- talks - lists.cs.princeton.edu

Talks on interpretability in multi-modal foundation models: Wednesday, 12/10 at 3:30pm
by Emily C. Lawrence 08 Dec '25

08 Dec '25

Guillermo Sapiro and Sanketh Vedula will be hosting guests at Princeton Precision Health's office at 252 Nassau St 2nd FL at 3:30pm - 5:00pm on Wednesday, December 10 who will present On decoding the inner workings of multimodal foundation models . Speakers-Alberto Cazzaniga, Lorenzo Basile & Diego Doimo all Researchers from Area Science Park Trieste, Italy. Alberto Cazzaniga will present, On image-text communication in vision language models Vision-language models (VLMs) integrate images and text efficiently, but how they transmit visual information into text generation remains poorly understood. We present two mechanistic findings that clarify image–text communication in modern VLMs. First, using counterfactual multimodal queries, we isolate a small set of attention heads that decide whether the model follows the image or its internal knowledge; editing these heads reliably shifts behavior and reveals the image regions that drive it. Second, comparing native and non-native VLMs, we show that they rely on distinct pathways for visual-to-text transfer: non-native models distribute information across many image tokens, whereas native models depend on a single gate-like token whose removal severely degrades image understanding. Together, these insights offer a clearer and more actionable view of how VLMs process visual evidence. Lorenzo Basile will present, Head Specialization in Vision, Language, and Multimodal Transformers Transformers often appear as opaque systems, but recent research shows that they contain meaningful internal structure. A key example is head specialization, where individual attention heads consistently encode specific concepts across language, vision, and multimodal models. Some heads capture visual properties like shape or color, while others represent linguistic or numerical information such as sentiment or toxic words. Identifying these specialized heads not only deepens interpretability but also provides practical tools for controlling model behavior. By selectively amplifying or suppressing head activity, we can adjust concept representations, adapt models to new tasks, and improve performance with minimal parameter changes. This talk presents methods for discovering head specialization, demonstrates its emergence across diverse architectures, and shows how these structures can be leveraged for precise, parameter-efficient interventions. Diego Doimo will present, The geometry of hidden representations of large transformer models In this talk, we will show how the geometric properties of hidden representations can help us understand the semantic information encoded by transformers. In the first part, we will focus on the intrinsic dimension of the internal transformer representations, showing that it is a valuable tool for identifying the layers encoding the semantic content of data across different domains, such as images, biological sequences, and text. In the second part, we will analyze the probability density of hidden representations. Specifically, we focus on how language models solve a question-answering task with few-shot learning and fine-tuning. We show that while both approaches can achieve similar performance, they create very different density distributions in the hidden representations, changing with a sharp geometrical transition in the middle of the network. For more information, please contact Sanketh Vedula- [ mailto:svedula@princeton.edu | svedula(a)princeton.edu ] Getting to the seminar space currently requires that you climb a set of stairs. If an accommodation is needed, please contact PPH in advance at: [ mailto:PrincetonPPH@princeton.edu | PrincetonPPH(a)princeton.edu ] Thank you, Princeton Precision Health

1 0

Tolulope Oshinowo present his MSE talk "Pantry: Reconceptualizing Social Media Feed Curation and Its Implications for User Agency" Wednesday, December 3, 2025 at 11:00am in CS 402.
by CS Grad Department 28 Nov '25

28 Nov '25

Tolulope Oshinowo present his MSE talk "Pantry: Reconceptualizing Social Media Feed Curation and Its Implications for User Agency" Wednesday, December 3, 2025 at 11:00am in CS 402. Thesis adviser: Andrés Monroy-Hernández, Manoel Horta Ribeiro (reader) Abstract: Social media feeds are among the most influential yet opaque infrastructures of everyday life online, typically optimized for engagement rather than user agency. In response, we present Pantry, a mobile app that reconceptualizes feed curation around user-authored control and in-the-moment preference expression. To evaluate this approach, we conducted a study combining preliminary surveys, tracked user activity logs, and debrief interviews to examine how users engaged with and perceived a more intentional feed curation workflow. Overall, our findings suggest that user-authored feeds are a viable and welcomed alternative, but also raise questions about how far such designs should diverge from familiar social media interface conventions. As both a functional system and research contribution, Pantry advances ongoing HCI discussions by grounding largely theoretical conversations in practical implementation. In doing so, we provide a concrete foundation for examining how alternative feed architectures can redistribute agency, and invite further inquiry through systems-building.

