset of image tokens that receive disproportionately high attention during generation. Based on this observation, we identify Hallucinatory Image Tokens (HITs) and propose EAZY, a training-free framework that detects and mitigates hallucinated objects by zeroing out token-level visual causes. Second, we study counting as a minimal yet revealing probe of visual reasoning, and introduce Visual Activation Patching and HeadLens to trace how visual information is grounded, routed across modalities, and aggregated into numerical predictions. This analysis reveals a structured counting circuit composed of visual grounding, cross-modal routing, counting aggregation, and awareness heads. Building on these findings, we develop lightweight interventions that improve both counting robustness and broader visual reasoning. Across experiments, our results show that LVLM failures and capabilities are both governed by sparse, structured, and interpretable internal mechanisms. These findings provide a unified account of how LVLMs process visual information, explain why they succeed or fail in reasoning-intensive settings, and demonstrate that mechanistic insights can directly enable more reliable and more capable multimodal models. Location: CoRE 305 Committee: Professor Vladimir Pavlovic Assistant Professor Ruixiang Tang Assistant Professor Chengzhi Mao Professor Zheng Zhang |
| Start Date: 26 May 2026; Start Time: 10:00AM - 12:00PM Title: From Modern Website Fingerprinting to Persona Fingerprinting: Realistic Threat Models, Behavior-Realistic Traffic Generation, and New Privacy Leakage from Encrypted Web Metadata Bio: Speaker: Abstract: Encrypted web traffic hides content but still leaks information through packet sizes, directions, and timing. Modern website fingerprinting around practical observation constraints and emerging privacy risks in three steps: (i) a boundary-free, site-level formulation that classifies fixed-length packet windows sampled from arbitrary points in long browsing sessions, with a systematic study of cross-environment degradation and robustness; (ii) evidence that data realism is a primary bottleneck—scripted automation produces brittle traces that fail to transfer to real users—followed by an LLM-driven multi-agent browsing framework that generates behavior-realistic traffic at scale via high-level decision-making and UI-grounded execution; and (iii) an extension from inferring visited sites to inferring browsing personas, showing that persona signals are learnable from encrypted metadata and can be amplified from standard WF representations via probing and multi-task learning. Results include >90% site identification in matched settings, sub-10% scripted→human transfer that recovers to ~80% with agent-generated traces, and ~84% mixed-site persona inference while retaining strong site accuracy (~93%). Location: CoRE 305 Committee: Professor Richard Martin Professor He Zhu Professor Hao Wang |
| Start Date: 27 May 2026; Start Time: 04:00PM - 05:00PM Title: A New SSSP of Graph and Geometric Data Representations: Structure, Sketching, Search, and Privacy Bio: Speaker: Abstract: Understanding complex data representations is a central challenge in modern algorithm design. I will present a collection of algorithmic advances across two domains: discrete graphs and continuous geometric spaces. At their core, both domains require us to efficiently compute and preserve distances, a task we approach through four algorithmic pillars: structure, sketching, search, and privacy. The first half explores the discrete combinatorics of shortest paths, using discrepancy bounds and hopsets to build differentially private graph algorithms and lower bounds. In the second half, I will introduce dimension reduction and hashing schemes that successfully preserve distance structures in non-Euclidean geometries. Location: CoRE 301 Committee: Professor Jie Gao Professor Mario Szegedy Professor Jalaj Upadhyay |
| Start Date: 28 May 2026; Start Time: 02:00PM - 04:00PM Title: Two Algorithms Pertaining to $s$ to $t$ Bio: Speaker: Abstract: This dissertation compiles together (with some revisions) the works \cite{ashvinkumar2025faster,ashvinkumar2026parallel} whose algorithms are concerned with $n$ vertex $m$ edge digraphs $G=(V,E)$ and paths within them (say from $s \in V$ to $t \in V$). We first discuss the restricted shortest paths problem where edges of $G$ are assigned two non-negative weights: lengths and costs. For a fixed source $s \in V$ and a cost threshold $C$ we wish to find, for each target $t \in V$, the length of the shortest $(s,t)$-path whose total cost is at most $C$. This problem is NP-hard --- for instance, the $0$-$1$ knapsack problem is a special case where $G$ is a dipath --- hence the focus on finding approximate solutions for the past 20 or so years. We give the first sub-$mn$ time $(1+\varepsilon, 1+\varepsilon)$-approximation algorithm for digraphs, where both the length-optimality and cost-constraint may be off by a multiplicative $(1 + \varepsilon)$ factor