- talks - lists.cs.princeton.edu

Carlos Jimenez will present his FPO "Agent in the Shell: Establishing Agentic Software Engineering" on Tuesday, April 28, 2026 at 3:30 in Friend 110.
by CS Grad Department 23 Apr '26

23 Apr '26

Carlos Jimenez will present his FPO "Agent in the Shell: Establishing Agentic Software Engineering" on Tuesday, April 28, 2026 at 3:30 in Friend 110. The members of Carlos’ committee are as follows: Examiners: Karthik Narasimhan (Adviser), Danqi Chen, Sanjeev Arora Readers: Karthik Narasimhan, H. Sebastian Seung 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 the talk. Abstract follows below: Neural language models were put to work on code almost as soon as they were useful for anything. Their capabilities have grown from auto-complete to chat interactions, and we establish a third paradigm, in which language model work as agents that autonomously complete full software engineering tasks. The first benchmarks for code language models focused on isolated function synthesis: a specification goes in, a function comes out. Real software engineering is a different activity; it requires directory navigation, running code, and editing files that already exist inside a larger system. We propose an evaluation task that measures progress using precise execution verification and demonstrate the benefits of the agentic approach, in which the model works on the repository interactively. We first introduce SWE-bench, a benchmark drawn from real GitHub issues across twelve open-source Python repositories, each scored by whether a system's patch passes the repository's own tests. The strongest retrieval-augmented baselines resolve under 2% of the test set, establishing that single-shot generation is not enough. We then present SWE-agent, an agent that gives a language model interactive access to the repository through the shell and an interface specially designed for software engineering agents. With model weights fixed, SWE-agent resolves 12.47% of SWE-bench, and ablations isolate the interface as an axis of progress separable from model scale. We finally introduce SWE-bench Multimodal, which reapplies the collection pipeline to visual JavaScript repositories and shows that the task definition transfers across language and modality while systems tuned to Python largely do not. In the years since its release, SWE-bench has been widely adopted as a core evaluation target for frontier models, and reported resolve rates have risen above 90% on a human-validated subset as of April 2026. We analyze the work that has followed and find that model improvement has far outpaced scaffold design: a minimal bash-only agent tracks the frontier within a few points, reflecting both raw capability gains and the tailoring of frontier models toward agentic software engineering as an explicit training target.

1 0

Tim Kosfeld will present his MSE thesis "Supervised linear modeling and integration of ATAC-seq" on Monday, April 27th, 2026 in Bendheim House 103 at 10:30am.
by CS Grad Department 23 Apr '26

23 Apr '26

Tim Kosfeld will present his MSE thesis "Supervised linear modeling and integration of ATAC-seq" on Monday, April 27th, 2026 in Bendheim House 103 at 10:30am. Advisor: Yuri Pritykin Reader: Ben Raphael All are welcome to attend. Abstract: Single-cell assay for transposase-accessible chromatin with sequencing (scATAC-seq) has been widely used to profile chromatin accessibility and explore the epigenomic properties in heterogeneous cell populations. Moreover, transcription factors (TFs) bind to accessible sites, and scATAC-seq analysis could associate TF binding specificities, typically available as sequence motifs, with variance in chromatin accessibility. Here we present a new supervised TF motif regression method, MotifReg, for computational analysis and interpretation of scATAC-seq data. We comprehensively compare MotifReg with other popular methods for motif and sequence analysis in scATAC-seq data and demonstrate superior performance and benefits of MotifReg. MotifReg can efficiently analyze large-scale scATAC-seq datasets and learn interpretable associations between TF motifs and accessible peaks. The resultant low-dimensional cell representations in the TF activity space facilitate downstream cell type-specific analyses of scATAC-seq data. The supervised formulation in MotifReg enables ranking TF motifs by significance of their association with chromatin accessibility, implicating cell type-specific TFs. Furthermore, by comparing MotifReg performance across data analysis settings, we develop guidelines and best practices for scATAC-seq peak calling, data preprocessing and normalization. In sum, MotifReg is a convenient and statistically sound new open source computational method for scATAC-seq data analysis and characterization of genome-wide TF activities.

