I am a Stanford PhD student working on machine learning. I'm fortunate to be advised by Serena Yeung-Levy and to be supported by the Quad Fellowship.
I build and evaluate LLMs and VLMs: benchmarks, RL training, agentic systems, and synthetic data. My work covers both general ML methodology and applications to biology and medicine. I graduate in June 2026.
Detecting artifacts in AI-generated images can improve generators. Older works fine-tune VLMs, but we find that scaffolding with small datasets (hundreds) can be enough. Key tools: in-context learning, prompt optimization, and multi-component design.
LLM agents that search and reason over documents can be trained with RLVR, but this requires training environments. We propose a scalable method to generate synthetic QA data from scientific papers.
Benchmark and eval methodology. MicroVQA is a benchmark for PhD-level visual reasoning in biological microscopy, used internally at frontier labs. RefineBot is a method for removing language shortcuts from 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 introduce Video Action Differencing (VidDiff), a task for comparing actions between videos via natural language. We release a benchmark covering diverse skilled actions, and a baseline method implemented as 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'. We use them for view-controlled image generation and novel view synthesis from a single image.
Unsupervised shape representations of cells and organelles are erroneously sensitive to image orientation. O2VAE fixes this with an architectural change that guarantees orientation invariance.
A proteomics map of human subcellular architecture, led by the Chan-Zuckerberg Biohub. I contributed embedding models for microscopy image classification.