for an arbitrary constant $\varepsilon > 0$. This answers a problem posed by Bernstein in 2012 after his breakthrough in the undirected case. Next, we turn to algorithmic versions of the so-called folklore shortcut set (and hopset), which is a set of edges $H$ such that $G \cup H$ has the exact same reachability relations as $G$, but shortest paths in $G \cup H$ use no more than $n^{1+o(1)}/\sqrt{|H|}$ edges. Close to 30 years after its conception, Jambulapati, Liu, and Sidford gave the first near-linear time construction for $|H| \approx n$. In a perfect world scenario where the matrix multiplication exponent $\omega = 2$, we complete this picture for $n \lesssim |H| \lesssim n^2$. In the world we live in right now (2026), our result still leads to a near-linear time construction of a shortcut set closer in parameters to the folklore. Like prior work, this is designed to parallelize, which then leads to faster work-efficient parallel algorithms for single-source reachability and shortest paths when $m \gg n$. Both algorithms use sparsification, but in a somewhat oblique way (typically, the input graph is sparsified); the first algorithm sparsifies an algorithm, and the second algorithm sparsifies a proof object. Location: CoRE 301 Committee: Professor Jie Gao (Chair / Co-Advisor) Professor Aaron Bernstein (Co-Advisor) Professor Mario Szegedy |
| Start Date: 10 Jun 2026; Start Time: 02:00PM - 03:00PM Title: An Open-Source, Reproducible Tensegrity Robot that can Navigate Among Obstacles Bio: Speaker: Abstract: Tensegrity robots, composed of rigid struts and elastic tendons, provide impact resistance, low mass, and adaptability to unstructured terrain. Their compliance and complex, coupled dynamics, however, present modeling and control challenges, hindering planning and obstacle avoidance. This paper presents a complete, open-source, and reproducible system that enables navigation for a 3-bar tensegrity robot. The system comprises: (i) an inexpensive, open-source hardware design, and (ii) an integrated, open-source software stack for physics-based modeling, system identification, state estimation, path planning, and control. All hardware and software are publicly available: https://tensegrity.yale.edu/The proposed system tracks the robot using a static overhead camera and executes collision-free paths to a goal among known obstacle locations. System robustness is demonstrated through experiments involving unmodeled environmental challenges, including a vertical drop, an incline, and granular media, culminating in an outdoor field demonstration. To validate reproducibility, experiments were conducted using robot instances at two different laboratories. This work provides the robotics community with a complete navigation system for a compliant, impact-resistant, and shape-morphing robot. This system is intended to serve as a springboard for advancing the navigation capabilities of other unconventional robotic platforms. Location: Room 403, 4th floor, 1 Spring Street, Downtown New Brunswick Committee: Professor Kostas Bekris Associate Professor Mridul Aanjaneya Associate Professor Jingjin Yu Professor Richard Martin |
| Start Date: 12 Jun 2026; Start Time: 09:00AM - 11:00PM Title: Approximation Algorithms under Informational and Structural Constraints: Robustness, Fairness and Privacy Bio: Speaker: Abstract: The classical theory of approximation algorithms assumes exact and complete access to the input, with solution quality as the main requirement on the output. Both premises fail in many algorithmic problems in the modern world: input access is often expensive, noisy, or evolving (\emph{informational} constraints), and outputs may be required to satisfy properties beyond approximation guarantees (\emph{structural} constraints). This thesis studies the design of approximation algorithms in such settings, treating informational and structural constraints as formal design parameters. The contributions cover three families of problems, each engaging a distinct combination of constraints.\emph{Informational.} In the weak-strong oracle model for metric optimization, we give a $(1+\eps)$-coreset for $(k,z)$-clustering and lower bounds and tight algorithm for metric MST. In the fully dynamic setting, we give an $O(1)$-approximation for correlation clustering in $O(\polylog n)$ amortized update time against an adaptive adversary. \emph{Structural.} We formalize districting as the packing of compact $c$-balanced subgraphs and characterize its hardness and approximability across various graph classes. On planar and minor-free graphs we obtain an $O(\log n)$-approximation via a new structural primitive, \emph{scattering separator}. \emph{Both.} For differentially private release, we design input perturbation mechanisms whose outputs natively satisfy structural requirements; integer-valued, invariant-preserving histograms and consistent, transparent range queries without relying on post-processing. For private combinatorial optimization, we characterize the privacy-utility tradeoff for Max-CSP under constraint-level differential privacy, with tight bounds for triangle-free bounded-degree CSPs and for Max-$k$XOR with odd $k$. Location: CoRE 301 Committee: Professor Jie Gao (advisor) Assistant Professor Sumegha Garg Assistant Professor Jalaj Upadhyay Assistant Professor Sandeep Silwal (external committee member) |
