Geoffrey Martin

I'm a Ph.D. student at Cornell University and Weill Cornell Medicine, where I work with Yifan Peng in the Peng Lab on language models for clinical decision-making. My work is supported by an NSF Graduate Research Fellowship.

Geoffrey Martin

Research

My primary line of research studies how language models can support clinical decision-making by following clinical practice guidelines, whose decision logic is spread across the text and flowcharts of long documents. I build systems that extract this logic into decision trees and evaluate how accurately language models follow them for patient cases. Secondarily, I work on natural language processing methods for clinical text.

Language Models for Clinical Decision-Making

My work here broadly draws on the following sub-areas:

  1. Multi-Agent Systems Coordinating language model agents to extract the decision logic in a clinical practice guideline into a decision tree, and then to traverse that tree for a patient case (CPGPrompt).
  2. Benchmarking and Evaluation Measuring how accurately language models follow guideline decision trees, with a multimodal benchmark of patient vignettes and an evaluation harness that scores each step of the path as well as the final recommendation.
  3. Fine-Tuning Training small open models to follow the decision paths of new clinical guidelines more closely.
  4. Analysis Placing hospital patients on the branches of a guideline tree from their records, to measure how closely the care they received follows the guideline.

Example

Adapted from the ankle portion of the Ottawa Ankle Rules (Stiell et al., JAMA 1993) and simplified for illustration. All patient details are invented.

NLP Methods for Clinical Text

This line of research develops core natural language processing methods for clinical and health text. Some of my work studies how to run language models efficiently on long clinical documents and limited hospital hardware, while other work studies how to extract information from clinical narratives and assess suicide risk from social media posts.

News

Publications

Comparing LLM and Fine-Tuned Model Performance on NVDRS Circumstance Extraction with Varying Prompt Complexity We compare large language models with a fine-tuned RoBERTa model for extracting the circumstances of suicides from death investigation narratives, and test when detailed coding guidelines in the prompt help. G. Martin, X.Z. Feng, Y. Peng IEEE International Conference on Healthcare Informatics (ICHI), 2026 arXiv
Budget-Aware Routing for Long Clinical Text We study how to choose which parts of a long clinical document to give a language model when its input length is capped, and propose a simple rule for picking the best selection strategy for each task. K. Qureshi, G. Martin, Y. Peng Findings of the Association for Computational Linguistics (ACL), 2026
Capacity Planning for Resource-Constrained Clinical LLM Serving We study how a hospital with limited computing hardware should divide it between small and large language models, and whether that choice matters more than how requests are routed between them. Y. Tang, G. Martin, Y. Zhang, Y. Peng IEEE International Conference on Big Data (BigData), Special Session on Healthcare Data, 2026 arXiv
A Survey on MLLM-based Visually Rich Document Understanding: Methods, Challenges, and Emerging Trends We survey how multimodal large language models are used to understand documents that combine text with visual layout, such as forms and scanned reports. Y. Ding, S. Luo, Y. Dai, Y. Jiang, Z. Li, Q. Sun, G. Martin, W. Liu, Y. Peng Findings of the Association for Computational Linguistics (ACL), 2026
CPGPrompt: Translating Clinical Guidelines into Large Language Model-Executable Decision Support In this work, we convert clinical practice guidelines into decision trees and use a language model to walk through them for each patient case, producing recommendations that follow the guideline. R. Deng*, G. Martin*, T. Wang, G. Zhang, Y. Liu, C. Weng, Y. Wang, J. Rousseau, Y. Peng Journal of the American Medical Informatics Association (JAMIA), 33(4), 2026
Explainable Suicide Risk Assessment on Social Media with Multi-Task QLoRA We fine-tune a language model on several related tasks at once to assess suicide risk in social media posts and to explain each assessment with evidence from the post. X.Z. Feng, G. Martin, H. Dong, Y. Peng IEEE International Conference on Big Data (BigData), BigData Challenge Cup, 2026 arXiv
Reducing the Prevalence of Alcohol-Exposed Pregnancies in the United States: A Simulation Modeling Study We build a simulation model of US women of reproductive age to estimate how much different public health strategies could reduce alcohol-exposed pregnancies. R. Yaesoubi, M. Mahin, G. Martin, A. D. Paltiel, M. Sharifi Medical Decision Making, 42(2), 2022

* equal contribution

Talks and Posters

Service

Fellowships

National Science Foundation Graduate Research Fellowship 2022
Caltech Summer Undergraduate Research Fellowship 2018

Before Cornell

I grew up in Singapore, with earlier stops at NASA's Jet Propulsion Laboratory building vision models for robotics and at Locus Analytics, a systems economics think tank, working as a data scientist for a few years. Just before Cornell I studied operations research and statistics at Yale, after completing my undergraduate degree at Yale-NUS College.

Related Academic Interests

Systems thinking, causal inference, public policy, and election science.