Claudio Mayrink Verdun

We characterize reward hacking in inference time alignment and demonstrate via three inference-time methods that we can leverage hedging to provably mitigate reward hacking.

Using coding theory and heavy tails to design powerful watermarks with zero distortion.