steerability

This paper investigates when activation steering succeeds or fails for LLMs by analyzing early decoding dynamics. The authors introduce ASTEER, a large testbed of steered generations, and train a GBDT classifier to predict steering outcomes from early hidden states, enabling efficient steering strength search.

This paper investigates the ability of LLMs-as-judges for safety to adapt to contextual information and varying safety definitions, finding that they are largely rigid and fail to adjust when the context contradicts their internal priors.