September 2026 - talks - lists.cs.princeton.edu

Taiming Lu will present his General Exam "Evaluating Agents on Open-Ended Problems" on Thursday, October 1, 2026 at 1:00 PM in CS 401.
by CS Grad Department 30 Sep '26

30 Sep '26

Taiming Lu will present his General Exam " Evaluating Agents on Open-Ended Problems " on Thursday, October 1, 2026 at 1:00 PM in CS 401 . Committee Members: Zhuang Liu (advisor), Danqi Chen , Karthik R. Narasimhan Abstract: Can language model agents solve problems when no one tells them how, or even what, to solve? Agents are becoming reliable at well-specified tasks, but many valuable problems are open-ended: the path to success is unknown, or the problem itself must first be discovered. In this talk, I study how to evaluate agents on both kinds of problems. I first introduce CEO-Bench, in which an agent runs a simulated startup for 500 days in a noisy, changing market. It reveals distinct strategic behaviors across frontier agents, and a large gap in their ability to consistently plan and execute long-horizon strategy in an open world. I then introduce SWEeper-Bench, in which an agent must find and fix unreported bugs in real web applications. The best agent passes 59% of tasks, yet agents pass 96% once told what the bug is. In both settings, agents execute well but lack the judgment to decide what matters. I conclude by arguing that spoken conversation offers a scalable way to close this gap: it can capture how people think across diverse, real-world settings, which current training data rarely contains. Reading List: https://docs.google.com/document/d/1vLPEA5_5p028XR5WAftHQIiBK-RANAqiWQJClOY… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

2 1

Today's ORFE Department Colloquia: 4:30 PM: Ankur Moitra, MIT
by Emily C. Lawrence 29 Sep '26

29 Sep '26

Today's ORFE Department Colloquia Please join us today for a light reception and pre-talk meet and greet at 4:00pm in Sherrerd Hall Atrium. Date & Time: Tuesday, September 29, 2026, 4:30 PM – 5:30 PM EDT Location: 101 - Sherrerd Hall Speaker: [ https://orfe.princeton.edu/events/2026/ankur-moitra-massachusetts-institute… | Ankur Moitra, Massachusetts Institute of Technology ] Title: Honeycombs and Sums of Hermitian Matrices, Revisited Abstract: Given the eigenvalues of Hermitian matrices A and B, can we determine all possible eigenvalues of A + B? The set of all possible eigenvalues turns out to be a polytope and a celebrated theorem of Knutson and Tao gives a complete characterization through geometric objects called honeycombs. The theorem can be explained to an undergraduate student in linear algebra, but can the proof be as well? We give a new proof. Instead of algebraic geometry, Schubert calculus and representation theory, we use induction and convexity and treat the problem from a polyhedral perspective. In the process, we give an intuitive explanation of how honeycombs arise as a type of model organism that distills the appropriate defining properties of Hermitian triples in more obvious ways. This is based on joint work with Alex Postnikov and Dora Woodruff. Bio: Ankur Moitra is the Norbert Wiener Professor of Mathematics as of July 2021. He was also appointed Director of the MIT Statistics and Data Science Center (SDSC). Moitra received his B.S. in electrical and computer engineering from Cornell in 2007. He completed his M.S. and Ph.D. degrees from MIT in computer science in 2009 and 2011 respectively, where he was advised by Tom Leighton and was supported by a Fannie and John Hertz Foundation Fellowship. He received a George M. Sprowls Award (best thesis) and a William A. Martin Award (best thesis) for his doctoral and master's dissertations. He then spent two years as an NSF CI Fellow at the Institute for Advanced Study, and jointly a senior postdoc in the computer science department at Princeton University. In 2013 he returned to MIT as an assistant professor in applied mathematics, and a principal investigator in CSAIL. He was promoted to Associate Professor in July, 2017, and received tenure in July 2019. His research spans a diverse set of topics, from statistical inference to optimization and approximation to codes and combinatorics. He is particularly interested in algorithmic problems with applications in machine learning and big data. In 2015 Ankur received an NSF CAREER Award and the Kelly Research Award. In 2016 he received the Sloan Research Fellowship and the Packard Fellowship in Science and Engineering. He also received the Rockwell International Career Development Professorship, July 2016 - June 2019. In 2018, Ankur received the Young Investigator Award from the Office of Naval Research, for his proposed research on "An Algorithmic Theory of Robustness." The MIT School of Science also selected Ankur for the Teaching Prize in Graduate Education.

