Tom Hartvigsen
Assistant Professor
Data Science
University of Virginia
(Office: 1919 Ivy Rd., Rm. 339)
★ News ★
Oct'24: New preprints on isolating math skills in LLM weights, attacking VLMs with model edits, LLMs resisting misinformation requests, and high-dimensional action spaces in Offline RL. Paper accepted to IEEE BigData on spike train classification.
Sep'24: Three papers accepted to NeurIPS'24!
Are Language Models Actually Useful for Time Series Forecasting? (Spotlight!) - Congrats Mingtian!
Test-Time Debiasing of Vision-Language Embeddings
UniTS: A Unified Multi-Task Time Series Model
Sep'24: Three papers accepted to EMNLP'24!
MATHWELL: Generating Educational Math Word Problems with Teacher Annotations - Congrats, Bryan!
Language Models Still Struggle to Zero-shot Reason about Time Series
Language Models are Surprisingly Fragile to Drug Names in Biomedical Benchmarks
Aug'23: Paper accepted to TMLR on using LLMs for robust text classification
July'24: Paper accepted to COLM'24 on multilingual toxicity in LLMs. Paper accepted to AIES'24 on detecting implicit social biases in VL models
July'24: New preprints on composable interventions for LLMs and extracting social determinants of health with LLMs
June'24: Paper accepted to MICCAI'24 on federated learning for medical imaging
May'24: Paper accepted to ACL'24 on categorical knowledge editing for LLMs
Apr'24: Nature Medicine paper on bias in computational pathology
Spring'24: Invited talks at Dartmouth, IBM Research, UCSF/UC Berkeley, and the University of Alabama, Birmingham
Hi! I'm a tenure-track Assistant Professor of Data Science and, by courtesy, Computer Science at the University of Virginia. I joined UVA in Fall 2023. Before that, I was a postdoc at MIT CSAIL working with Marzyeh Ghassemi. I did my PhD in Data Science at WPI where I was advised by Elke Rundensteiner and Xiangnan Kong.
I am recruiting 1-2 creative and driven PhD students to start Fall 2025 (see the Research Group page for details)
Research
My research group works on machine learning and natural language processing. We work to enable responsible model deployment in ever-changing environments, with applications to biomedical data science.
Active directions and highlights:
Continually monitoring and editing knowledge and behavior of big models
Aging with GRACE: Lifelong Model Editing with Discrete Key-Value Adapters (NeurIPS'23 + code + blog post)
TAXI: Evaluating Categorical Knowledge Editing for Language Models (ACL'24 + data)
BendVLM: Test-Time Debiasing of Vision-Language Embeddings (NeurIPS'24)
Composable Interventions for Language Models (preprint'24)
Time series and multi-modality
Are Language Models Actually Useful for Time Series Forecasting? (NeurIPS'24 + code)
UniTS: A Unified Multi-Task Time Series Model (NeurIPS'24 + code)
Language Models Still Struggle to Reason about Time Series (EMNLP'24 + code)
Detecting and mitigating harmful biases in language and language models
PolygloToxicityPrompts: Multilingual Evaluation of Neural Toxic Degeneration in Large Language Models (COLM'24 + Leaderboard + blog post)
ToxiGen: Using LLMs to detect and mitigate implicit social biases (ACL'22 + dataset). ToxiGen has been used while training Llama2, Code Llama, phi-1.5, phi-2, and other LLMs, and to detect toxicity in Econ Forums and Laws.
Biomedical Data Science
Demographic Bias in Misdiagnosis by Computational Pathology Models (Nature Medicine)
Language Models are Surprisingly Fragile to Drug Names in Biomedical Benchmarks (EMNLP'24 + leaderboard)
In the News
Our work drawing lessons from aviation safety for health AI was covered by MIT News and Innovate Healthcare
GRACE was featured in the Microsoft Research blog
ToxiGen was covered by TechCrunch and Microsoft Research
Our work on Fair Explainability was covered by MIT News
Misc
Outside of research, I enjoy bouldering, biking, books (science fiction/science fact), birding, juggling, vegan cooking, and playing guitar. I also spent a summer living at BioSphere 2 in Arizona.