- talks - lists.cs.princeton.edu

Rohit Agarwal will present his General Exam "AI Alignment in Agentic Pipelines via Incentives and Correction" on Friday, May 15, 2026 at 2:00 PM in CS 301 and via zoom.
by CS Grad Department 12 May '26

12 May '26

Rohit Agarwal will present his General Exam "AI Alignment in Agentic Pipelines via Incentives and Correction" on Friday, May 15, 2026 at 2:00 PM in CS 301 and via zoom. Zoom link: https://princeton.zoom.us/j/7714894406 Committee Members: Elad Hazan (advisor), Sanjeev Arora, Peter Henderson Abstract: We propose a mechanism-design perspective on AI alignment in agentic pipelines. The key observation is that alignment in a multi-agent setting is a fixed-point problem: penalizing misbehavior may deter a solver, but it can also reduce an auditor's incentive to monitor, since auditing then mainly incurs cost on a population that already appears aligned. Thus, robust alignment requires not only discouraging bad behavior, but also preserving incentives for oversight. We also discuss the applications of this viewpoint in real-life Large Language Model settings, in coding through an empirical reinforcement learning setup, and in mathematics through empirical agentic pipeline design. Reading List: https://docs.google.com/document/d/1Ji1xpqK-0S5saeVjwodH07aXmU_iQo0XvYYVTcV… 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

Kara Schechtman will present her General Exam "Causal Evaluation of Algorithmic Decision Support’s Organizational Role" on Thursday, May 14, 2026 at 10:30 AM in Sherrerd 306.
by CS Grad Department 11 May '26

11 May '26

Kara Schechtman will present her General Exam "Causal Evaluation of Algorithmic Decision Support’s Organizational Role" on Thursday, May 14, 2026 at 10:30 AM in Sherrerd 306. Committee Members: Lydia Liu (advisor), Aleksandra Korolova, Arvind Narayanan Abstract: In high stakes organizational decision-making settings such as education, healthcare, and hiring, algorithms are increasingly used to support human decision-makers or even automate certain decisions entirely. Building a full scientific understanding of algorithmic decision support’s impacts requires grappling with the complexity of the organizations in which they operate, and the organizational goals we aim to achieve. In this talk, I present on recent and ongoing work which aim to contribute to this growing understanding. First, I present recent work on the role of human expertise in algorithmic intervention targeting. Algorithmic decision support (ADS) is widely used to help human-decision makers target interventions, yet little is known about how human decision makers integrate algorithmic suggestions with contextual knowledge unavailable to algorithmic tools in settings where their actions influence outcomes. We develop a causal framework to test whether human decision-makers use contextual information beyond algorithmic inputs in order to improve outcomes. Applying the framework to rich longitudinal data from an algorithm-assisted college advising program, we estimate advisors relied on “non-algorithmic context” that 2 out of 3 recorded interventions plausibly relied on non-algorithmic context, corroborating this finding by systematic qualitative analysis of advisor meeting notes. These results suggest that human discretion remains a dominant driver of educational interventions even in algorithmic-assisted settings, and highlight the importance of scaling up human capacity to contextualize and act on algorithmic recommendations when expanding ADS. Second, I outline the methodological commitments that inform my work — using a causal approach to evaluate organizational ADS, while drawing on social theory and qualitative research on street-level bureaucrats’ use of ADS to inform problem selection for empirical research. I explain how these methodologies coalesce in prior and ongoing work on the role of ADS in college advising, the role of correspondence studies in language model discrimination auditing, and the role of data collection in medical and public health decision making, as well as my goals for future work in these areas. Reading List: https://docs.google.com/document/d/1ZnR1Mmiep0u_pQ22eAMykzH4Evpep5tzmR5t-rO… 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

Sijia Liu will present her General Exam "Towards Autonomous Language Agents for Long Horizon Tasks" on Tuesday, May 12, 2026 at 3:00 PM in CS 401 and via zoom.
by CS Grad Department 07 May '26