1 0

TODAY: 10:30 am, Nov. 24, John (Can) Bostanci, Columbia, “A computational lens on quantum information,” Computer Science building, 105
by Emily C. Lawrence 24 Nov '25

24 Nov '25

Speaker: John (Can) Bostanci , Columbia University 10:30 am, Monday, Nov. 24, 2025 Computer Science Building, Room 105 Host: Prof. Alex Lombardi, CS “A computational lens on quantum information” Abstract: Quantum information is known to behave in ways completely unlike its classical counterpart. For example, quantum information can be entangled across long distances, and in general cannot be cloned. This discrepancy leads to fundamental differences between quantum and classical provers, cryptographic primitives, and transformations. In this talk I will discuss recent advances in our understanding of the ways that these properties can affect the nature of computation itself. My focus will be on our work giving the first classical oracle separation between QMA and QCMA, providing strong evidence that quantum proofs confer genuine computational advantages over classical ones. The result, surprisingly, identifies the reusability of classical proofs as the reason they must be less useful than quantum proofs. Our proof introduces new ideas in quantum query complexity, including a bosonic compressed oracle framework, which provides a structured way to reason about verifiers that process quantum states. I will also discuss broader themes connecting quantum verification, unclonable cryptography, and the complexity of quantum transformations.

1 0

Special Quantum Seminar: John (Can) Bostanci, Columbia, “A computational lens on quantum information,” 10:30 am, Mon. Nov. 24 (Computer Science, 105)
by Emily C. Lawrence 20 Nov '25

20 Nov '25

Speaker: John (Can) Bostanci , Columbia University 10:30 am, Monday, Nov. 24, 2025 Computer Science Building, Room 105 Host: Prof. Alex Lombardi, CS “A computational lens on quantum information” Abstract: Quantum information is known to behave in ways completely unlike its classical counterpart. For example, quantum information can be entangled across long distances, and in general cannot be cloned. This discrepancy leads to fundamental differences between quantum and classical provers, cryptographic primitives, and transformations. In this talk I will discuss recent advances in our understanding of the ways that these properties can affect the nature of computation itself. My focus will be on our work giving the first classical oracle separation between QMA and QCMA, providing strong evidence that quantum proofs confer genuine computational advantages over classical ones. The result, surprisingly, identifies the reusability of classical proofs as the reason they must be less useful than quantum proofs. Our proof introduces new ideas in quantum query complexity, including a bosonic compressed oracle framework, which provides a structured way to reason about verifiers that process quantum states. I will also discuss broader themes connecting quantum verification, unclonable cryptography, and the complexity of quantum transformations.

1 0

Tianyu Gao will present his FPO "Enabling Language Models to Process Information at Scale" on Thursday, November 20, 2025 at 1:30 PM in CS 402.
by CS Grad Department 19 Nov '25