| Start Date: 22 Jun 2026; Start Time: 12:00PM - 01:00PM Title: Tag-Filtered Approximate Nearest Neighbor Search Bio: Speaker: Abstract: Approximate Nearest Neighbor Search (ANNS) has become ubiquitous in the search and recommendation of objects represented with high-dimensional vectors. In these scenarios, objects are often associated with other metadata such as dates, color, manufacturer, etc. Therefore, it is common to perform ANNS with filters related to objects’ metadata. Existing approaches either suffer from high computation overhead or have poor performance on queries with low-to-medium frequency tags. In this qualifying exam, I will present a tag-frequency-aware vector index, which dynamically allocates resources to different tags and strikes a balance between performance and memory footprint. Extensive experiments showed that our approach achieved up to 13 times speedup over the baseline. Location: CoRE 305 Committee: Associate Professor Dong Deng Associate Professor Yongfeng Zhang Assistant Professor He Zhu Assistant Professor Xintong Wang |
| Start Date: 23 Jun 2026; Start Time: 03:00PM - 04:00PM Title: Exploring Memory Management and Virtualization Optimization Opportunities in Modern Tiered Memory Systems Bio: Speaker: Abstract: The growing memory demands of modern data-intensive applications have pushed systems toward tiered memory architectures, as scaling local DRAM alone is increasingly limited by capacity, cost, and platform constraints. In this context, CXL enables large-scale memory expansion, but accessing CXL-attached memory via a host-centric solution can incur significant latency, underutilized bandwidth, and energy overhead. To better utilize expanded memory, near-CXL processing offers a promising alternative by placing lightweight cores close to CXL memory, allowing data-intensive computation to run near where data resides. However, existing operating-system mechanisms for virtualization, scheduling, and data movement are primarily designed for conventional host-centric execution, and can become inefficient in near-CXL processing, where lightweight cores operate under different compute capabilities and memory access characteristics. This research investigates near-CXL processing as a concrete use case for rethinking OS support in modern tiered memory systems. We study how memory management and virtualization can be redesigned to reduce address translation overhead, enable cooperative execution between host and near-CXL cores, and more efficiently manage data movement. Through this work, we aim to identify general OS principles and mechanisms for future tiered and disaggregated memory architectures, where computation and memory are increasingly distributed across heterogeneous resources. Location: CoRE 305 Committee: Associate Professor Sudarsun Kannan Assistant Professor Minesh Patel Professor Santosh Nagarakatte Associate Professor Mridul Aanjaneya Professor Kevin Skadron (external) |
| Start Date: 13 Jul 2026; Start Time: 10:30AM - 11:30AM Title: Informed Collective Decision-Making via Voting Bio: Speaker: Abstract: When groups must make decisions, voting provides a procedurally fair method to aggregate preferences and knowledge. In our polarized, AI-integrated world plagued by misinformation and echo chambers, making fair, informed, and incentive-aware decisions has become both more critical and more challenging. The challenge grows as AI systems increasingly participate in group decisions, online deliberation platforms, and decentralized autonomous organizations scale democratic participation, and algorithmic filter bubbles fragment our information environment.In real voting scenarios, the preferences of agents are shaped by their social backgrounds as well as their information through different networks. The incomplete information creates a tension between the preferences agents perceive themselves to have and those they would hold under complete information. Similar challenges arise in political elections, corporate board decisions, expert panels, and human-AI collaborationacross domains. Motivated by such scenarios, we would like to ask: How can we make informed, efficient, fair, and incentive-aware collective decisions?This problem sits at a gap between two theoretical traditions. Classical social choice theory offers a rich arsenal of axioms for fair and efficient aggregation, but presupposes agents have well-defined, complete preferences. Epistemic social choice and the Condorcet jury Theorem~\cite{condorcet1785} provide guarantees for information aggregation, but assume agents share common interests and cannot extend to strategic agents. The combination of preference heterogeneity and