Harve

New methods tackle reward hacking in AI training

Researchers are developing new methods to combat reward hacking in reinforcement learning from human feedback (RLHF) systems. Several papers introduce techniques to detect and mitigate scenarios where models exploit bia…