I am a Stanford PhD student working on computer vision and machine learning. I'm fortunate to be advised by Serena Yeung-Levy and to be supported by the Quad Fellowship.
I work on vision-language models: evaluation, agentic systems, and training environments. I often apply these methods to biology and medicine, and to questions about how AI can advance science.
Detecting artifacts in AI-generated images can improve generators. Older works fine-tunes VLMs, but we find that scaffolding with small datasets can be enough. Key tools: in-context learning, prompt optimization, and multi-component design—we improve each.
LLM agents that search and reason over documents can be trained with reinforcement learning with verifiable rewards (RLVR), but require training environments. We propose a scalable method to generate synthetic QA datasets from scientific papers.
MicroVQA is an expert-curated benchmark for research-level reasoning in biological microscopy. We also propose a method called RefineBot for removing language shortcuts from multiple-choice VQA.
The BIOMEDICA dataset has 6 million scientific articles and 24 million image-text pairs for training vision-language models in biomedicine. We use it to train state-of-the-art embedding models for biomedical images.
We propose Video Action Differencing (VidDiff), a new task for detecting subtle variations in how actions are performed between two videos. We release a benchmark spaning diverse skilled actions, and a baseline method that is a simple agentic workflow.
We show that 2D diffusion models like StableDiffusion have 3D control in their text input space which we call '3D view tokens'.
Unsupervised shape representations of cells and organelles are erroneously sensitive to image orientation, which we mitigate with equivariant convolutional network encoders in our method, O2VAE.
A proteomics map of human subcellular architecture, led by the Chan-Zuckerberg Biohub.