Can RLHF be More Efficient with Imperfect Reward Models? A Policy Coverage Perspective | Read Paper on Bytez