epistemic uncertainty under strategic behavior introduces complications that existing theory cannot accommodate.My research develops a unified game-theoretic framework that bridges social choice theory and information economics to address this challenge. The foundation is a combined utility model that formally decomposes preferences into subjective components (dependent on social groups) and objective components (dependent on uncertain ground truth), providing the mathematical structure to analyze this rich decision environment. Building on this, I introduce the informed-X desiderata---a benchmark requiring that outcomes match what agents would choose under complete information. Finally, through game-theoretic analysis of strategic behavior, I establish that strategic voting can actually \textit{help} rather than hinder informed decisions, and characterize precisely when this occurs. Location: CoRE 301 Committee: Professor Lirong Xia Professor David Pennock Professor Kangning Wang |
| Start Date: 15 Jul 2026; Start Time: 10:00AM - 11:00AM Title: Physics-Informed Learning for Robot Motion Planning and Control Bio: Ahmed Qureshi is an Assistant Professor in the Department of Computer Science at Purdue University, where he directs the Cognitive Robot Autonomy and Learning (CoRAL) Lab. His research pursues a physics-first philosophy for robot motion learning: rather than relying on large expert demonstrations or trial-and-error interaction, his group develops methods that embed the governing laws of physics directly into learning algorithms. This approach has produced self-supervised methods that train in minutes, require no expert annotation, and plan in near real-time across high-dimensional, manipulation, and unknown environments. His broader research spans scalable motion planning, dexterous manipulation, active perception, and multi-agent task and motion planning. Dr. Qureshi's work has been recognized with spotlight and best paper awards at top academic venues. He serves as an Associate Editor for IEEE Transactions on Robotics and IEEE Robotics and Automation Letters, and received the Outstanding Associate Editor Award from RA-L in 2024. He has served on the program committees of RSS, ICRA, IROS, and CoRL. He earned his B.S. in Electrical Engineering from NUST, M.S. from Osaka University, and Ph.D. in Intelligent Systems, Robotics, and Control from UC San Diego. Website: https://qureshiahmed.github.io/ Speaker: Abstract: Modern robot motion learning often relies on large datasets and expensive expert demonstrations to learn structure that physics can already provide. This talk discusses an alternative approach: instead of asking neural networks to rediscover physical principles from data, PDE-based priors can be embedded directly into the learning process. The talk focuses on motion planning through continuous value functions governed by the Eikonal PDE, a special case of Hamilton-Jacobi equations that captures shortest-path and minimum-time behavior. This formulation enables self-supervised robot motion learning without relying on expert trajectories, exhaustive graph search, or trial-and-error interaction. The resulting methods train quickly, generalize across environments, and infer motion plans in near real time. The talk also shows how these ideas scale to high-dimensional systems and constraint-rich manipulation tasks, and how Eikonal priors can provide a motion-aware mapping representation that is more directly useful for planning than standard occupancy grids or signed distance fields. Finally, it discusses how incorporating these priors into reinforcement learning improves data efficiency and scalability. Together, these results suggest that physics can serve as a useful foundation for learning robot motion, with data used to adapt and extend these priors rather than replace them. Location: SPR-403, at 1 Spring Street, Downtown New Brunswick Committee: |
| Start Date: 24 Jul 2026; Start Time: 11:00AM - 12:00PM Title: Efficient and Scalable Near-Duplicate Text Alignment Algorithms Bio: Speaker: Abstract: Near-duplicate text alignment locates and aligns the specific subsequences of one text that are near-duplicates of subsequences of another—a core operation for large-language-model data-contamination detection, training-data deduplication, and retrieval-augmented generation. The problem is computationally hard because a text has quadratically many subsequences and the underlying Jaccard / weighted Jaccard similarities are costly to evaluate. This exam studies near-duplicate text alignment from three complementary algorithmic directions: (i) one-permutation hashing, which computes passage signatures in a single pass under Jaccard similarity; (ii) consistent weighted sampling, which extends alignment to weighted Jaccard similarity with provably optimal grouping; and (iii) LSHAlign, a locality-sensitive-hashing scheme that further supports the all-pair alignment setting with linear time and space complexity. Extensive experiments on real text collections demonstrate that the proposed methods substantially reduce index-construction time, index size, and query latency compared with prior approaches, while preserving alignment accuracy. This work offers practical algorithmic foundations for scaling near-duplicate text analysis to the large corpora used in modern language-model pipelines. Location: CoRE 301 Committee: Associate Professor Dong Deng Associate Professor Yongfeng Zhang Professor Lirong Xia Assistant Professor Qiong Zhang |
