Tag
This paper systematically compares fine-tuned encoder classifiers (ModernBERT family) against decoder-based safety judges for LLM adversarial evaluation, finding that encoders can offer a cost- and latency-efficient alternative without significant performance loss.