19 Nov '25

Tianyu Gao will present his FPO "Enabling Language Models to Process Information at Scale" on Thursday, November 20, 2025 at 1:30 PM in CS 402. The members of Tianyu’s committee are as follows: Examiners: Danqi Chen (Adviser), Sanjeev Arora, Tri Dao Readers: Karthik Narasimhan, Yoon Kim (MIT) A copy of his thesis is available upon request. Please email gradinfo(a)cs.princeton.edu if you would like a copy of the thesis. Everyone is invited to attend his talk. Abstract follows below: Language models (LMs) can effectively internalize knowledge from vast amounts of pretraining data, enabling them to achieve remarkable performance on exam-style benchmarks. Expanding their ability to compile, synthesize, and reason over large volumes of information on the fly will further unlock transformative applications, ranging from AI literature assistants to generative search engines. In this thesis, I will present my research on advancing LMs for processing information at scale. (1) I will first discuss my contributions to creating efficient pre-training and customization methods for LMs, which enable scalable deployment of LM-powered applications across diverse settings. (2) I will then introduce my foundational work on using contrastive learning to produce high-performing text embeddings, which form the cornerstone of effective and scalable search. (3) Finally, I will present my evaluation framework for LM-based information-seeking systems, emphasizing the importance of providing citations for verifying the model-generated answers. Our evaluation highlights shortcomings in LMs’ abilities to reliably process long-form texts (e.g., dozens of webpages), which I address by developing state-of-the-art long-context LMs that outperform leading industry efforts while using a small fraction of the computational budget. I will conclude the thesis by sharing my vision for the next generation of autonomous information processing systems and outline the foundational challenges that must be addressed to realize this vision.

2 1

Howard Chen will present his FPO "Memory Adaptation in Language Models for Long-Horizon and Continual Tasks" on Wednesday, November 19, 2025 at 2:00 PM in CS 105.
by CS Grad Department 18 Nov '25

18 Nov '25

Howard Chen will present his FPO "Memory Adaptation in Language Models for Long-Horizon and Continual Tasks" on Wednesday, November 19, 2025 at 2:00 PM in CS 105. The members of Howard’s committee are as follows: Examiners: Karthik Narasimhan (Co-Adviser), Danqi Chen (Co-Adviser), Tom Griffiths Readers: Peter Henderson, Jason Weston (Meta) A copy of his thesis is available upon request. Please email gradinfo(a)cs.princeton.edu if you would like a copy of the thesis. Everyone is invited to attend his talk. Abstract follows below: Humans can remember events from decades ago and are not overwhelmed by the constant flood of sensory input throughout our lifetime. By contrast, while modern language models (LMs) can memorize massive amounts of data, they fall short in dynamically incorporating the stream of information that emerges over time. As language models operate in a fast-changing world and can now accumulate their own experiences, this limitation becomes acute. It stems from an inflexible memory design: deployed models assume fixed parameters updated only intermittently; larger context windows remain bounded and unevenly effective. This thesis addresses both forms of inflexibility through memory adaptation: deliberately shaping LM’s memory over time. Two questions guide the work: (1) how to organize in-context (working) memory when accumulated context exceeds what the model can process, and (2) how to update parametric memory with minimal disruption as continual weight changes often induce forgetting. I first survey how memory is implemented from early to contemporary AI systems and where they fall short. I then proceed in two parts. Part I studies long-horizon tasks in realistic web navigation, showing that standard training does not extend the effective horizon. I introduce a framework that treats long context as a managed data structure the model can traverse. This organization mitigates long-context failures by prioritizing salient information through reasoning and interaction. Part II studies how to incorporate new data into parameters while avoiding forgetting. I analyze dynamics for two data types—factual knowledge and task specific capability. I then trace failure patterns across stages, and propose mitigation, identifying when forgetting is limited. I also examine tradeoffs between gains in the target domain and drops on other tasks under different learning objectives, and identify conditions with minimal forgetting. Taken together, the results support memory adaptation as a unifying principle: in-context memory organization and deliberate parametric memory updates enable long-horizon and continual tasks. I close with a roadmap toward self-improving agents: a “machine metabolism” framework where the model decides what to ingest, where to store it, and when to consolidate. I discuss the opportunities and implications for building such a system.

