Scott K. Geng
sgeng@cs.washington.edu

Hi! I am a PhD student at the University of Washington, where I am very fortunate to be advised by Pang Wei Koh and Ranjay Krishna. I am broadly interested in computer vision and natural language processing. My doctoral work is supported by an NSF Graduate Research Fellowship.

Previously, I graduated with a BA in Math and Computer Science from Columbia University, where I was lucky to be introduced to research by Carl Vondrick and Junfeng Yang.

Email  /  CV  /  Google Scholar  /  Twitter  /  GitHub

Our hypothesis: the relative quality delta between two weak data points suffices to improve a stronger student modeling. Scaling up, delta learning enables dead simple state-of-art language model post-training with only weak (cheap) data.

RLVR with very silly rewards (assign rewards randomly, reward incorrect labels) can massively boost math perf. in Qwen models—but not other models! Suggests that RLVR (at current scales) mostly elicits existing knowledge from model.

Does synthetic data from generative AI truly allow us to bootstrap and surpass the original real data used to train the generator? We propose a principled baseline to ground this question empirically, and find no — not yet.

We introduce a language-augmented vision framework for modeling social interactions in videos of two-person conversations. To study this problem, we create the RealTalk video dataset with 100+ hours of in-the-wild conversations.

We identify the novel problem of zero-shot adversarial robustness and propose a new text-grounded adversarial training objective that can help make CLIP robust while preserving its ability to generalize.

The semantic meaning of code is explictly measureable as the CPU's runtime memory values. Predicting execution traces is thus a natural self-supervised task, which we leverage to learn good code representations.

Low-frequency brain waves are correlated with symptom severity in sporadic essential tremor but not familial (i.e. genetic). Suggests difference in mechanism.

At Columbia.

Course Assistant (Spring 2021, Fall 2021): COMS 4771 Machine Learning

Jon Barron has a very clean website.
Last updated: July 15th, 2025.