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

Colin Wang will present his masters' thesis "Evaluating Multimodal Models for Chart Understanding" on April 21, 2025 at 2pm in the AI Lab (41 William Street) Conference Room (Room 274).
by CS Grad Department 16 Apr '25

16 Apr '25

Colin Wang will present his masters' thesis "Evaluating Multimodal Models for Chart Understanding" on April 21, 2025 at 2pm in the AI Lab (41 William Street) Conference Room (Room 274). The members of his committee are as follows: Danqi Chen (Adviser) and Sanjeev Arora (reader). All are welcome to attend. Please see the title and abstract below. Title : Evaluating Multimodal Models for Chart Understanding Abstract : Chart understanding and reasoning play a pivotal role when applying Multimodal Large Language Models (MLLMs) to real-world tasks such as analyzing scientific papers or financial reports. In this work, we propose CharXiv, a comprehensive evaluation suite involving 2,323 natural, challenging, and diverse charts from arXiv papers. CharXiv includes two types of questions: 1) descriptive questions about examining basic chart elements and 2) reasoning questions that require synthesizing information across complex visual elements in the chart. To ensure quality, all charts and questions are handpicked, curated, and verified by human experts. Our results reveal a substantial, previously underestimated gap between the reasoning skills of the strongest proprietary model and the strongest open-source model. All models lag far behind human performance of 80.5%, underscoring weaknesses in the chart understanding capabilities of existing MLLMs. Since its release, CharXiv has gained significant real-world adoption. It has been used internally by organizations such as OpenAI and Anthropic, and serves as an official benchmark for models including QwenVL, InternVL, and the Doubao series. CharXiv has also become a standard testbed for academic research in visual reasoning and has inspired the development of more challenging and visually-diverse evaluations in the AI community. In this talk, I will present the design and curation process of CharXiv, share key findings and failure modes from model evaluations, and reflect on CharXiv's ongoing impact within the broader research community.

1 0

Brendan Wang will present his MSE talk “Computational Modeling of Local Insulation of Chromatin in Drosophila” on Monday, April 21, 2025 at 1:00 pm in Carl Icahn 101.
by CS Grad Department 16 Apr '25

16 Apr '25

Brendan Wang will present his MSE talk “ Computational Modeling of Local Insulation of Chromatin in Drosophila ” on Monday, April 21, 2025 at 1:00 pm in Carl Icahn 101. The members of his committee are as follows: Yuri Pritykin(Adviser) and Mona Singh (reader) All ae welcome to attend. Please see abstract below. In Drosophila melanogaster, chromatin is partitioned into self-interacting regions called topologically-associated domains (TADs). At TAD boundaries, architectural proteins called insulators can provide insulation or the ability to prevent interactions between gene regulatory elements between neighboring domains. While existing methods have analyzed the sequence patterns underlying TAD boundaries, they have not thoroughly characterized how such patterns contribute to the insulatory function of TADs. Characterizing how the binding of insulators to DNA influences insulation in a genome-wide manner is crucial to deciphering the cis-regulatory rules cells use to regulate gene expression through chromatin organization. To this end, we propose Domino, a sequence-based deep learning method that predicts local insulation. Our results demonstrate that Domino can uncover the sequence patterns underlying local insulation in Drosophila and accurately predict in-silico the effects of experimental perturbations.

1 0

Natalia Khotiaintseva will present her MSE talk "GPU Acceleration of Adaptive Mesh Refinement via a Hashtable" on Monday, April 21, 2025 at 3:20pm in Friend Center 004.
by Gradinfo 15 Apr '25

