Doctoral Thesis Proposal - Yash Savani
— 5:00pm
Location:
In Person and Virtual - ET
-
Gates Hillman 9115 and Zoom
Speaker:
YASH SAVANI
,
Ph.D. Student
Computer Science Department
Carnegie Mellon University
https://yashsavani.com/
This thesis proposal presents controlled generation methods for foundation models, focusing on gradient-based steering to improve training and robustness. For diffusion and flow models, I introduce Diffusing Differentiable Representations (NeurIPS 2024), which guides the training of differentiable representations, such as Neural Radiance Fields, by pulling back the score function through the differentiable render function. I also present work with Adobe Research on temporal credit assignment for policy gradient methods, enabling more effective training of flow models via GRPO-style reinforcement learning.
For large language models, I present two methods for controlled generation. The first maximizes resource utilization in GRPO-style reinforcement learning by selectively dropping low-variance trajectories (in submission). The second, Antidistillation Sampling (NeurIPS 2025), steers generation to defend against distillation attacks using precomputed proxy gradients. Together, these contributions establish a unified framework for controlled generation across modalities, with applications spanning creative content synthesis, model protection, and efficient training.
Thesis Committee
J. Zico Kolter (Chair)
Aviral Kumar
Nicholas M. Boffi
Krishna Kumar Singh (Adobe Research)
Additional Information
In Person and Zoom Participation. See announcement.
For More Information:
matthewstewart@cmu.edu