Postdoctoral Research Associate, Machine Unlearning and Model Editing for AI Biosecurity

About the School

The University of Virginia School of Data Science, the first of its kind in the nation, advances discovery, innovation, and societal impact through collaborative, open, and responsible data science research and education. Founded in 2019, the School brings together expertise across business, computation, engineering, humanities, law, mathematics, social sciences, and statistics to address complex, real-world challenges. Its academic offerings include a B.S. in Data Science, an undergraduate minor, residential and online M.S. in Data Science programs, and a Ph.D. in Data Science, all designed to prepare students for a rapidly evolving data-driven world.

About the Position

This position develops and evaluates machine unlearning and model editing methods that selectively reduce hazardous biological capabilities in AI systems while preserving beneficial scientific functions. The researcher reports to Assistant Professor Tom Hartvigsen and will work closely with other SDS faculty members including Chirag Agarwal, and Stephen Turner, and works with faculty in interpretability and with a laboratory partner that leads adversarial red teaming. The role centers on implementing, innovating, and comparing model editing and unlearning methods, measuring safety--utility tradeoffs against both benchmarks and realistic task batteries, and leading technical development of an open evaluation suite for AI biosecurity. Strong familiarity with biology and biosecurity is important, as the work targets biological capabilities and connects to a human-subjects evaluation running in parallel.

Key Responsibilities

• Implement and compare machine unlearning and model editing methods, including gradient-based fine-tuning, representation-level edits, and inference-time steering

• Design and run experiments that measure how interventions affect benchmark scores and real-world task performance, producing safety-utility curves

• Develop adversarial testing protocols with the laboratory partner, including prompt-based jailbreaks, fine-tuning recovery, and ensemble attacks

• Lead engineering of the open-source UBS-Bio evaluation suite, including baselines, metrics, and documentation

• Support interpretability analyses that identify which model representations encode hazardous versus beneficial capabilities

• Prepare and present manuscripts and publish and maintain reproducible code releases

Minimum Qualifications

• Doctoral degree (PhD or equivalent) in data science, computer science, machine learning, or a related field, completed at the time of hire

• Strong programming in Python and hands-on experience with modern ML frameworks such as PyTorch and Hugging Face Transformers

• Track record of publications in machine learning, natural language processing, and/or biosecurity

• Demonstrated experience training, finetuning, or post-training for large language models

• Software engineering practices that support reproducible and reusable research tools

Preferred Qualifications

• Understanding of biology, biosecurity, or dual-use research considerations

• Experience with machine unlearning, model editing, or related capability-mitigation methods

• Experience with mechanistic interpretability or representation analysis

• Familiarity with adversarial robustness, red-teaming, or jailbreak evaluation

• Experience releasing and maintaining open-source ML evaluation tooling

• Familiarity with secure computing environments and controlled-access model arrangements