- talks - lists.cs.princeton.edu

Yebin Chon will present her FPO "LILAC: Enhancing Parallel Code Naturalness through Hybrid Compiler-LLM Translation" on Monday, January 26, 2026 at 1:30pm in CS 402.
by Gradinfo 20 Jan '26

20 Jan '26

Yebin Chon will present her FPO "LILAC: Enhancing Parallel Code Naturalness through Hybrid Compiler-LLM Translation" on Monday, January 26, 2026 at 1:30pm in CS 402. Readers: Margaret Martonosi, Zachary Kincaid, David August (adviser) Examiners: Sneha Goenka, Aarti Gupta, David August All are welcome to attend. Title: "LILAC: Enhancing Parallel Code Naturalness through Hybrid Compiler-LLM Translation" Abstract: Modern parallel hardware is becoming increasingly diverse, and there is a following increase in the heterogeneity of parallel programming models (PPMs), often tied to a particular parallel architecture. This brings a need for translation between different programming models to properly port and maintain code for various hardware systems. Source-to-source translators may automate this step and provide production compiler-level correctness guarantees. However, previous attempts to build translators that can go between different PPMs often result in unnatural code, hindering programmer comprehension and long-term maintenance. More recent Large Language Model (LLM) based methods produce highly natural, artful code, but lack the correctness guarantees of traditional compiler-based methods. One work provides a formal proof of correctness for the LLM-generated translation, but is severely limited in the code that it can process. This creates a critical trade-off: developers must choose between the correctness and applicability of compilers and the artfulness of LLMs, both of which are essential. This thesis provides a new way to bring the two methods together, creating cross-PPM translations that are both correct and natural. To achieve this, it introduces LILAC, a hybrid framework that combines the strengths of traceable and widely applicable compiler transformations and the stylistic capabilities of LLMs. LILAC builds on top of an existing compiler-based translator to provide functionally correct parallel translations and to simultaneously detect areas of potential unnaturalness. It then uses an LLM through an API to obtain suggestions for naturalizing the areas identified by the compiler. The compiler then checks these suggestions for correctness and applies the transformations to the code. In this way, LILAC performs several hybrid compiler-LLM passes to naturalize the code without compromising correctness. In a human study with 26 expert C/C++ programmers evaluating CUDA-to-OpenMP translations, LILAC-generated code was judged to be significantly more natural and maintainable than the output of the baseline compiler. The benchmarks used are unable to be translated by the formally verified LLM method. By building on a correct compiler translation, LILAC ensures that enhancements in naturalness do not come at the cost of correctness or applicability.

2 1

Yinghui He will present her General Exam "Skill-Targeted Adaptive Learning for LLMs" on Tuesday, January 20, 2026 at 3:00 PM in CS 401 and via zoom.
by CS Grad Department 15 Jan '26

15 Jan '26

Yinghui He will present her General Exam "Skill-Targeted Adaptive Learning for LLMs" on Tuesday, January 20, 2026 at 3:00 PM in CS 401 and via zoom. Zoom link: https://princeton.zoom.us/my/yinghuihe Committee Members: Sanjeev Arora (advisor), Danqi Chen, Zhuang Liu Abstract: Large language models (LLMs) can often “learn” at test time from carefully chosen in-context examples and at train time from supervised fine-tuning (SFT), yet both settings suffer from a common failure mode: generic instruction or data additions frequently yield diminishing returns and can even harm performance in smaller models. We propose skill-targeted adaptive learning as a unifying approach to closing these gaps by using an LLM teacher’s metacognition. First, we study adaptive skill-based in-context learning. We discovered that naive skill prompting can create a “cognitive overload” effect on easy questions by injecting unnecessary information. To address this, we present AdaptMI, which introduces skill-targeted in-context math instructions only when a model struggles, and AdaptMI+, which further targets examples to the specific skills missing from the model’s intermediate reasoning and final responses. Second, we explore skill-targeted adaptive training for settings where vanilla supervised fine-tuning saturates. We present STAT, a fine-tuning strategy in which a strong teacher labels each training instance with required skills, builds a Missing-Skill Profile by auditing the student’s errors, and then either (i) reweights/selects existing data or (ii) synthesizes new examples emphasizing missing skills. Together, these skill-targeted adaptive learning approaches provide a principled way to overcome performance saturation and systematically close capability gaps in language models. Reading List: https://docs.google.com/document/d/17ZEbYvjn-D7WVGy66YVyoRsUWGr4-pisie41dI4… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

Yebin Chon will present her FPO "LILAC: Enhancing Parallel Code Naturalness through Hybrid Compiler-LLM Translation" on Monday, January 26, 2026 at 1:30pm in CS 402.
by Gradinfo 14 Jan '26