2 1

Sadhika Malladi will present her FPO "Deep Learning Theory in the Age of Generative AI" on Thursday, November 20, 2025 at 11:00 AM in CS 401.
by CS Grad Department 17 Nov '25

17 Nov '25

Sadhika Malladi will present her FPO "Deep Learning Theory in the Age of Generative AI" on Thursday, November 20, 2025 at 11:00 AM in CS 401. The members of Sadhika’s committee are as follows: Examiners: Sanjeev Arora (Adviser), Danqi Chen, Chi Jin Readers: Tri Dao, Zhiyuan Li (TTIC) A copy of her thesis is available upon request. Please email gradinfo(a)cs.princeton.edu if you would like a copy of the thesis. Everyone is invited to attend her talk. Abstract follows below: This thesis develops theoretical frameworks for understanding modern deep learning in realistic training regimes. As model architectures, optimization methods, and data pipelines continue to evolve, much of the existing theory no longer describes how large-scale systems behave in practice. Exhaustive empirical studies, once central to understanding these systems, have become infeasible at current computational scales. The goal of this work is to bridge that gap by constructing mathematical analyses that remain faithful to real training scenarios while offering prescriptive guidance for practice. The first part establishes stochastic differential equation (SDE) approximations that accurately model stochastic optimization at finite learning rates. These analyses clarify when SDE-based descriptions are valid and yield general scaling rules for efficient large-batch training. The second part examines the fine-tuning of pre-trained language models, formalizing it as a small, structured modification to an existing representation. This perspective explains the effectiveness of parameter-efficient adaptation methods and motivates a memory-efficient fine-tuning algorithm derived from the theoretical analysis. The third part studies preference-based alignment, identifying why commonly used objectives fail to fit preference data and introducing diagnostic measures that link training objectives to observed model behavior. Together, these results demonstrate how mathematically grounded analysis can extend beyond idealized assumptions to capture the dynamics of contemporary machine learning systems. The thesis argues that such theory can play a prescriptive and interpretive role—guiding large-scale training, diagnosing failure modes, and providing a stable foundation for understanding as empirical practices continue to evolve.

2 1

Abhishek Joshiwill present his MSE talk "Procedural Generation of Articulated Simulation-Ready Assets" TODAY (Monday, November 17, 2025) at 2:30pm in Friend 125.
by gradinfo＠cs.princeton.edu 17 Nov '25

17 Nov '25

Abhishek Joshiwill present his MSE talk "Procedural Generation of Articulated Simulation-Ready Assets" TODAY (Monday, November 17, 2025) at 2:30pm in Friend 125. Thesis adviser: Jia Deng, Tom Silver (reader) Abstract: While simulated environments have proven effective for training robots, the lack of high-quality articulated assets makes it difficult for robots to learn new skills involving such objects. To this end, we introduce Infinigen-Articulated, a toolkit for creating realistic, diverse, procedurally generated articulated assets for robot simulation. We introduce procedural generators for 18 common household articulated object categories, along with tools to create custom articulated assets in Blender and export them directly to standard simulation formats including information such as joint dynamics, material physics, collision meshes, and semantic metadata. Experiments demonstrate that assets sampled from these generators are effective for movable part segmentation, training generalizable reinforcement learning policies, and sim-to-real transfer of imitation learning policies.

1 0

CS Colloquium Speaker: Christian Borgs - Monday, November 17 at 12:10pm
by Emily C. Lawrence 17 Nov '25