15 Apr '25

Natalia Khotiaintseva will present her MSE talk "GPU Acceleration of Adaptive Mesh Refinement via a Hashtable" on Monday, April 21, 2025 at 3:20pm in Friend Center 004 (basement) Title: GPU Acceleration of Adaptive Mesh Refinement via a Hashtable Abstract: In this work we explore acceleration of computations in AMR for astrophysical simulations. Modern scientific codes widely use Adaptive Mesh Refinement, and they need to leverage GPU supercomputing power. Specifically, this thesis demonstrates a hashtable implementation on GPU to store the grid structure. On a toy problem, we compare the performance of a typical matrix implementation against a hashtable implementation in a worst-case scenario — when all grid cells are uniformly refined. Acceleration of such large-scale astrophysical hydrodynamics simulations is critically important, making optimization techniques such as GPU-based hashtable storage essential. Our result is that there is no disadvantage in the hashtable implementation versus the matrix implementation. This result is crucial for future implementation of a hashtable-based AMR with dynamic refinement and complete hydrodynamics simulation. We also explore the application of the Grace Hopper superchip to AMR code and observe a significant speedup of ~2.3 in gradient calculation, and ~14.7 speedup in hashtable creation. Additionally, we design a linter program to streamline CUDA C++ debugging process, in particular for our AMR implementation. These approaches together aim to speed up scientific code; they combine large-scale scientific computing with cutting-edge tools in computer science.

1 0

Lucy He will present her General Exam "The Hidden Risks of Benign Data: Generative AI’s Safety and Copyright Challenges" on Tuesday, April 15, 2025 at 10:30 AM in CS 302 and via zoom.
by CS Grad Department 14 Apr '25

14 Apr '25

Lucy He will present her General Exam "The Hidden Risks of Benign Data: Generative AI’s Safety and Copyright Challenges" on Tuesday, April 15, 2025 at 10:30 AM in CS 302 and via zoom. Zoom link: https://princeton.zoom.us/my/lucyhe Committee Members: Danqi Chen (co-advisor), Peter Henderson (co-advisor), Tom Griffiths Abstract: Large language models and text-to-image models can be vulnerable to seemingly harmless inputs, resulting in safety degradation or copyright risks. In the first part of the talk, I will discuss critical issues related to preserving the safety of open-weight models. I will describe a general approach to identify toxicity-promoting data. Using this approach, we find that superficially benign data, such as a collection of bullet lists and math, can be even more harmful than explicitly toxic text. In the second part of the talk, I will introduce our work examining copyright concerns in text-to-image models. We introduce an evaluation framework to evaluate both the model’s copyright compliance and consistency with user input. We then apply the framework to study how benign prompts trigger copyrighted character generation and the potential (in)effectiveness of current mitigation strategies. These works critically examine how benign data can compromise model safety and trigger copyright risks, shedding light on a better understanding and improvement of these systems for deployment. Reading List: https://docs.google.com/document/d/1JHuGtMwAplGjamaEB_v8RuBHUUrZRnxGeZwceUT… 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

Jiayi Geng will present her MSE thesis "Understanding the Failure Modes in Large Language Models"on 4/16 at 2pm in Friend 005.
by Gradinfo 14 Apr '25

14 Apr '25

Jiayi Geng will present her MSE thesis "Understanding the Failure Modes in Large Language Models"on 4/16 at 2pm in Friend 005. Date & Time: April 16th, 2:00–2:30 PM Location: Friend Center, Room 005 Thesis Title: Understanding the Failure Modes in Large Language Models Thesis advisors: Danqi Chen, Tom Griffiths Presentation Abstract: Using AI to create autonomous researchers has the potential to accelerate discovery. A prerequisite for this vision to become reality is assessing how well an AI model can identify the underlying structure of a system from its behavior. In this paper, we explore the question of whether a large language model (LLM) can learn from passive observations and actively collect informative data to refine its own hypotheses. To answer this question, we investigate the ability of LLMs to reverse-engineer four types of black-box systems, chosen to represent problems that might appear in different domains of research: list-mapping programs, formal languages, mathematical equations and board games. We use the Bayesian agents as a normative reference to study the gap between LLMs and ideal inference in reverse-engineering black-box systems. Through extensive experiments, we show that while LLMs have difficulty reverse-engineering these systems from observations alone, data generated by LLM-driven interventions can be effective. By testing edge cases, the LLM is able to refine its own hypotheses and avoid failure modes such as overcomplication, where the LLM falsely assumes prior knowledge about the black box, and overlooking, where the LLM fails to incorporate observations. These insights provide practical guidance for helping LLMs more effectively reverse-engineer black-box systems, supporting their use in making new discoveries.

1 0

Yaxin Tu will will present her General Exam "Separating Succinct Non-Interactive Arguments from Cryptographic Assumptions" on April 15, 2025 at 12pm in CS 401.
by Nicki Mahler 10 Apr '25

10 Apr '25

Yaxin Tu will will present her General Exam "Separating Succinct Non-Interactive Arguments from Cryptographic Assumptions" on April 15, 2025 at 12pm in CS 401. Committee: Alex Lombardi (adviser), Ran Raz(adviser), and Pravesh Kothari . Link to abstract and reading list: [ https://docs.google.com/document/d/1FmkljJIThGRgeHYmuChrSgzIas9uP2mUL41lulM… | https://docs.google.com/document/d/1FmkljJIThGRgeHYmuChrSgzIas9uP2mUL41lulM… ] Abstract: An argument system (computationally sound proof) for NP is succinct and non-interactive, if it contains a single prover-to-verifier message whose length is polylogarithmic the instance and witness sizes. In the past decade, the question of what cryptographic assumptions are sufficient and/or necessary for constructing SNARG, has received answers in both directions. On the one hand, the work of Gentry-Wichs’13 exhibits a barrier of using standard assumptions to construct SNARG, showing that efficient black-box reductions cannot prove the security of any SNARG construction based on any falsifiable cryptographic assumption. On the other hand, a recent line of work bypasses the Gentry-Wichs barrier by constructing SNARG based on iO and one-way functions, with a super-polynomial time security reduction. We prove an impossibility result that rules out black-box constructions of SNARG from polynomially-secure iO and one-way functions. Our result extends the Gentry-Wichs barrier, showing that there’s no black-box SNARG construction from even non-falsifiable assumptions like iO, with an efficient security reduction. Our result also shows that the super-polynomial time reduction is inherent for black-box constructions of SNARG from iO.

1 0

Simon Park will present his MSE talk “Instruct-SkillMix: A Powerful Pipeline for LLM Instruction Tuning” on Tuesday, April 15, 2025 at 12:00pm in Room 238, 41 William Street (AI Lab).
by CS Grad Department 10 Apr '25

10 Apr '25

Simon Park will present his MSE talk “Instruct-SkillMix: A Powerful Pipeline for LLM Instruction Tuning” on Tuesday, April 15, 2025 at 12:00pm in Room 238, 41 William Street (AI Lab). The members of his committee are as follows: Sanjeev Arora (Adviser) and Danqi Chen (reader) All ae welcome to attend. Please see abstract below. We introduce INSTRUCT-SKILLMIX, an automated approach for creating diverse, high quality SFT data for instruction-following. The pipeline involves two stages, each leveraging an existing powerful LLM: (1) Skill extraction: uses the LLM to extract core “skills” for instruction-following by directly prompting the model. This is inspired by “LLM metacognition” of (Didolkar et al., 2024); (2) Data generation: uses the powerful LLM to generate (instruction, response) data that exhibit a randomly chosen pair of these skills. Here, the use of random skill combinations promotes diversity and difficulty. The estimated cost of creating the dataset is under $600. Vanilla SFT (i.e., no PPO, DPO, or RL methods) on data generated from INSTRUCT-SKILLMIX leads to strong gains on instruction following benchmarks such as AlpacaEval 2.0, MT-Bench, and WildBench. With just 4K examples, LLaMA-3-8B-Base achieves 42.76% length-controlled win rate on AlpacaEval 2.0, a level similar to frontier models like Claude 3 Opus and LLaMA-3.1-405B-Instruct. Ablation studies also suggest plausible reasons for why creating open instruction-tuning datasets via naive crowd-sourcing has proved difficult. In our dataset, adding 20% low quality answers (“shirkers”) causes a noticeable degradation in performance. The INSTRUCT-SKILLMIX pipeline seems flexible and adaptable to other settings.

1 0

Shuyao Zhou will present her MSE talk “Code It, See It, Touch It: Vision-Powered AR Programming Tool for Children” on Tuesday, April 15, 2025 at 11:00am in CS 402.
by CS Grad Department 10 Apr '25

10 Apr '25

Shuyao Zhou will present her MSE talk “Code It, See It, Touch It: Vision-Powered AR Programming Tool for Children” on Tuesday, April 15, 2025 at 11:00am in CS 402. The members of her committee are as follows: Andrés Monroy-Hernández (Adviser) and Parastoo Abtahi (reader) All ae welcome to attend. Please see abstract below. We present an augmented reality (AR) authoring tool that enables children to create interactive experiences by linking visual code blocks to physical objects recognized by vision-based AI models. Building on the accessibility of block-based programming, our system encourages creativity and exploration by extending traditional visual authoring tools into the physical environment. We evaluated the system through workshops with 20 children (ages 7–16) across the United States and Argentina, where participants collaboratively designed AR experiences. Our findings suggest that the system supports playful, explorative learning, while also highlighting challenges in real-time object tracking accuracy and responsiveness.

1 0

CS Colloquium speakers: week of February 26
by Emily C. Lawrence 10 Apr '25

10 Apr '25

CS Colloquium speakers Speaker: Jialin Ding, Amazon Web Services Date: Monday, February 26 Time: 12:30pm EST Location: CS 105 Host: Wyatt Lloyd Event page: https://www.cs.princeton.edu/events/26578 Register for live-stream online here: https://princeton.zoom.us/webinar/register/WN_10z_l8MYRr21ZOuPqA8Odw Title: Instance-Optimization: Rethinking Database Design for the Next 1000X Abstract: Modern database systems aim to support a large class of different use cases while simultaneously achieving high performance. However, as a result of their generality, databases often achieve adequate performance for the average use case but do not achieve the best performance for any individual use case. In this talk, I will describe my work on designing databases that use machine learning and optimization techniques to automatically achieve performance much closer to the optimal for each individual use case. In particular, I will present my work on instance-optimized database storage layouts, in which the co-design of data structures and optimization policies improves query performance in analytic databases by orders of magnitude. I will highlight how these instance-optimized data layouts address various challenges posed by real-world database workloads and how I implemented and deployed them in production within Amazon Redshift, a widely-used commercial database system. Bio: Jialin Ding is an Applied Scientist at AWS. Before that, he received his PhD in computer science from MIT, advised by Tim Kraska. He works broadly on applying machine learning and optimization techniques to improve data management systems, with a focus on building databases that automatically self-optimize to achieve high performance for any specific application. His work has appeared in top conferences such as SIGMOD, VLDB, and CIDR, and has been recognized by a Meta Research PhD Fellowship. To learn more about Jialin’s work, please visit [ https://jialinding.github.io/ | https://jialinding.github.io/ ] . CSML/CS/PSY Colloquium Speaker: Brenden Lake, New York University Date: Tuesday, February 27 Time: 12:30pm EST Location: CS 105 Host: Tom Griffiths Event page: https://csml.princeton.edu/events/csmlpsycs-seminar Register for live-stream online here: https://princeton.zoom.us/webinar/register/WN_aUfLfVQFSIqV2RDSYu4big Title: Towards more human-like learning in machines: Bridging the data and generalization gaps Abstract: There is an enormous data gap between how AI systems and children learn: The best LLMs now learn language from text with a word count in the trillions, whereas it would take a child roughly 100K years to reach those numbers through speech (Frank, 2023, "Bridging the data gap"). There is also a clear generalization gap: whereas machines struggle with systematic generalization, children can excel. For instance, once a child learns how to "skip," they immediately know how to "skip twice" or "skip around the room with their hands up" due to their compositional skills. In this talk, I'll describe two case studies in addressing these gaps. 1) The data gap: We train deep neural networks from scratch (using DINO, CLIP, etc.), not on large-scale data from the web, but through the eyes and ears of a single child. Using head-mounted video recordings from a child as training data (<200 hours of video slices over 26 months), we show how deep neural networks can perform challenging visual tasks, acquire many word-referent mappings, generalize to novel visual referents, and achieve multi-modal alignment. Our results demonstrate how today's AI models are capable of learning key aspects of children's early knowledge from realistic input. 2) The generalization gap: Can neural networks capture human-like systematic generalization? We address a 35-year-old debate catalyzed by Fodor and Pylyshyn's classic article, which argued that standard neural networks are not viable models of the mind because they lack systematic compositionality -- the algebraic ability to understand and produce novel combinations from known components. We'll show how neural networks can achieve human-like systematic generalization when trained through meta-learning for compositionality (MLC), a new method for optimizing the compositional skills of neural networks through practice. With MLC, neural networks can match human performance and solve several machine learning benchmarks. Given these findings, we'll discuss the paths forward for building machines that learn, generalize, and interact in more human-like ways based on more natural input. Related articles: Vong, W. K., Wang, W., Orhan, A. E., and Lake, B. M (2024). Grounded language acquisition through the eyes and ears of a single child. Science, 383, 504-511. Orhan, A. E., and Lake, B. M. (in press). Learning high-level visual representations from a child’s perspective without strong inductive biases. Nature Machine Intelligence. Lake, B. M. and Baroni, M. (2023). Human-like systematic generalization through a meta-learning neural network. Nature, 623, 115-121. Bio: Brenden M. Lake is an Assistant Professor of Data Science and Psychology at New York University. He received his M.S. and B.S. in Symbolic Systems from Stanford University in 2009, and his Ph.D. in Cognitive Science from MIT in 2014. He was a postdoctoral Data Science Fellow at NYU from 2014-2017. Brenden is a recipient of the Robert J. Glushko Prize for Outstanding Doctoral Dissertation in Cognitive Science and the MIT Technology Review 35 Innovators Under 35. His research was also selected by Scientific American as one of the 10 most important advances of 2016. Brenden's research focuses on computational problems that are easier for people than they are for machines, such as learning new concepts from just a few examples, learning by asking questions, learning by generating new goals, and learning by producing novel combinations of known components. CS Colloquium Speaker: Eric Mitchell, Stanford University Date: Thursday, February 29 Time: 12:30pm EST Location: CS 105 Host: Karthik Narasimhan Event page: https://www.cs.princeton.edu/events/26587 Register for live-stream online here: TBA Talk info TBA

