Ellen Vitercik

Congrats to my awesome postdoc @xizhi_tan, as well as @GolowichNoah and Yifan!

Looking forward to visiting Penn tomorrow!

@DimitrisPapail @boazbaraktcs Yes, I have a no-laptops/devices policy for my PhD-level discussion classes and highly recommend it! I think it's demoralizing for a student to give a seminar presentation to an audience of students distracted on their laptops.

@boazbaraktcs ++ no laptops/phones or devices allowed

Tempted to announce that my AI safety course will: 1. Have mandatory attendance. 2. Projects will be expected to be research paper quality. 3. Won't satisfy any departmental requirements 4. No one will get more than A- to get right kind of self selection. (3&4 would be new)

@nhaghtal @Berkeley_EECS Woohooo! 💃🥳🪇

This week I was promoted to the rank of Associate Professor at @Berkeley_EECS ! In a remarkable show of enthusiasm, the committee apparently tore a hole in spacetime to make me an Associate Professor 9 months ago!

Proud to share that my awesome PhD student Yingxi Li has received an Amazon AI PhD Fellowship. Congratulations, Yingxi, and thank you, @AmazonScience @amazon!

@_vaishnavh I have a spreadsheet I maintain where I use exactly this idea to estimate my "load" on the reviewing system! Then I make sure the number of papers I review each year is at least my load.

To clarify, my proposal is a system where, if a group of authors submit K papers, they should guarantee N*K reviews (where N is say, 3 or 4).

Curious why conferences don't have a system where the authors of every paper together guarantee to provide N reviews (and they can distribute the load amongst themselves). This way wouldn't we tax authors in proportion to the number of papers they burden the system with?

Excited to share our position paper on neural algorithmic reasoning. NAR is a promising and fast-growing area, and in this paper we highlight several opportunities to sharpen its scope, deepen its theory, and strengthen its empirical evaluation.

Slides: vitercik.github.io/assets/pdf/jun… LLMs for Cold-Start Cutting Plane Separator Configuration (CPAIOR’25): arxiv.org/abs/2412.12038 Algorithms with Calibrated Machine Learning Predictions (ICML’25): arxiv.org/abs/2502.02861

Last week, I spoke at Juna AI about our recent research on machine learning for optimization, which aims to leverage historical structure to make large, discrete decision-making faster and more robust. Two themes: 1. LLMs for “cold-start” integer programming solver configuration: we generate cutting-plane separator settings after testing only a handful of candidate configs, often with large speedups vs solver defaults. (CPAIOR’25; w/ @LawlessOpt, Yingxi Li, Anders Wikum, @madeleineudell) 2. Decision-making under uncertainty: calibrated uncertainty estimates allow us to move beyond “trust-the-predictor” optimization, with applications to scheduling. (ICML’25 spotlight; w/ @judyhshen, Anders Wikum) Slides in the first reply.

arXiv: arxiv.org/abs/2510.20288 Talk: youtu.be/yEBPI9c7OE8?si…

New at ITCS this week: we obtain an O(1)-competitive algorithm for online metric matching in d-dimensional Euclidean metrics (d > 2) using only a single sample from each request distribution. Mingwei Yang is presenting our paper: Smoothed Analysis of Online Metric Matching with a Single Sample: Beyond Metric Distortion Yingxi Li, Ellen Vitercik, Mingwei Yang (arXiv and talk in first reply) Setting: n servers (e.g., rideshare drivers) are available in advance, and n requests (e.g., riders) arrive online. Each request must be matched immediately to an available server, paying the distance in an underlying metric. Contributions: • We move beyond i.i.d.: each request is drawn from its own distribution (under a mild smoothness condition). • No distributional knowledge: we use one sample per request distribution. • O(1) competitive ratio for Euclidean metrics when d > 2. Conceptual idea: We sidestep the usual Ω(log n) barrier from probabilistic metric embeddings by bounding our algorithm’s cost directly on a deterministic embedding, and relating it to OPT via majorization arguments.

@LawlessOpt and I are excited to present our #AAAI2026 tutorial on “LLMs for Optimization: Modeling, Solving, and Validating with Generative AI.” When: Tuesday, Jan 20, 2026, 8:30am–12:30pm (Singapore time) Where: Garnet 216 (Singapore EXPO) Optimization is central to planning, scheduling, and decision-making, but deploying solvers requires deep expertise. Our tutorial covers how LLMs can support the end-to-end optimization pipeline (model formulation, solver configuration, and model validation) and highlights open research directions. Tutorial page (agenda + reading list): conlaw.github.io/llm_opt_tutori… (Connor’s intro slides are shown here.) Thanks to @leo_bix and @madeleineudell for helping put the proposal together. CC @RealAAAI

Structural reasoning—order, hierarchy, connectivity—is a prerequisite for reliable multi-step decision-making. DSR-Bench (leads: @yuhe441 and Yingxi Li) probes this without tools via data-structure operations and surfaces clear failure modes in SOTA LLMs.

4) Theoretical Guarantees - Optimizing Solution-Samplers for Combinatorial Problems: The Landscape of Policy-Gradient Methods (@cmcaram et al., NeurIPS’23) - Approximation Algorithms for Combinatorial Optimization with Predictions (Antoniadis et al., ICLR’25)

3) Math Optimization - Using LLMs to Model and Solve Optimization Problems at Scale (@aliteshnizi et al., arXiv’25) - Contrastive Predict-and-Search for Mixed Integer Linear Programs (Huang et al., ICML’24) - Differentiable Integer Linear Programming (Geng et al., ICLR’25)

Excited to share the materials (slides + reading list) from my Stanford seminar “AI for Algorithmic Reasoning & Optimization.” We covered formal frameworks for LLM reasoning, GNNs for combinatorial optimization, ML for math optimization and theory guarantees. Link in first reply.