07 May '26

Sijia Liu will present her General Exam "Towards Autonomous Language Agents for Long Horizon Tasks" on Tuesday, May 12, 2026 at 3:00 PM in CS 401 and via zoom. Zoom link: https://princeton.zoom.us/j/3816009387?pwd=TfbXJiyYLsJpDhsnkmVa6waA5VbLQs.1 Committee Members: Karthik Narasimhan (advisor), Danqi Chen, Zhuang Liu Abstract: LLM agents have made rapid progress on complex reasoning and software engineering tasks, yet they still struggle to make sustained, autonomous progress over long horizons. Recent work has begun using state-of-the-art coding agents (e.g., Claude Code) to automate AI research, with promising signals of agents iteratively hill-climbing toward superhuman performance. However, these efforts remain confined to well-specified subproblems and rely on substantial human-in-the-loop supervision to gauge progress. To better understand and address this gap, we study MLE-Bench, a widely used benchmark that evaluates LLM agents on Kaggle-style machine learning engineering competitions over ultra-long horizons (24 hours) under realistic resource constraints (data, memory, GPUs, and network access). Preliminary analysis shows that even frontier proprietary models fall short at proposing novel ideas to iteratively improve their solutions, and often get stuck in debugging loops or commit to suboptimal model choices. Scaffold design remains crucial even for the strongest proprietary models. Open-weights models lag substantially behind, and naive trajectory distillation from strong teachers fails to close the gap. Together, these results illuminate the core challenges on the path to fully autonomous language agents for long-horizon tasks and open-ended discovery. Reading List: https://docs.google.com/document/d/1Uby4uBlan6YTQXwx1wTXJvcaB74q7iWU_e-BZQy… 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

Zeyu Shen will present his General Exam "Towards Reliable and Aligned AI Systems" on Tuesday, May 12, 2026 at 2:00 PM in CS 301.
by CS Grad Department 07 May '26

07 May '26

Zeyu Shen will present his General Exam "Towards Reliable and Aligned AI Systems" on Tuesday, May 12, 2026 at 2:00 PM in CS 301. Committee Members: Aleksandra Korolova (advisor), Peter Henderson, Prateek Mittal Abstract: In this qualification exam, I examine how to build AI systems that remain reliable, safe, and aligned under adversarial pressure. It connects foundational work on adversarial robustness with recent evidence that fine-tuning and reward optimization can induce misalignment in language models, and studies retrieval-augmented generation as a systems setting where safety depends not only on the model itself but also on the trustworthiness of retrieved information. My main presentation will focus on ReliabilityRAG, which proposes a reliability-aware, graph-theoretic defense that uses contradiction detection, maximum independent set selection, and weighted sampling to provide both provable robustness and strong empirical protection against retrieval corruption and prompt injection. Reading List: https://docs.google.com/document/d/1cfk_AQjsqQruH3n_AO_-bVOg92UnISnlowJd0jZ… 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

Patty Liu will present her General Exam "Using AI to Augment Access to Justice" on Wednesday, 5/13/26 at 1:15pm in Sherrerd Hall 306.
by Gradinfo 07 May '26

07 May '26

Patty Liu will present her General Exam "Using AI to Augment Access to Justice" on Wednesday, 5/13/26 at 1:15pm in Sherrerd Hall 306. Committee: Peter Henderson (adviser), Lydia Liu, and Manoel Horta Ribeiro Abstract: The rapid adoption of AI in legal contexts presents both opportunities and risks. For example, attorneys, judges, and pro se litigants are increasingly turning to large language models to draft briefs and conduct legal research, yet these systems hallucinate legal citations at non-negligible rates and are rarely evaluated against the realities of legal practice. My research aims to audit AI systems in high-stakes legal settings and to understand how AI can responsibly expand access to justice. To audit AI in law, we study how AI tools can help verify legal citations to alleviate the growing problem of legal hallucinations in courts. We first show that hallucination rates are not consistently declining across model generations. We also introduce a benchmark of legal brief excerpts with injected hallucinations grounded in a taxonomy of citation errors derived from real court filings and evaluate an agentic verification system on this benchmark. We find that while agentic retrieval substantially improves recall, structural barriers, particularly paywalled legal databases, limit what these verification systems can achieve. To understand how AI can expand access to justice, we conduct a qualitative study with public defenders, a population underserved by existing legal AI research. Through in-depth interviews, we develop a task-level map of public defense work and surface the ethical constraints, confidentiality risks, and organizational barriers that shape responsible AI adoption in this context. The combination of these works show that progress in legal AI requires both empirical auditing of AI systems and direct engagement with the practitioners who use them. Reading List: [ https://docs.google.com/document/d/1uipmhQd2fgYn14u9BfNmqh-s3BBjoSollOW9X4e… | https://docs.google.com/document/d/1uipmhQd2fgYn14u9BfNmqh-s3BBjoSollOW9X4e… ] 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

Jack Nugent will present his General Exam "Benchmarking the Robustness of Monocular Depth Estimation with Procedural Scene Perturbations" on Friday, May 8, 2026 at 2:00 PM in Friend 202.
by CS Grad Department 06 May '26

06 May '26

Jack Nugent will present his General Exam "Benchmarking the Robustness of Monocular Depth Estimation with Procedural Scene Perturbations" on Friday, May 8, 2026 at 2:00 PM in Friend 202. Committee Members: Jia Deng (advisor), Felix Heide, Szymon Rusinkiewicz Abstract: Monocular depth estimation, predicting depth from a single image, is a fundamental computer vision task that has seen rapid progress in recent years. Though the task is inherently ambiguous, models are increasingly capable of producing accurate metric estimates, sharp predictions, and precise local geometry. As these models are deployed, it becomes increasingly important to rigorously measure their accuracy and their robustness to conditions that may appear in the real world. To systematically analyze model performance, we introduce a new Procedural Depth Evaluation benchmark. We draw on recent advances in photorealistic procedural scene generation to programmatically generate 3D scenes. We then apply controlled perturbations that vary targeted scene factors while preserving the relevant depth-estimation challenge. For example, we alter light intensity, color, and direction while ensuring the entire scene remains visible. By measuring the error on these perturbed scenes, we quantify robustness to a particular perturbation. Our benchmark tests robustness to various object, camera, material, and lighting changes. We find that no model is robust to all these perturbations, and that the most accurate models are not the most robust. This benchmark provides a way to diagnose failure modes in depth estimation models that are not captured by standard evaluations. Reading List: https://docs.google.com/document/d/1LQd3WEYciYDrX_uSbtOBH0IsUuQPGlsOqkPw7Er… 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

Ching-Yi Tsai will present his General Exam "Shaping Interaction with Situated AI through Representations of Uncertainty and Privacy" on Thursday, May 7, 2026 at 1:30 PM in CS 401 and via zoom.
by CS Grad Department 05 May '26

05 May '26

Ching-Yi Tsai will present his General Exam "Shaping Interaction with Situated AI through Representations of Uncertainty and Privacy" on Thursday, May 7, 2026 at 1:30 PM in CS 401 and via zoom. Zoom link: https://princeton.zoom.us/j/98708013821?pwd=p5QPT9IYvKsQUwvcEq8fJPXCsJ2xab.1 Committee Members: Parastoo Abtahi (advisor), Andrés Monroy-Hernández, Janet Vertesi Abstract: Situated AI systems, including AR glasses, robots, and other embodied assistants, are increasingly entering everyday life. Yet interactions with these systems are still shaped by conventional designs that treat interpretations, controls, and consequences as static and deterministic, making alternative interpretations and actions difficult to perceive or refine in the dynamic conditions of everyday life. A single-target pointer, for example, can miss a user’s intended referent under noisy everyday conditions, while a simple camera on/off switch is too coarse to selectively regulate privacy-sensitive video content without sacrificing application functionality. My research investigates how interaction with situated AI can shift from static, implicit handling to in-situ, dynamic representations at the right granularity for both awareness and action. I study situated representations that bring ambiguity and privacy risks into users’ awareness while also allowing users to perceive, change, and resolve them in the moment. I will present two projects that explore scenarios for doing so: Uncertain Pointer, which investigates non-deterministic, situated feedforward visualizations for communicating and disambiguating ambiguity in users’ input, such as speech and gesture; and PrivacyLoupe, which explores a dynamic representation of privacy control that balances policy expressiveness and context sensitivity with interaction cost. I will also discuss future directions in in-situ dynamic abstraction and other interaction demands that could benefit from in-situ representation. Together, these projects examine how interaction with situated AI can be shaped by moving from static and implicit handling to situated representations that support user agency, awareness, and action. Reading List: https://docs.google.com/document/d/1chCb6CSHrHFQeaQiQUzavIZmrn6HczRD_3_bkQ2… 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

Henri Schmidt will present his FPO "Inferring cellular phylogenies using optimization” on Thursday, May 7, 2026 at 10:00 AM in CS 402.
by CS Grad Department 05 May '26