1 22

Yanjin Chen will present her MSE talk “Predictive modeling reveals novel sequence features of prime editing efficiency and precision” on Thursday, April 10, 2025 at 1:30pm in LSI 253.
by CS Grad Department 07 Apr '25

07 Apr '25

Yanjin Chen will present her MSE talk “ Predictive modeling reveals novel sequence features of prime editing efficiency and precision ” on Thursday, April 10, 2025 at 1:30pm in LSI 253. The members of hercommittee are as follows: Yuri Pritykin (Adviser) and Britt Adamson (reader) All ae welcome to attend. Please see abstract below. Prime editing (PE) is an exciting genome editing technology that enables precise genetic modifications without requiring donor DNA or inducing double-strand breaks. However, its relatively low efficiency limits broader applications. PE efficiency largely depends on the PE guide RNA (pegRNA) design, which consists of a primer binding site (PBS) and a reverse transcriptase template (RTT) that contains the intended edit. Various computational models have been developed to predict pegRNA efficiencies across different target sites and edit types. Most existing models, however, rely on deep learning, making it difficult to interpret the contributions of individual features and their combinatorial effects on editing efficiency. Besides, existing datasets provide limited coverage in terms of the sequence space for each editing length at a single target DNA locus, offering few insights into how the intended edit sequence influences PE efficiency. Recently, in collaboration with Britt Adamson’s lab at Princeton, we acquired new PE screen datasets including 1-25 bps substitutions, insertions, and deletions at two target loci, in the K562 cell line and a variant of this cell line that is mismatch repair (MMR)-deficient. This thesis aims to dissect the influence of sequence attributes of the desired edits and cellular repair pathways on PE outcomes. We introduce generalized linear models with LASSO feature selection to identify important positional sequence features predictive of editing rates and unintended indel (insertion or deletion) rate at each target locus. Separate models were trained on insertion and substitution datasets. Our results demonstrate that linear models achieve performance comparable to that of deep neural networks, suggesting that sequence features alone are sufficient determinants of editing efficiency and precision. Beyond predictive accuracy, linear models offer greater interpretability. Notably, we observe a new factor in both our dataset and previously published datasets: for insertion edits, incorporating cytosine (C) at the first editing position increases editing efficiency, providing a practical guideline for optimizing pegRNA design. Additionally, our model recaptures the C-C mismatch preference in parental cell lines versus in MMR-deficient cell lines. In summary, this work offers novel insights into the sequence composition and cellular repair mechanisms in PE editing efficiency, which can inform researchers to design PE experiments to achieve higher efficiency.

1 0