Mitigating Reward Hacking in Reinforcement Learning from Human Feedback via Advantage Sign Robustness
arXiv:2604.02986v2 Announce Type: replace-cross Abstract: Reward models (RMs) used in reinforcement learning from human feedback (RLHF) are vulnerable to reward hacking: as the policy maximizes a learned proxy reward, true quality plateaus or…
Read the full story at arXiv cs.AI ↗
ImpactMinor
Published11 h ago (Tue, 10 Nov 2026 05:00:00 GMT)
AuthorShinnosuke Ono, Johannes Ackermann, Soichiro Nishimori, Takashi Ishida, Masashi Sugiyama
MoreLess
Impact score3/100
RatingProbability of 95%.
RegionsGlobal
Retrievedvia rss
ClassifiedTue, 10 Nov 2026 14:02:44 GMT by classifier