Scaling Laws for Safety Probes

New scaling laws predict safety probe performance for language models

Researchers have developed scaling laws to predict the performance of safety probes used to monitor language models. By analyzing six new safety datasets and training probes on 20 base models across five families, they …