| Start Date: 28 Jul 2026; Start Time: 12:00PM - 01:30PM Title: Cognitive Memory Mechanisms for Understanding and Improving Large Language Models Bio: Speaker: Abstract: We introduce a cognitively grounded framework for understanding and improving memory in neural language models. First, we present a mechanistic connection between attention-based sequence-to-sequence models and cognitive models of human memory search. This work maps components of neural machine translation architectures onto context-based memory models and shows that learned attention mechanisms can support human-like retrieval behavior in free recall, including systematic effects of context and item history. Second, we study the “lost-in-the-middle” phenomenon in long-context language models, in which models retrieve information more reliably from the beginning and end of a context than from the middle. Through experiments on language models trained on tasks inspired by human short-term and long-term memory paradigms, we show that this U-shaped positional bias can emerge from competing retrieval demands, autoregressive model structure, and learned attention dynamics rather than from context length alone. Together, these results suggest that cognitive memory theory can help explain when and why neural language models succeed or fail at retrieval. Building on this foundation, our ongoing work investigates how memory-augmented LLM architectures can incorporate cognitive principles such as hierarchical organization, episodic structure, and memory encoding and retrieval. This research aims to develop more structured, adaptive, and reliable model memory mechanisms for long-context reasoning and retrieval-augmented generation. Location: CoRE 305 Committee: Assistant Professor Qiong Zhang (Chair) Assistant Professor Hao Wang Assistant Ruixiang Tang Professor Lirong Xia |
| Start Date: 29 Jul 2026; Start Time: 10:00AM - 12:00PM Title: Extending established memory recall theories and findings beyond the laboratory Bio: Speaker: Abstract: To understand how humans encode and retrieve information, memory research traditionally uses well-controlled laboratory experiments. For example, studying and recalling a list of discrete items. Although useful in uncovering fundamental memory mechanisms, lab experiments do not fully capture the range of dynamics and behaviors shown in real-world scenarios. In our work, we extend well-established theories from lab experiments to account for this broader range of behaviors revealed in naturalistic settings.The first naturalistic setting we considered was people remembering information with others in a group. To model group recall, we extended a cognitive model that captures how individuals recall a list of items alone. This model encodes studied items in a latent space, and recalls are made by searching for memories in this space. Our extended model captures and explains group memory behaviors, including collaborative inhibition: the counterintuitive finding that a group of people recalling information together remembers less than the same number of individuals recalling alone. Collaborative inhibition is captured through a model mechanism in which group members are influenced by each other's recalls. This influence synchronizes their memory search trajectories in the latent space, negatively affecting group performance.We next studied how people summarize stories and news articles. To understand the cognitive processes involved in summarization, we analyzed large datasets of human summaries for behavioral patterns studied in memory literature. Our work demonstrates that summarization is closely related to recall: the process of summarizing is constrained by mechanisms of memory recall, displaying behaviors that align with well-known memory effects of primacy, recency, and temporal contiguity. By additionally analyzing a dataset of story recall, we showed that naturalistic recall also takes the form of summarization, showing higher memorability for central information from a story. In more realistic settings, since recall is constrained by limited time and cognitive resources, it is rational to prioritize central information, maximizing informativeness while minimizing effort.Our ongoing work studies why the human memory system segments continuous experiences into discrete events at multiple levels of abstraction. To examine whether this event segmentation serves an adaptive purpose, we draw parallels between the outputs of neural network models, which are optimized solely for task performance, and human memory behaviors. Location: CoRE 301 Committee: Assistant Professor Qiong Zhang Professor Amélie Marian Professor Jacob Feldman Associate Professor Yongfeng Zhang |
