Tag
This paper presents RLCDAlignBench, a benchmark for evaluating Jev, a model trained with reinforcement learning for calibrated decisions, as a zero-shot detector for ten AI alignment failures with high performance and cost efficiency.