My research is in Machine Learning and Artificial Intelligence. I study how to build AI systems that reuse learned knowledge and adapt to new tasks and goals through context and feedback. My long-term goal is to build general-purpose robots and design molecules with desired biological functions.
Model Architecture and Pretraining. I believe the right architecture and pretraining lay the foundation for generalization. The success of language models motivates me to bring their task flexibility to robotics and molecular design. I study how transformers support in-context learning (ICL) and computation. I prove that transformers can train deep neural networks through in-context gradient descent without updating their own weights. I also prove that fixed-weight transformers can emulate prompt-specified algorithms and compose simple computations into more complex ones. The prompt acts as a program, and the transformer as an interpreter. Next, I aim to understand when pretraining learns these capabilities and how training data and objectives affect generalization. I aim to use these insights to design architectures that separate task descriptions and data from the computation needed to solve a task. I also aim to build in physical and biological inductive biases, so models can learn and adapt with less data and computation across new tasks and unfamiliar environments.
Post-Training and Inference-Time Guidance. I study how pretrained models adapt to new tasks and objectives through fine-tuning, reinforcement learning (RL), and inference-time guidance. I develop theory to understand when these methods work and how efficiently they use data and computation, and use these insights to design better algorithms. A central perspective in my work is to treat a pretrained model’s distribution as a prior and combine it with task-specific information to define a target distribution. This target brings together knowledge learned during pretraining and the requirements of a new task. I study how to construct these target distributions and sample from them efficiently. Within this framework, RL uses rewards to define task-specific objectives and fine-tunes the model to approximate the resulting target distribution. Inference-time guidance instead steers sampling toward the target without updating the pretrained model. My goal is to develop reliable and efficient adaptation methods for new robotic tasks and molecular design goals and constraints.
Data for Generalization. A future direction is to study what data models need to generalize. Beyond data size and quality, I want to understand which tasks, examples, and outcomes a training set should include. I am interested in how to organize data so models learn shared structure across tasks, and how to choose new data when existing experience is not enough. For robots, this includes demonstrations, interaction data, and failed attempts. For molecular design, it includes molecular structures, measured properties, and experimental results. My goal is to connect data collection with model design and adaptation, and help models learn from fewer interactions and experiments.