14 Jan '26

Yebin Chon will present her FPO "LILAC: Enhancing Parallel Code Naturalness through Hybrid Compiler-LLM Translation" on Monday, January 26 , 2026 at 1:30pm in CS 402. Readers: Margaret Martonosi, Zachary Kincaid, David August (adviser) Examiners: Sneha Goenka, Aarti Gupta, David August Title: "LILAC: Enhancing Parallel Code Naturalness through Hybrid Compiler-LLM Translation" All are welcome to attend. Abstract: Modern parallel hardware is becoming increasingly diverse, and there is a following increase in the heterogeneity of parallel programming models (PPMs), often tied to a particular parallel architecture. This brings a need for translation between different programming models in order to properly port and maintain code for various hardware systems. Source-to-source translators may automate this step and provide production compiler-level correctness guarantees. However, previous attempts to build translators that can go between different PPMs often result in unnatural and hard-to-read code, hindering programmer comprehension and long-term maintenance. More recent Large Language Model (LLM) based methods produce highly natural, artful code, but lack the correctness guarantees of traditional compiler-based methods. One work provides a formal proof of correctness for the LLM-generated translation, but is severely limited in the code that it can process. This creates a critical trade-off: developers must choose between the correctness and applicability of compilers and the artfulness of LLMs, both of which are essential. This thesis provides a new way to bring the two methods together, creating cross-PPM translations that are simultaneously correct and natural. To achieve this, it introduces LILAC, a hybrid framework that combines the strengths of traceable and widely applicable compiler transformations and the stylistic capabilities of LLMs. LILAC builds on top of an existing compiler-based translator to provide functionally correct parallel translations, and at the same time detect areas of potential unnaturalness. It then uses an LLM through an API to obtain suggestions for naturalizing the areas identified by the compiler. The compiler then checks these suggestions for correctness and applies the transformations to the code. In this way, LILAC performs several hybrid compiler-LLM passes to naturalize the code - refactoring control flow structures, improving variable names, and inserting comments - without compromising correctness. In a human study with 26 expert C/C++ programmers evaluating CUDA-to-OpenMP translations, LILAC-generated code was judged to be significantly more natural and maintainable than the output of the baseline compiler. The benchmarks used are unable to be translated by the formally verified LLM method. By building on a correct compiler translation, LILAC ensures that enhancements in naturalness do not come at the cost of correctness or applicability.

1 0

Chongyi Zheng will present his General Exam "Towards Unsupervised Pre-training for Reinforcement Learning" on Wednesday, January 14, 2026 at 11:00 AM in CS 301 and via zoom.
by CS Grad Department 09 Jan '26

09 Jan '26

Chongyi Zheng will present his General Exam "Towards Unsupervised Pre-training for Reinforcement Learning" on Wednesday, January 14, 2026 at 11:00 AM in CS 301 and via zoom. Zoom link: https://princeton.zoom.us/j/93974458862?pwd=1OO9PebfnbDu9vPe6XmmaP5TukAS7G.1 Committee Members: Benjamin Eysenbach (advisor), Peter Henderson, Karthik Narasimhan Abstract: Unsupervised reinforcement learning (RL) promises to solve new tasks without additional environment interaction, by leveraging prior experience and powerful representations. However, when and why unsupervised RL works remains poorly understood. In this presentation, I will explore one framework that enables unsupervised pre-training in RL, focusing on how temporal representations, training objectives, and policy adaptation jointly shape an agent’s ability to solve previously unseen tasks. I will propose new ways to understand a prior state-of-the-art unsupervised RL algorithm and develop a simplified version. I will also discuss some theoretical tools and diagnostic experiments that help us probe information acquired by unsupervised RL agents. Reading List: https://docs.google.com/document/d/1kEEdCLEyvFqUQ4bFDP-go0sBWNGIh7YBi2rwpSE… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

ECE Seminar: Mohammad Shahrad, The University of British Columbia | Dec 17th, B205 EQuad 4:30 pm
by Emily C. Lawrence 10 Dec '25

10 Dec '25

Taming The Cloud: For Providers, Developers, And The Planet ECE Seminar Date: Wednesday, December 17, 2025 Time: 4:30 pm Location: B205 Engineering Quadrangle Speaker: Mohammad Shahrad, The University of British Columbia Abstract: Cloud computing has made deploying applications at planetary scale almost effortless, but that very ease often creates inefficiencies in how resources are allocated and used. Characterizing production could-native settings at Azure and IBM Cloud, we found many sources of inefficiencies as well as opportunities for automated configuration and resource management. When provider-level automation falls short, lightweight developer tools and middleware can help close the gap, improving efficiency and predictability. As the global footprint of this sector is growing rapidly, efficiency can no longer be measured by performance and cost alone. We show why incorporating sustainability and energy impact into our metrics is increasingly essential during design and operation. Bio: Mohammad Shahrad is an Assistant Professor of Electrical and Computer Engineering at the University of British Columbia (UBC), where he leads the Cloud Infrastructure Research for Reliability, Usability, and Sustainability (CIRRUS) Lab. Prior to joining UBC, he was a Computer Science Lecturer at Princeton University, and he earned his Ph.D. in Electrical Engineering also from Princeton in 2020 under the supervision of Prof. David Wentzlaff. During his Ph.D., he spent a year at Microsoft Research working on cloud efficiency projects. Mohammad’s recent research has focused on advancing the sustainability, efficiency, and user experience of cloud-native systems. His research has been supported by industry and government partners including IBM, Rogers, Huawei, NSERC, and MITACS. Mohammad's research has won the USENIX Community Award, been deployed in production at two large cloud providers, and been featured in the press.

1 0

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