Robin Linzmayer

Classes start this week! I'm TAing COMS 4705: Natural Language Processing this fall. Office hours aren't set yet, check back soon.

About

I study how to evaluate increasingly capable AI systems in high-stakes settings.

I'm a PhD student in Computer Science at Columbia University, advised by Noémie Elhadad, part of the Columbia Core AI Lab (CAIL), and supported by the NSF CISE Graduate Fellowship. My work combines safety-critical evaluation, reasoning interpretability, and health data modeling to build foundations for more reliable and aligned AI systems.

Previously, I spent four years as a Machine Learning Scientist at Flatiron Health, where I built machine learning systems from electronic health records to unlock large-scale oncology research.

Papers

Untangling the Mechanisms of Misleading Context in Medical Question Answering
R Linzmayer, N Elhadad
Under review @ ML4H 2026
Examines how misleading context in medical question answering corrupts clinical reasoning. Assesses models' susceptibility to the misleading information, whether they disclose its influence, the mechanism of corrupted reasoning, and the monitorability of the final decision.
2026
AcuityBench: Evaluating Clinical Acuity Identification and Uncertainty Alignment
R Linzmayer, G Lin, D Coneybeare, J Chu, T Cloyd, M Garg, …, N Elhadad
Under review @ NeurIPS 2026
Introduces a multi-source, physician-annotated benchmark for evaluating whether models identify the right level of care across QA and conversational formats while exposing safety-critical error patterns and uncertainty behavior.
2026
Aggregate benchmark scores obscure patient safety implications of errors across frontier language models
R Linzmayer, A Ramaswamy, H Hugo, G Nadkarni, N Elhadad
medRxiv preprint
Shows that aggregate triage scores can obscure clinically important model behavior, including directional error patterns and a consistent shift toward lower-acuity recommendations when a third party minimizes the patient’s symptoms.
2026
A foundation model for capturing complexity of menstrual health data
R Linzmayer, C Pang, I Urteaga, G Gürsoy, A Shea, VJ Vitzthum, et al.
npj Women's Health
Introduces a foundation model for longitudinal menstrual health data that captures temporal and symptomatic patterns, improves downstream forecasting, and enables privacy-sensitive data sharing.
2026
Artificial Intelligence in the Prediction of Ovarian Torsion
S Guang, R Linzmayer, J Baker, S Seaman, AP Advincula, N Elhadad
Journal of Minimally Invasive Gynecology 32 (11), S121
Uses LLM-based extraction from unstructured EHR notes and interpretable machine learning to predict ovarian torsion, showing that locally trained models outperform zero-shot LLM classification for this clinical task.
2025
Approach to machine learning for extraction of real-world data variables from electronic health records
B Adamson, M Waskom, A Blarre, J Kelly, K Krismer, S Nemeth, …, R Linzmayer, …, A Cohen
Frontiers in Pharmacology 14, 1180962
Describes how machine learning pipelines can transform unstructured oncology records into research-ready real-world data, reducing reliance on manual abstraction while preserving clinical utility.
2023
Metastatic patterns and outcomes by HER2 and hormone receptor (HR) status in patients with metastatic breast cancer
Q Yuan, E Castellanos, E Fidyk, K Schwed, M Estevez, S Nemeth, …, R Linzmayer, A Cohen
Journal of Clinical Oncology 41 (16_suppl), 1031–1031
Uses machine-learning-extracted EHR variables to study metastatic patterns and survival by HER2 and hormone receptor status in a real-world metastatic breast cancer cohort.
2023
Implementation of an EHR-embedded decision support tool in community oncology practices
R Maniago, SS Richey, S DeVincenzo, S Jou, R Linzmayer, J Donegan, et al.
Journal of Clinical Oncology 39 (28_suppl), 274–274
Evaluates the first year of an EHR-embedded clinical decision support tool across community oncology practices, measuring real-world use and guideline-concordant treatment selection.
2021
Capturing student feedback and emotions in large computing courses: A sentiment analysis approach
M Neumann, R Linzmayer
Proceedings of the 52nd ACM Technical Symposium on Computer Science Education
Applies sentiment analysis to anonymous course feedback, showing that NLP can help identify student struggle and emotional trends in large computing courses.
2021