1 0

Ryan Oet will present his MSE thesis "Organizing the Organizer: Developing Digital Tools to Support Gig Worker Labor Advocacy" on Friday, April 24th, 2026 in Friend 110 at 11:30am and Zoom.
by CS Grad Department 23 Apr '26

23 Apr '26

Ryan Oet will present his MSE thesis "Organizing the Organizer: Developing Digital Tools to Support Gig Worker Labor Advocacy" on Friday, April 24th, 2026 in Friend 110 at 11:30am and Zoom. Zoom Link: https://princeton.zoom.us/j/94830827553 Advisor: Andrés Monroy-Hernández Reader: Parastoo Abtahi All are welcome to attend. Abstract: Algorithmic management practices in the ridehail industry cause financial, physical, and emotional harm to drivers. New regional policies in the United States mandate partial protections for platform workers, including minimum pay guarantees, paid sick leave, and appeals processes for wrongful deactivations. Organizers require easily accessible data and efficient workflows to effectively advocate for workers, enforce legal compliance from platforms, and support drivers navigating bureaucratic systems. Existing workflows are tedious and error-prone. To address this challenge, we collaborated with the labor association Drivers Union to design and implement automated data access tools for policy advocacy and sick leave auditing. We conducted semi-structured evaluative interviews with organizers and, through thematic analysis, found that these tools accurately target their current operational needs. Organizers also anticipate that the tools will substantially expand the efficiency and scale of worker-support processes. Looking ahead, they expressed a clear need for descriptive analytical solutions to answer increasingly complex questions and new methods of data collection to capture key metrics missing from existing aggregation approaches. We contribute reflections on researcher-organizer collaborations in the ridehail industry, two computational artifacts for supporting labor advocacy, and directions for further work in this space.

2 1

Lana Glisic will present her MSE thesis "LiDAR for Children's Coding and Creativity in Augmented Reality" on Friday, April 24th, 2026 in Green Hall 3-S-14 at 11:00am.
by CS Grad Department 23 Apr '26

23 Apr '26

Lana Glisic will present her MSE thesis "LiDAR for Children's Coding and Creativity in Augmented Reality" on Friday, April 24th, 2026 in Green Hall 3-S-14 at 11:00am. Advisor: Andrés Monroy-Hernández Reader: Parastoo Abtahi All are welcome to attend. Abstract: Abstract: Children often encounter Augmented Reality (AR) through social media and video games, where engagement is largely limited to consumption rather than creation. An AR authoring environment is presented that enables the construction of interactive experiences by programming a virtual character to navigate and climb obstacles within the physical environment. The system leverages LiDAR-based depth sensing to reconstruct surrounding geometry and provides a block-based programming interface to define character behaviors and interactions. By combining spatial sensing with visual programming, expanded opportunities for coding, creativity, and the development of spatial reasoning are supported. A user study involving 6 children aged 5-11 in the United States was conducted to evaluate the usability and engagement of the LiDAR-driven features. Participants successfully created interactive AR narratives featuring virtual characters that traversed and climbed real-world structures in three-dimensional space. The implementation and potential applications of the system in educational contexts are examined. Findings indicate that the approach promotes learning through creative exploration while also revealing challenges related to the accuracy and robustness of real-time scene reconstruction and geometric traversal.

2 1

Tanvi Namjoshi will present her MSE thesis "Major Decisions: Evaluating LLM Homogenization and Personalization in Educational Advisory Tasks" on Friday, April 24th, 2026 in CS 401 at 2:30 pm.
by CS Grad Department 22 Apr '26

22 Apr '26

Tanvi Namjoshi will present her MSE thesis "Major Decisions: Evaluating LLM Homogenization and Personalization in Educational Advisory Tasks" on Friday, April 24th, 2026 in CS 401 at 2:30 pm. Advisor: Lydia T. Liu Reader: Manoel Horta Ribeiro All are welcome to attend. Abstract: Large language models (LLMs) are increasingly being used for advisory tasks in high-stakes settings such as education. While researchers have evaluated their agreement with human advisors, there has been limited work looking at the challenges that emerge with adoption at scale. Unlike factual question-answering, where homogeneous outputs are desirable, and creative generation, where scale is rarely a concern, advisory tasks require recommendations that are both individually relevant and collectively diverse. In this work, we introduce a formal evaluation framework for homogenization and personalization for this class of problems, and apply it to the task of undergraduate major selection. To study the homogenization of responses we use a Shannon entropy-based approach to study 17 open-source LLMs across three levels of provided student context. We find that global recommendation entropy remains below our real-world enrollment baseline across all context levels, and several majors -- disproportionately humanities disciplines -- are never recommended by most models. While models are semantically responsive to the individual student profiles, they exhibit systematic narrowing; across all models, over 50% of the recommendations are concentrated within the top five majors per model. These findings suggest that naive prompt-based advising is insufficient for institutional deployment and highlight the need for population-level metrics to complement capability assessments in future AI system evaluations.

1 0

Rohan Baskar Prabhakar will present his FPO "Hardware-Aware Software Optimizations for and with Machine Learning" on Monday, May 4, 2026 at 10am in CS 302.
by Gradinfo 22 Apr '26

22 Apr '26

Rohan Baskar Prabhakar will present his FPO "Hardware-Aware Software Optimizations for and with Machine Learning" on Monday, May 4, 2026 at 10am in CS 302. His committee is as follows: Examiners: David Wentzlaff (adviser), Aarti Gupta and Ravi Netravali Readers: Jialin Ding and Prateek Mittal Abstract: In recent years, fundamental limits in semiconductor manufacturing have caused a gradual decline in the steady cadence of hardware performance scaling. With pivotal trends like Moore's law and Dennard scaling ending, there is a growing need to ensure that software workloads execute at peak efficiency. Towards this objective, this dissertation describes three hardware-aware optimizations that either accelerate machine learning inference or use machine learning to optimize single-threaded CPU workloads. First, this dissertation presents Kraken, a variation of the Transformer architecture designed to improve the efficiency of tensor parallelism during inference. The new model architecture incorporates an innate notion of model parallelism that complements the topology of multi-device inference hardware and allows communication to overlap with compute. Experiments demonstrated that while preserving the language modeling performance of standard Transformers, the Kraken architecture improves Time To First Token by a geomean of 35.6% across a range of model configurations. Second, the dissertation investigates the feasibility of integrating a binary classifier to increase the efficiency of the verification phase in speculative decoding. Although speculative decoding is effective in accelerating the decode step of Transformer inference, performance gains are limited to small batch sizes where kernels are memory-bound as opposed to compute-bound. Using n-gram matching as the draft method and intermediate activations from early Transformer layers as input allows binary classifiers to filter 75% of draft tokens that would otherwise be rejected. Doing so decreases the effective batch size of the verification step, expanding the range of scenarios in which speculative decoding is effective. Finally, the dissertation introduces Toggle, a dynamic optimization system that enables single-threaded CPU programs to switch both compilers and optimization choices at runtime. Relying on the premise that the best choice of compiler and optimizations is a function of the current program phase and input, the system uses otherwise idle inference accelerators and statistics from hardware performance counters to perform continuous optimization. When evaluated on the SPEC CPU 2017 benchmark suite, integrating Toggle improved program runtime by a geomean of 4.32%, effectively extracting a year's worth of hardware performance gains.

1 0

Max Gupta will present his MSE talk "Inductive Biases for Relational Reasoning in Neural Networks" on Thursday 4/23/26 at 9:30am in Friend 004.
by Gradinfo 22 Apr '26