1 0

Jonathan Williams will present his General Exam "Latent Reasoning via Looped Transformers" on Friday, October 2, 2026 at 10:30 AM in Friend 202 and via zoom.
by CS Grad Department 29 Sep '26

29 Sep '26

Jonathan Williams will present his General Exam "Latent Reasoning via Looped Transformers" on Friday, October 2, 2026 at 10:30 AM in Friend 202 and via zoom. Zoom link: https://princeton.zoom.us/j/4688637940 Committee Members: Karthik Narasimhan(advisor), Danqi Chen, Zhuang Liu Abstract: Looped Language Models (LoopLMs) perform multi-step latent reasoning prior to token generation and outperform conventional LLMs on reasoning benchmarks at smaller parameter budgets. However, attempts to further improve LoopLM reasoning with reinforcement learning have failed - standard objectives such as Group Relative Policy Optimization (GRPO) only assign credit to the final latent state, creating a fundamental mismatch with the model's internal computation. To resolve this, we introduce RLTT (Reward Latent Thought Trajectories), a reinforcement learning framework which distributes reward across the full latent reasoning trajectory. RLTT provides dense, trajectory-level credit assignment without relying on external verifiers and can directly replace GRPO with negligible overhead. Across extensive experiments with Ouro-1.4B/2.6B-Thinking under identical training and inference conditions, RLTT yields statistically significant improvements over GRPO on challenging mathematical reasoning benchmarks, improving mean accuracy over MATH-500, AIME24/26, and BeyondAIME by +5.8% on the 1.4B scale, and +10.9% on the 2.6B scale. Despite being trained exclusively on mathematics, RLTT also transfers effectively to non-mathematical reasoning benchmarks, demonstrating the effectiveness of trajectory-level credit assignment for reinforcement learning in LoopLMs. Reading List: https://docs.google.com/document/d/1vyxUCHi_cSj5DxAxKw5eGEDNmomCkWh6yoXiYei… 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

Sachin Konan will present his General Exam "On the Automation of Scientific Evaluation and Discovery" on Thursday, October 1, 2026 at 10:00 AM in CS 401 and via zoom.
by CS Grad Department 28 Sep '26

28 Sep '26

Sachin Konan will present his General Exam "On the Automation of Scientific Evaluation and Discovery" on Thursday, October 1, 2026 at 10:00 AM in CS 401 and via zoom. Zoom link: https://princeton.zoom.us/j/92647117538 Committee Members: Zhuang Liu (advisor), Manoel Horta Ribeiro, Danqi Chen Abstract: Large language models (LLMs) have become more capable of reasoning, tool use, and agentic interaction, but can they conduct end-to-end research? This talk examines two central challenges: evaluating research beyond task-specific rewards and enabling agents to acquire and apply knowledge throughout experimentation. First, I introduce PaperScorer, a family of models that predict conference acceptance from screenshots or text of research papers. We find that screenshots provide a stronger predictive signal than text and that PaperScorer performs well on both low- and high-rated papers, whereas existing frontier reasoning models struggle with the latter. This work offers a step toward automated research evaluation and reward modeling, using acceptance as a proxy rather than a complete measure of scientific value. Second, I present MOMO (Mixture of Model Optimization), a framework for diverse, adaptive search. Prior approaches combining LLMs with evolutionary search have typically focused on optimizing a single model. MOMO instead jointly optimizes multiple models with different pretraining distributions, with channels for sharing discoveries. This approach is motivated by the hypothesis that reinforcement learning with verifiable rewards sharpens a model’s pretraining distribution, while effective research requires diversity in the search space. By combining optimization across models with information sharing, MOMO aims to preserve diversity while allowing discoveries to inform subsequent exploration. The approach has achieved a world-record result on the Erdős minimum overlap problem. Reading List: https://docs.google.com/document/d/1GA-sBzGUf3aJTspBj1EOPzo09aHbmORLzBDWuFW… 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

Sanghyuk Chun will present his General Exam "CoMet: Context and Multiplicity Decomposition for Multimodal Uncertainty Estimation" on Monday, September 28, 2026 at 12:00 PM in CS 402 and via zoom.
by CS Grad Department 23 Sep '26

23 Sep '26

Sanghyuk Chun will present his General Exam "CoMet: Context and Multiplicity Decomposition for Multimodal Uncertainty Estimation" on Monday, September 28, 2026 at 12:00 PM in CS 402 and via zoom. Zoom link: [ https://princeton.zoom.us/j/3423252796?omn=95071290639 | https://princeton.zoom.us/j/3423252796?omn=95071290639 ] Committee Members: Olga Russakovsky (advisor), Ryan P. Adams, Zhuang Liu Abstract: Uncertainty estimation has been a long-standing challenge in AI models; it amounts to "knowing what you don't know," and metacognition is notoriously difficult even for humans (cf. the Dunning-Kruger effect). Although it is still far from solved even in simpler classification systems, tackling it in multimodal large language models (MLLMs) is becoming increasingly important. Within MLLMs, uncertainty can stem from any of the diverse sources as well as from their relationships, and further can stem from the unbounded answers in the open-ended setting. To tackle the issues, we propose CoMet, an MLLM uncertainty estimation method by decomposing uncertainty into a context-specific term and a multiplicity-specific term. The former captures ambiguity induced by the given context (e.g., task or prompt), while the latter captures how many plausible answers determined by the context remain compatible with the given input. We train a lightweight post-hoc uncertainty module to estimate these quantities, which enables efficient uncertainty estimation without autoregressive answer generation or repeated sampling. Experiments on various open-ended multimodal benchmarks, hallucination detection, and multiple-choice visual question answering benchmarks show that CoMet consistently improves uncertainty estimation over existing baselines while remaining efficient in practice. Reading List: https://docs.google.com/document/d/1vWS_333RK-l4NOXwzlYi6ffVFobgt5sWpf6vmbf… 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

Linzhan Mou will present his General Exam "Towards a Universal Animation Foundation Model" on Friday, September 25, 2026 at 1:00 PM in Friend 202 and via zoom.
by CS Grad Department 22 Sep '26

22 Sep '26

Linzhan Mou will present his General Exam "Towards a Universal Animation Foundation Model" on Friday, September 25, 2026 at 1:00 PM in Friend 202 and via zoom. Zoom link: https://princeton.zoom.us/j/2298042162 Committee Members: Szymon Rusinkiewicz (advisor), Adam Finkelstein, Zhuang Liu Abstract: Recent advances in automatic rigging now deliver animation-ready 3D assets at scale, yet generating the motion to drive them remains a bottleneck. Existing learned animators are topology-constrained: they rely on category-specific templates or require per-skeleton fine-tuning and reference motions at inference. In this talk, I will present UniMate, a unified foundation model that synthesizes articulated motion for arbitrary skeletons from a rigged 3D asset and a text prompt, with no test-time optimization or per-skeleton retraining. At its core is a Topology-Aware Diffusion Transformer (TADiT), which integrates skeletal topology into attention through three mechanisms: a graph-aware attention bias from pairwise joint relations and geodesic distances; Spec-RoPE, a spectral rotary position embedding that generalizes RoPE to arbitrary kinematic trees via the graph Laplacian; and a global topological conditioner attention pooled from the rest-pose skeleton. We also curate UniML3D, a dataset comprising 20 hours of text-paired motion sequences spanning bipedal, quadrupedal, avian, marine, insectoid, serpentine, and articulated rigid objects. Trained on this dataset, UniMate achieves state-of-the-art quality, generalization, and efficiency, while supporting zero-shot cross-topology transfer, in-betweening, expansion, and text-guided editing. I will conclude by discussing applications of UniMate in robotics, including simulation and motion tracking. Reading List: https://docs.google.com/document/d/1Q9KMTWEC1CssZZeu23N_Mn1d243AuGv8-5Cdltj… 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

Boya Zeng will present his General Exam "Towards Strong, Open, and Efficient Visual Generative Models" on Monday, September 21, 2026 at 9:00 AM in Friend 202 and via zoom.
by CS Grad Department 16 Sep '26

16 Sep '26

Boya Zeng will present his General Exam "Towards Strong, Open, and Efficient Visual Generative Models" on Monday, September 21, 2026 at 9:00 AM in Friend 202 and via zoom. Zoom link: https://princeton.zoom.us/j/92298024735 Committee Members: Zhuang Liu (advisor), Olga Russakovsky, Danqi Chen Abstract: Recent text-to-image models have achieved remarkable improvements in image quality and prompt following, but the sources of these gains are often difficult to identify. Leading models typically combine many changes in architecture, data, and training recipe, while providing limited ablations and incomplete information about their training data and settings. As a result, modern text-to-image diffusion models still lack consensus on many important design choices. In this talk, I will present a systematic study of the modeling and data choices that drive text-to-image diffusion models. Through more than 300 controlled experiments, we investigated choices spanning model architecture, text and noise conditioning, training data composition and scale, captioning, and inference. These experiments reveal several simple but consequential design principles. Guided by these findings, we develop i1, a 3B-parameter text-to-image model trained entirely on publicly available datasets. i1 substantially improves over previous fully open models and is competitive with leading open-weight systems across a range of representative text-to-image benchmarks, demonstrating that strong performance does not necessarily require highly complex architectures or inaccessible training data. I will conclude by discussing future directions in building stronger and more efficient open models. Reading List: https://docs.google.com/document/d/1-CPI_66vLNMLr9FKIzEJZAy_Napf5iy-IbkoLW8… 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

Today's ORFE Department Colloquium at 4:30pm: John Duchi, Stanford University
by Emily C. Lawrence 15 Sep '26

15 Sep '26

Today's ORFE Department Colloquium: Please join us today for a light reception and pre-talk meet and greet at 4:00pm in Sherrerd Hall Atrium. Date & Time: Tuesday, September 15, 2026, 4:30 PM – 5:30 PM EDT Location: 101 - Sherrerd Hall Speaker: [ https://orfe.princeton.edu/events/2026/john-duchi-stanford-university | John Duchi, Stanford University ] Title: Finding Stationary Points in Stochastic Convex Optimization: Why and How Abstract: In this talk, I will discuss finding stationary points of possibly non-differentiable stochastic convex functions. Motivation for solving such problems arises in modern “distribution free” statistical learning problems. We will highlight some of the challenges of solving such problems, including that subgradients of stochastic convex functions do not converge uniformly, and also demonstrate some new geometric tools to show that solving the problem is in fact possible. Based on the paper “ [ https://arxiv.org/abs/2607.06883 | Finding a stationary point of a stochastic convex problem ] ”, a joint work with Felipe Areces and Malo Sommers. Bio: [ https://web.stanford.edu/~jduchi/ | John Duchi ] is an associate professor of Statistics and Electrical Engineering at Stanford University. His work spans statistical learning, optimization, information theory, and computation, with a few driving goals. (1) To discover statistical learning procedures that optimally trade between real-world resources — computation, communication, privacy provided to study participants — while maintaining statistical efficiency. (2) To build efficient large-scale optimization methods that address the spectrum of optimization, machine learning, and data analysis problems we face, allowing us to move beyond bespoke solutions to methods that robustly work. (3) To develop tools to assess and guarantee the validity of — and confidence we should have in — machine-learned systems.