17 Nov '25

CS Colloquium Speaker Speaker: Christian Borgs, University of California, Berkeley Date: Monday, November 17 Time: 12:10pm EST Location: CS 105 Host: Adji Bousso Dieng Event page: https://www.cs.princeton.edu/events/are-travel-bans-effective-containing-sp… Register for live-stream online here: https://princeton.zoom.us/webinar/register/WN_CGmNJkqmS76ML-byDkxNLA Title: Are Travel Bans Effective in Containing the Spread of a Disease? Abstract: In this talk, I present a mathematical model for the spread of an epidemic from one community to another via travel. Here each community is modeled by a random network (for simplicity, we assume it is an Erdos-Renyi random graph), with the epidemic spread inside the community given by the SIR model on this graph. Travel is modeled by individuals moving from one community to the other at some rate \eta_T, and returning home at another rate \eta_H. We assume that the return rate is of the same order as the recovery rate of the epidemic, while \eta_T is much smaller. Under this assumption, we rigorously prove that if an epidemic starts in the first community, and the second community enacts a travel ban at the moment the epidemic is large enough to be detectable, such a travel ban is ineffective in preventing a large outbreak in the second community. But contrast, other mitigation measure like social distancing or vaccinations (modeled by reducing the rate of infections in the second community) are effective. This is joint work with my students Karissa Huang and Geng Zhao. Bio: Christian Borgs is professor in the Berkeley AI Research Group (BAIR) in the EECS department at UC Berkeley, and faculty director of the Bakar Institute of Digital Materials for the Planet. Borgs was trained as a physicist, holding a Ph.D. from the University of Munich and a Habilitation from the Free University in Berlin, and then becoming chair of Statistical Mechanics at the University of Leipzig. After that, he worked at Microsoft Research for about 20 years, first as co-founder and co-director of the Theory Group, and then as co-founder and Deputy Managing Director of Microsoft Research New England in Cambridge, MA, which was one of the first labs integrating research from areas as diverse as theoretical CS, ML, economics and qualitative social science. He joined Berkeley in 2020. Borgs is a Fellow of the American Mathematical Society, and the American Association for the Advancement of Science. Among the earlier awards he has received are the Karl-Scheel Prize of the German Physical Society, and the Heisenberg Fellowship of the German Research Council. Borgs’ current research focuses on both AI for science and the science of networks, including mathematical foundations, particularly the theory of graph limits aka Graphons (which he co-invented about 15 years ago), graph processes, graph algorithms, and applications of graph theory from economics to systems biology and epidemics. In addition, he has worked on various aspects of responsible AI, from differential privacy to questions of bias in automatic decision making or polarization in online communities. In AI for Science, he is interested in applications of AI to material science, with the aim to create materials to address urgent problems of society like climate change or scarcity of water in many regions of the world.

1 2

Mengzhou Xia will present her FPO "Advancing the Pareto Frontier of Training Open Language Models" on Wednesday, November 19, 2025 at 4:30 PM in CS 105.
by CS Grad Department 14 Nov '25

14 Nov '25

Mengzhou Xia will present her FPO "Advancing the Pareto Frontier of Training Open Language Models" on Wednesday, November 19, 2025 at 4:30 PM in CS 105. The members of Mengzhou’s committee are as follows: Examiners: Danqi Chen (Adviser), Sanjeev Arora, Peter Henderson Readers: Karthik Narasimhan, Pang Wei Koh (University of Washington) A copy of his thesis is available upon request. Please email gradinfo(a)cs.princeton.edu if you would like a copy of the thesis. Everyone is invited to attend his talk. Abstract follows below: Large language models (LLMs) have reshaped AI by enabling breakthroughs in language understanding, reasoning, and diverse applications. However, their massive computational demands and the proprietary nature of leading models hinder broad accessibility and customization. My work addresses these challenges by optimizing the use of existing compute, data, and models to push the Pareto frontier for LLM training. In doing so, it not only produces stronger language models but also offers universal approaches that support effective customization and advance our scientific understanding of model training. First, we study how structured pruning can be leveraged to pre-train compact, high performing models at a fraction of the usual pre-training cost, demonstrating its effectiveness in pushing the Pareto frontier for general-purpose pre-training. Next, we turn to the post-training phase to explore the critical role of data in shaping model behavior, presenting principled data optimization techniques that enhance models’ capabilities, safety, and transparency—showing that “less is more” when it comes to constructing effective training datasets. Finally, we introduce novel post-training approaches that more effectively align language models with desired behaviors and objectives. By revealing gaps in the reasoning abilities of even proprietary models, we outline future directions for building AI systems with enhanced reasoning capabilities—focusing on broader data synthesis through agentic processes and enabling advanced applications.

1 0