Lucas Rosenblatt

Assistant Professor of Computer Science, Williams College.

lucas_photo.jpeg
Download CV
lr21 [at] williams [dot] edu

I am an Assistant Professor of Computer Science at Williams College, and my work focuses on responsible AI/ML. My research develops methods for generating high-quality, privacy-preserving synthetic data, designs algorithms that reduce bias and promote fairness in machine learning, and improves our understanding of large language models in order to identify and mitigate systemic risks to users.

I completed my PhD in Computer Science at NYU, supported by an NSF Graduate Research Fellowship. More about me and my research →

Interested in working with me? I am not taking on new student researchers this fall, but I will be taking students on in the spring. If you are an interested Williams student, a great first step is to take my course, CSCI 382: Responsible AI/ML.

News

Nov 2, 2026 New pre-print that I’m quite excited about: Differentially Private Synthetic Data from Sliced Chebyshev Moments with Teal Witter and Chris Musco. Many practical DP synthetic data methods only preserve low-dimensional marginals, and so can only capture global structure by (essentially) luck. Our algorithm, SPLICE, gets around this by matching noisy sliced Chebyshev moments. We actually prove a Sliced Wasserstein-1 guarantee that is independent of the data dimension, and we show that SPLICE is state-of-the-art on some measures in an extensive benchmark. This paper ties together a few different threads from the past few years, please reach out with questions / comments!
Nov 2, 2026 Explanation Multiplicity in SHAP: Characterization and Assessment was accepted to the NeurIPS 2026 Datasets & Benchmarks track. Much of the credit here goes to Daniel (Hyunseung) Hwang, who led the project, alongside Seungeun Lee. This work shows that SHAP explanations can differ substantially across repeated runs even with the model, input, and prediction held fixed. Thus, a single-run SHAP output is best interpreted as a draw from a distribution over possible outputs, not as an some authoritative artifact.
May 4, 2026 Two papers accepted at ICML 2026:
May 2, 2026 I defended my PhD! Huge thanks to my committee: Julia Stoyanovich, Christopher Musco, Rachel Cummings, Bill Howe. And to everyone who came out to support :)

Recent Publications

2026

  1. Privately Fine-Tuned LLMs Preserve Temporal Dynamics in Tabular Data
    Lucas Rosenblatt, Peihan Liu, Ryan McKenna, Natalia Ponomareva
    International Conference on Machine Learning (ICML) 2026
  2. Exactly Computing do-Shapley Values
    R. Teal Witter, Álvaro Parafita, Tomas Garriga, Maximilian Muschalik, Fabian Fumagalli, Axel Brando, Lucas Rosenblatt
    International Conference on Machine Learning (ICML) 2026
  3. Explanation Multiplicity in SHAP: Characterization and Assessment
    Hyunseung Hwang, Seungeun Lee, Lucas Rosenblatt, Steven Euijong Whang, Julia Stoyanovich
    NeurIPS 2026 (Datasets & Benchmarks Track)
  4. Epistemic Parity: Reproducibility as an Evaluation Metric for Differential Privacy
    Lucas Rosenblatt, Bernease Herman, Anastasia Holovenko, Wonkwon Lee, Joshua R. Loftus, Elizabeth McKinnie, Taras Rumezhak, Andrii Stadnik, Bill Howe, Julia Stoyanovich
    Communications of the ACM, 69(7), 2026 · CACM Research Highlight
  5. How to DP-Fy Your Data: A Practical Guide to Generating Synthetic Data with Differential Privacy
    Natalia Ponomareva, Zheng Xu, H. Brendan McMahan, Peter Kairouz, Lucas Rosenblatt, Vincent Cohen-Addad, Cristóbal Guzmán, Ryan McKenna, Galen Andrew, Alex Bie, Da Yu, Alex Kurakin, Morteza Zadimoghaddam, Sergei Vassilvitskii, Andreas Terzis
    Journal of Artificial Intelligence Research (JAIR), 86, 2026
  6. Differentially Private Synthetic Data from Sliced Chebyshev Moments
    Lucas Rosenblatt, R. Teal Witter, Christopher Musco
    arXiv preprint, 2026
  7. ContinuousBench: Can Differentially Private Synthetic Text Improve Capabilities?
    Peihan Liu, Lucas Rosenblatt, Weiwei Kong, Natalia Ponomareva, Gautam Kamath, Rachel Cummings, Roxana Geambasu, Yu Gan, Lillian Tsai, Alex Bie
    arXiv preprint, 2026
  8. Towards Provably Unbiased LLM Judges via Bias-Bounded Evaluation
    Benjamin Feuer, Lucas Rosenblatt, Oussama Elachqar
    arXiv preprint, 2026
  9. Trustworthy Differentially Private Data Generation and Predictive Modeling
    Lucas Rosenblatt
    PhD dissertation, New York University, 2026