1 0

Zeyu Ma will present his FPO "Procedural Generation of Unbounded 3D Scenes" on Friday, September 18, 2026 at 1:30 PM in CS 402.
by CS Grad Department 14 Sep '26

14 Sep '26

Zeyu Ma will present his FPO "Procedural Generation of Unbounded 3D Scenes" on Friday, September 18, 2026 at 1:30 PM in CS 402. The members of Zeyu’s committee are as follows: Examiners: Jia Deng (Adviser), Felix Heide, Adam Finkelstein Readers: Zhuang Liu, Qixing Huang (UT Austin) 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: Synthetic data, meaning images and labels generated by computers, is a promising alternative to collecting datasets in the real world and annotating them manually. Procedural generation, which creates content using algorithms, is an effective way to produce it: compact rules express complex, unbounded geometry; parameters give strong controllability; and unlimited varied instances can be generated automatically. This thesis addresses the challenges that unbounded scenes pose for procedural generation. Such scenes are an essential part of diverse environments, yet existing methods cannot handle them efficiently. Producing images from a procedural scene means extracting a mesh from it, and no uniformly spaced mesh serves an unbounded scene: a spacing fine enough for the geometry near the camera is prohibitively dense in the distance, and one coarse enough for the distance is far too sparse up close. Our contributions are in two parts. The first is an efficient procedural system for unbounded terrain, developed as part of Infinigen and published at CVPR 2023. It represents terrain with pseudo-signed distance functions, which keep only the sign of the true distance and thereby unify height maps, volumetric elements such as caves and rocks, and simulated tiles into one composable representation. The second is mesh extraction for a moving camera, a line of work that led to two papers. OcMesher, published at 3DV 2025,builds a camera-aware octree that is refined only where surfaces are visible and close to some camera, and produces a single static mesh valid for a whole set of camera poses. BinocMesher, published at SIGGRAPH Asia 2025, partitions4D spacetime with a binary-octree and slices the extracted 4D mesh at each timestamp, giving temporally smooth meshes along long camera trajectories. In all of them, detail is allocated against where the camera can see, and the allocation is a property of the coordinate system or data structure.

2 1

Next Week's ORFE Department Colloquium: Tuesday, September 15, 2026 at 4:30 PM: John Duchi, Stanford University
by Emily C. Lawrence 11 Sep '26

11 Sep '26

Next Week's ORFE Department Colloquium: Date & Time: Tuesday, September 15, 2026, 4:30 PM – 5:30 PM EDT Location: 101 - Sherrerd Hall Speaker: [ https://orfe.princeton.edu/events/2026/john-duchi-stanford-university | John Duchi, Stanford University ] Title: Finding Stationary Points in Stochastic Convex Optimization: Why and How Abstract: In this talk, I will discuss finding stationary points of possibly non-differentiable stochastic convex functions. Motivation for solving such problems arises in modern “distribution free” statistical learning problems. We will highlight some of the challenges of solving such problems, including that subgradients of stochastic convex functions do not converge uniformly, and also demonstrate some new geometric tools to show that solving the problem is in fact possible. Based on the paper “ [ https://arxiv.org/abs/2607.06883 | Finding a stationary point of a stochastic convex problem ] ”, a joint work with Felipe Areces and Malo Sommers. Bio: [ https://web.stanford.edu/~jduchi/ | John Duchi ] is an associate professor of Statistics and Electrical Engineering at Stanford University. His work spans statistical learning, optimization, information theory, and computation, with a few driving goals. (1) To discover statistical learning procedures that optimally trade between real-world resources — computation, communication, privacy provided to study participants — while maintaining statistical efficiency. (2) To build efficient large-scale optimization methods that address the spectrum of optimization, machine learning, and data analysis problems we face, allowing us to move beyond bespoke solutions to methods that robustly work. (3) To develop tools to assess and guarantee the validity of — and confidence we should have in — machine-learned systems.

1 0