RLHF

Mitigating Reward Hacking in RLHF via Advantage Sign Robustness
We address reward hacking in RLHF by identifying flipped advantage signs as a key cause and proposing Sign-Certified Policy Optimization (SignCert-PO), a lightweight method that down-weights non-robust completions during policy optimization.
Shinnosuke Ono, Johannes Ackermann, Soichiro Nishimori, Takashi Ishida, Masashi Sugiyama
PDF DOI
Mitigating Reward Hacking in RLHF via Advantage Sign Robustness