I’m currently interested in the privacy, vulnerabilities, and behavior of generative models. I’m also exploring problems in attribution and reinforcement learning. My Google Scholar is linked here.

Publications

Wang, Jeffrey G.*, Marvin Li*, Jason Wang*, and Seth Neel. “Pandora’s White-Box: Pushing the Limit of Privacy Attacks Against Open LLMs.” Under Submission. [Link]

Li, Marvin*, Jason Wang*, Jeffrey G. Wang*, and Seth Neel. “MoPe: Model Perturbation-based Privacy Attacks on Language Models.” Empirical Methods for Natural Language Processing (2023). [Link]

Chakraborty, Abhijit*, Jeffrey G. Wang*, and Ferhat Ay. “dcHiC detects differential compartments across multiple Hi-C datasets.” Nature Communications (2022). [Link]

Details. Our privacy research is ongoing and was also featured in the NeurIPS SoLaR workshop, where I served on the program committee.¹

The software developed for dcHiC is open-source on GitHub. It was featured as an oral presentation in the Regulatory and Systems Genomics track of ISMB, a poster at RECOMB, and an oral presentation in the 4D Nucleome Working Group. For this work, I was a Finalist of the 2021 Regeneron Science Talent Search.

A Sketch of my Research Philosophy

A unifying theme of my research is computation. I find it incredible that we can just throw more and more of the right compute toward approximating some function undergirding the mystery we wish to solve, and usually that mystery shatters.

In my opinion, the best research consists of theoretically grounded work that is empirically validated through capable engineering. In our current explosion of machine learning research, this means that I particularly value building robust systems, creating efficient implementations, and writing clean code.

In general, I take a systems approach toward epistemology. I have found that I love building knowledge by repeatedly depth-first searching into new topics, gleaning as much as I can, and building infrastructure for maintaining/revisiting this knowledge. In research, I enjoy taking approaches that synthesize techniques between several domains, and often find–in the process of DFS’ing—-that hidden structure emerges between seemingly disparate areas.