05 May '26

Henri Schmidt will present his FPO "Inferring cellular phylogenies using optimization” on Thursday, May 7, 2026 at 10:00 AM in CS 402. The members of Henri’s committee are as follows: Examiners: Ben Raphael (Adviser), Yuri Pritykin, Pravesh Kothari Readers: Olga Troyanskaya, Mohammed El-Kebir (UIUC) 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: Reconstructing the phylogenetic tree of cells is a central challenge in developmental and cancer biology. Advances in sequencing and experimental design have brought this goal within reach, producing high-resolution single-cell measurements that enable phylogenetic reconstruction. However, phylogenetic reconstruction from such measurements is difficult and requires phylogenetic models that balance biological plausibility and computational tractability. We present a collection of research whose goal is, broadly, to design tractable phylogenetic models along with efficient algorithms to reconstruct phylogenetic trees. In particular, we adapt modern tools and techniques from combinatorial and convex optimization to the phylogenetic setting, leading to highly efficient and accurate algorithms for large-scale phylogenetic inference.

2 1

Anand Brahmbhatt will present his General Exam "Spectral and Improper Methods for Learning and Control of Dynamical Systems" on Friday, May 8, 2026 at 9:30 AM in CS 302.
by CS Grad Department 05 May '26

05 May '26

Anand Brahmbhatt will present his General Exam "Spectral and Improper Methods for Learning and Control of Dynamical Systems" on Friday, May 8, 2026 at 9:30 AM in CS 302. Committee Members: Elad Hazan (advisor), Mark Braverman, Sanjeev Arora Abstract: Classical control methods such as LQR provide powerful tools for known linear dynamical systems, but typically assume fixed quadratic costs and stochastic noise. In contrast, prior work in online control uses improper learning to handle changing convex cost functions and bounded oblivious adversarial disturbances, achieving optimal regret guarantees. In this talk, we revisit this line of work and show that, by imposing a suitable structural restriction on the policy class, one can design a spectral control algorithm that retains optimal regret while significantly improving computational efficiency, with per-iteration runtime scaling only polylogarithmically in the stability margin. We then describe extensions of this approach to partially observed systems and to bandit feedback, where only scalar cost information is available. Finally, we discuss how similar improper learning ideas extend to learning in nonlinear dynamical systems, yielding prediction guarantees via high-dimensional linear surrogates. Reading List: https://docs.google.com/document/d/1OLqfFM92wM6I1dQ5gbRBkZxCfgR_hOA8FS7mxuH… 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

Kincaid MacDonald will present his General Exam "Epistemic Empowerment and Belief Curvature in LLM Agents" on May 14, 2026 at 1pm in 306 Sherrerd Hall and Zoom.
by Gradinfo 05 May '26

05 May '26

Kincaid MacDonald will present his General Exam "Epistemic Empowerment and Belief Curvature in LLM Agents" on May 14, 2026 at 1pm in 306 Sherrerd Hall and Zoom. Zoom: [ https://princeton.zoom.us/j/4739233869 | https://princeton.zoom.us/j/4739233869 ] Committee: Peter Henderson (advisor), Aleksandra Korolova, Tom Griffiths Abstract We study the agentic behavior of LLMs through the geometry of their beliefs, viewing reasoning trajectories as paths traced across a manifold of belief representations. This opens a quantitative window on fickle matters of agentic temperament, like a model's tendency to fixate on one hypothesis, or its capacity to take actions that challenge its beliefs. Both of these are captured by epistemic empowerment, a measure of empowerment in belief space. We prove that the epistemic empowerment is upper-bounded by the local curvature of the model's beliefs, and devise a curvature estimator which is more accurate than existing methods in high dimensions. Using this, we can measure LLM's epistemic empowerment and increase it through RL fine-tuning, resulting in increased performance on tests of exploration capacity and open-mindedness. We also posit that an agent high in epistemic empowerment makes a better collaborator than one high in traditional empowerment. Reading List: [ https://docs.google.com/document/d/1-W5ZGTqXyhEzTKbGaO2FJVc69Tgl4c95nCXSHcA… | https://docs.google.com/document/d/1-W5ZGTqXyhEzTKbGaO2FJVc69Tgl4c95nCXSHcA… ]

1 0