Doctoral Speaking Skills Talk - Yash Savani
April 30, 2026 11:00AM—12:00PM
Location:
4303
-
Gates and Hillman Centers
Speaker:
YASH SAVANI,
Ph.D. Student, Computer Science Department, Carnegie Mellon University
https://yashsavani.com/
Reasoning models generate extended chain-of-thought traces that simultaneously serve users and expose valuable training signals to potential competitors. A competitor's model can be fine-tuned on these traces to replicate frontier capabilities at a fraction of the original training cost, while stripping away the safety features that took significant effort to build. Current defenses, such as suppressing reasoning traces or blocking suspected users, impose costs on legitimate users and undermine safety monitoring. In this talk, I will present Antidistillation Sampling, a decoding-time method that modifies the teacher model's token sampling distribution to make its outputs less useful for distillation while preserving their utility for end users.
Presented in Partial Fulfillment of the Speaking Skills Requirement
Contact
Matt Stewart