Zoe Wu

Hi, I'm Zoe! I'm currently a Post-baccalaureate Researcher at the Kempner Institute at Harvard University.

My research is focused on developing algorithmic techniques that enable inference-time control over generative modeling. My research seeks to use the theoretical arguments of advanced statistical inference techniques to create empirically performant, principled algorithms with applications ranging across robotics, protein generation, reasoning, and computer vision.

A major challenge now in generative AI research has been how we can steer the qualitative properties of generated content, incorporate external control into the output, or adjust for biases in the training data. Such aims find uses not only across language and image generation, but also across protein generation — where samples must maximize performance metrics (e.g. stability, affinity in proteins, closeness to target structures) — and robotic manipulation — where robotic policies must align with real-world constraints.