2025

  1. Do You Really Need Public Data? Surrogate Public Data for Differential Privacy on Tabular Data
    Shlomi Hod, Lucas Rosenblatt, Julia Stoyanovich
    NeurIPS 2025 (Datasets & Benchmarks Track)
  2. Sharper Bounds for Chebyshev Moment Matching, with Applications to Differential Privacy and Beyond
    Cameron Musco, Christopher Musco, Lucas Rosenblatt, Apoorv Vikram Singh
    Conference on Learning Theory (COLT) 2025
  3. Fragments to Facts: Partial-Information Fragment Inference from LLMs
    Lucas Rosenblatt, Bin Han, Robert Wolfe, Bill Howe
    International Conference on Machine Learning (ICML) 2025
  4. Differential Privacy Under Class Imbalance: Methods and Empirical Insights
    Lucas Rosenblatt, Yuliia Lut, Eitan Turok, Marco Avella-Medina, Rachel Cummings
    International Conference on Machine Learning (ICML) 2025
  5. SimKey: A Semantically Aware Key Module for Watermarking Language Models
    Shingo Kodama, Haya Diwan, Lucas Rosenblatt, R. Teal Witter, Niv Cohen
    arXiv preprint, 2025

2024

  1. A Simple and Practical Method for Reducing the Disparate Impact of Differential Privacy
    Lucas Rosenblatt, Julia Stoyanovich, Christopher Musco
    AAAI 2024
  2. I Open at the Close: A Deep Reinforcement Learning Evaluation of Open Streets Initiatives
    R. Teal Witter, Lucas Rosenblatt
    AAAI 2024
  3. Laboratory-Scale AI: Open-Weight Models are Competitive with ChatGPT Even in Low-Resource Settings
    Robert Wolfe, Isaac Slaughter, Bin Han, Bingbing Wen, Yiwei Yang, Lucas Rosenblatt, Bernease Herman, Eva Maxfield Brown, Zening Qu, Nic Weber, Bill Howe
    ACM Conference on Fairness, Accountability, and Transparency (FAccT) 2024
  4. Epistemic Parity: Reproducibility as an Evaluation Metric for Differential Privacy
    Lucas Rosenblatt, Bernease Herman, Anastasia Holovenko, Wonkwon Lee, Joshua R. Loftus, Elizabeth McKinnie, Taras Rumezhak, Andrii Stadnik, Bill Howe, Julia Stoyanovich
    SIGMOD Record, 53(1), 2024 · SIGMOD Research Highlight
  5. FairlyUncertain: A Comprehensive Benchmark of Uncertainty in Algorithmic Fairness
    Lucas Rosenblatt, R. Teal Witter
    arXiv preprint, 2024
  6. Are Data Experts Buying into Differentially Private Synthetic Data? Gathering Community Perspectives
    Lucas Rosenblatt, Bill Howe, Julia Stoyanovich
    arXiv preprint, 2024

Publications before 2024 are listed on my Google Scholar profile.

Teaching

Courses I’m teaching/have taught at Williams:

Courses I taught or TA’d in grad school: