The TIER benchmark assesses LLM safety across four risk domains and four levels of threat implicitness. Results indicate that safety responses evolve gradually from refusal to compliance rather than following a binary shift, particularly in response to sophisticated jailbreaks.
HOW THIS AFFECTS YOU
●
researcherThis allows for more granular study of the transition between refusal and compliance in LLM safety training.
●
policyYou can use more nuanced, non-binary metrics to evaluate the safety robustness of deployed models.