22 Apr '26

Max Gupta will present his MSE talk " Inductive Biases for Relational Reasoning in Neural Networks" on Thursday 4/23/26 at 9:30am in Friend 004. Advisor: Tom Griffiths Reader: Brenden Lake Title: Inductive Biases for Relational Reasoning in Neural Networks Abstract: This thesis examines how the inductive biases of neural networks (architecture, training data, training objective, and weight initialization) affect the emergence of visual relational reasoning capabilities. I present both prior published work and ongoing experimental work under review. By examining the effect of each of these four inductive biases, the thesis aims to disentangle the relative influence of each bias on the relational reasoning abilities of vision models. I cover a series of increasingly large vision models across a suite of increasingly abstract relational reasoning tasks. Along the task complexity axis, I start with the most fundamental relation: same-different reasoning, classifying two objects as 'same' or 'different'. Scaling up task complexity along this axis to higher order relational tasks involves composing intermediate same-different judgements in tasks like odd-one-out and relational match to sample, tasks often used in cognitive science and visual IQ tests. I compare with human data to understand certain perceptual biases towards regularity and simplicity in humans versus neural networks. Throughout, I focus on meta-learning (learning to learn) as a particularly effective training method for endowing neural networks with relational reasoning capabilities. In particular, the episodic partitioning of training data into mini train/test sets enables strong, generalizable relational reasoning even in tiny convolutional networks and at a fraction of the scale of data as pre-trained vision models. Throughout, I employ interpretability techniques for understanding what is going on inside models meta-trained from scratch and inside off-the-shelf pre-trained models. One compelling finding is the recurrence of a two-stage perceptual-to-relational reasoning pipeline that vision models across architectures and scales employ to reason relationally. Moreover, we find that the concepts of 'same' and 'different' are more strongly represented as orthogonal dimensions in smaller, meta-trained models than larger off-the-shelf models like CLIP and DINOv2. In these larger models, I also find correlational evidence that similar attention heads are employed for relational reasoning across this suite of tasks. I highlight this evidence as a promising avenue for future work investigating causal evidence for a shared subspace encoding first and second order relations. Initial results show that this is easiest to do in smaller models for simpler relations like same-different, which live in relatively low-dimensional subspaces. I conclude by addressing the possibility for intervention in these subspaces and downstream consequences for model controllability.

1 0

Lauren Wang will present her General Exam "Explainable and Controllable Augmented Reality Interfaces through Embodied Interaction and Situated Visualization" on Monday, April 27, 2026 at 10:30 AM in CS 401.
by CS Grad Department 22 Apr '26

22 Apr '26

Lauren Wang will present her General Exam "Explainable and Controllable Augmented Reality Interfaces through Embodied Interaction and Situated Visualization" on Monday, April 27, 2026 at 10:30 AM in CS 401. Committee Members: Parastoo Abtahi (advisor), Andrés Monroy-Hernández, Tom Griffiths Abstract: As technologies become more sophisticated, users are expected to interact with intelligent systems whose reasoning, capabilities, and limitations are opaque. I identify two shortcomings in such human-computer interfaces: a gulf of execution[1], where users struggle to communicate abstract, multimodal, or spatial intent, and a gulf of evaluation[1], where limited feedback makes it difficult to understand how inputs influence outcomes. Current language-based interfaces lack the bandwidth to convey non-textual information and are ambiguous in spatial contexts. My work investigates augmented reality (AR) as a unifying interaction paradigm to bridge both gaps by enabling embodied multimodal input (voice, gesture, gaze) to better capture user intent, and by providing in-situ, spatially grounded visualizations that help establish shared mental models. I demonstrate this approach across real-world applications. In human-robot interaction, failures arise when users do not understand what robots can and cannot do. I develop Explainable OOHRI[2] to expose robot constraints and action possibilities through AR visualizations, and enable users to issue personalized instructions via direct manipulation of GhostObjects[3] —life-size, world-aligned virtual twins. Extending this paradigm to interaction with generative AI, I explore embodied 3D/4D scene editing for content creation and enhancing human-human communication. I study methods that fuse multimodal inputs to infer user intent across multiple levels of abstraction, while producing immediate, interpretable outputs that support counterfactual exploration and maintain controllability throughout iterative design processes. Taken together, my research aims to advance AR interfaces that make intelligent systems more explainable, controllable, and aligned with human intent, grounded in cognitive theories and designed to drive impacts in real-world settings. Reading List: https://docs.google.com/document/d/1RLaQ7eUAq7MStWIkt8QvtzNj-UkaHyJQJboRh7n… 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

Alexander Wettig will present his FPO "Building Better Language Models with Data Curation" on Tuesday, April 28, 2026 at 2:00 PM in FC 110.
by CS Grad Department 21 Apr '26

21 Apr '26

Alexander Wettig will present his FPO "Building Better Language Models with Data Curation" on Tuesday, April 28, 2026 at 2:00 PM in FC 110. Zoom Link: [ https://princeton.zoom.us/j/98043325388?pwd=IIFV4labMYR0YyBXNC5ocNaaZ2i8aQ.… | https://princeton.zoom.us/j/98043325388?pwd=IIFV4labMYR0YyBXNC5ocNaaZ2i8aQ.… ] The members of Alex’s committee are as follows: Examiners: Danqi Chen (Adviser), Sanjeev Arora, Karthik Narasimhan Readers: Peter Henderson, Ludwig Schmidt (Stanford) 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: Training data is among the most consequential choices when developing a language model, and the key ingredients of successful data curation have become closely guarded secrets of industrial research labs. This thesis works toward making data curation an open and reproducible science, developing tools and methods for shaping the training data distribution of language models and studying how quality filtering, domain mixing, and data composition affect downstream capabilities. Our methodology is grounded in data ablations: controlled experiments that vary the training data while holding all other variables fixed. The major part of this thesis focuses on data curation for pre-training, where we introduce three complementary techniques. (1) QuRating selects documents by quantifying human notions of text qualities, such as writing style, educational value, and required expertise. With the right balance of quality and diversity, we curate training datasets that produce strong language models that perform on par with models trained with 50% more compute. (2) WebOrganizer operates at the corpus level, introducing topic and format taxonomies that classify web pages into meaningful categories to enable explicit optimization of the data mixture. This allows us to achieve well-calibrated performance across tasks, and the best results are achieved when domain mixing is combined with quality selection. (3) Finally, we consider the relationship of quality filtering and deduplication as part of the Olmo-3 project and develop quality-aware upsampling strategies used in curating trillions of tokens of high-quality pre-training data. In the second part, we extend our methodology to the setting of continued pre-training for long-context understanding. Through systematic ablations guided by reliable downstream evaluation, we identify good sources of long-context data and uncover details such as the importance of mixing both long and short documents. Our final recipe produces the ProLong model, which achieves state-of-the-art long-context performance at a fraction of the data budget of comparable efforts. We conclude by discussing how data curation will co-evolve with future progress in AI.

1 0

Alexander Wettig will present his FPO "Building Better Language Models with Data Curation" on Tuesday, April 28, 2026 at 2:00 PM in FC 110.
by CS Grad Department 21 Apr '26

21 Apr '26

Alexander Wettig will present his FPO "Building Better Language Models with Data Curation" on Tuesday, April 28, 2026 at 2:00 PM in FC 110. The members of Alex’s committee are as follows: Examiners: Danqi Chen (Adviser), Sanjeev Arora, Karthik Narasimhan Readers: Peter Henderson, Ludwig Schmidt (Stanford) 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: Training data is among the most consequential choices when developing a language model, and the key ingredients of successful data curation have become closely guarded secrets of industrial research labs. This thesis works toward making data curation an open and reproducible science, developing tools and methods for shaping the training data distribution of language models and studying how quality filtering, domain mixing, and data composition affect downstream capabilities. Our methodology is grounded in data ablations: controlled experiments that vary the training data while holding all other variables fixed. The major part of this thesis focuses on data curation for pre-training, where we introduce three complementary techniques. (1) QuRating selects documents by quantifying human notions of text qualities, such as writing style, educational value, and required expertise. With the right balance of quality and diversity, we curate training datasets that produce strong language models that perform on par with models trained with 50% more compute. (2) WebOrganizer operates at the corpus level, introducing topic and format taxonomies that classify web pages into meaningful categories to enable explicit optimization of the data mixture. This allows us to achieve well-calibrated performance across tasks, and the best results are achieved when domain mixing is combined with quality selection. (3) Finally, we consider the relationship of quality filtering and deduplication as part of the Olmo-3 project and develop quality-aware upsampling strategies used in curating trillions of tokens of high-quality pre-training data. In the second part, we extend our methodology to the setting of continued pre-training for long-context understanding. Through systematic ablations guided by reliable downstream evaluation, we identify good sources of long-context data and uncover details such as the importance of mixing both long and short documents. Our final recipe produces the ProLong model, which achieves state-of-the-art long-context performance at a fraction of the data budget of comparable efforts. We conclude by discussing how data curation will co-evolve with future progress in AI.

1 0