Large Language Model

Mitigating Reward Over-Optimization in RLHF via Behavior-Supported Regularization
Reinforcement learning from human feedback (RLHF) is an effective method for aligning large language models (LLMs) with human values. …
PDF Cite