| Start Date: 29 Jul 2026; Start Time: 10:30AM - 12:00PM Title: Practical Formal Methods for the Linux eBPF Verifier: Scalable Verification and Witness Generation Bio: Speaker: Abstract: The eBPF framework allows user-supplied programs to run inside the Linux kernel for tasks such as networking, tracing, observability, access control, load balancing, and security monitoring. Because these programs execute in a safety-critical kernel context, every eBPF program is checked by the eBPF verifier before it is loaded. The verifier uses static analysis to prove properties such as termination, safe memory access, and well-defined arithmetic behavior. Its analysis must therefore be sound, precise enough to accept useful safe programs, and efficient enough for use in a production kernel. This dissertation focuses on the verifier's value-tracking logic, a central part of its static analysis, where bugs have led to serious kernel vulnerabilities.Specifically, we develop formal techniques for reasoning about the soundness and precision of the eBPF verifier's value-tracking logic as implemented in the Linux kernel. We first introduce a divide-and-conquer verification strategy for implementation-level soundness checking. Rather than verifying each full abstract operator together with the shared reduction operator as one monolithic proof obligation, we separate instruction-specific reasoning from shared reduction reasoning and compose the results. This significantly reduces repeated solver time across operators and kernel versions, while also exposing latent unsoundness in several bitwise operators that were previously masked by the shared reduction operator.We then introduce differential synthesis, an SMT-based method for turning failed soundness checks into executable eBPF witness programs. When a soundness check fails, the SMT solver produces a counterexample that may involve abstract inputs that never arise during verifier analysis, so differential synthesis searches for bounded instruction sequences that reach the relevant abstract values from valid initial verifier inputs. The resulting programs demonstrate reachable mismatches between concrete execution and the verifier's abstract analysis, making verification failures reproducible and useful for debugging and regression testing. Finally, we extend this synthesis framework from soundness to precision; we develop a precision-comparison framework for comparable abstract operators in the verifier. Our framework checks whether one operator is at least as precise as another for all abstract inputs and, when their precision differs, searches for real eBPF programs that illustrate the difference. We use this framework to evaluate precision-improving kernel patches, measure the precision gained from the shared reduction operator, and develop a more precise multiplication operator that has been upstreamed to the Linux kernel. Location: Hill 350 Committee: Professor Santosh Nagarakatte Associate Professor He Zhu Associate Professor Srinivas Narayana Associate Professor Rich Martin |
| Start Date: 12 Aug 2026; Start Time: 01:00PM - 03:00PM Title: Learning to See, Generate, and Act for Scalable Robotic Manipulation Bio: Speaker: Abstract: A fundamental challenge in scaling robotic manipulation is reconciling broad generalization across diverse tasks with precise dexterity for fine-grained control. To resolve this tension, my dissertation instantiates a dual-system architecture where a deliberative System-1 predicts high-level action primitives through world modeling, and a reactive System-0 translates them into low-level motor commands. I advance this vision through the complementary contributions across my PhD publications, spanning efficient visual world models that learn latent dynamics directly from offline videos, a scalable data generation paradigm that synthesizes natural hand-object interactions from multi-modal sensing gloves to supply the rich data needed for dexterous robot learning, and a policy architecture that frames robot actions as language tokens to bridge high-level planning with low-level execution. Together, these contributions build towards a unified framework of robust generalization and dexterity across diverse manipulation tasks. Location: Room 402, 4th floor, 1 Spring Street, Downtown New Brunswick Committee: Professor Abdeslam Boularias Professor Kostas Bekris Professor Chengzhi Mao |
| Start Date: 10 Sep 2026; Start Time: 12:00PM - 03:00PM Title: 2026 Annual CS Fall Fair Bio: Speaker: Abstract: Location: Busch Student Center Committee: |
Upcoming Events
| 29 Jul 2026; - 10:00AM - 12:00PM Extending established memory recall theories and findings beyond the laboratory |
| 12 Aug 2026; - 01:00PM - 03:00PM Learning to See, Generate, and Act for Scalable Robotic Manipulation |
| 10 Sep 2026; - 12:00PM - 03:00PM 2026 Annual CS Fall Fair |