Extending Activation Steering to Broad Skills and Multiple Behaviours
This work addresses safety risks from dangerous capabilities in language models, but it is incremental as it builds on existing activation steering techniques.
The paper investigated activation steering for broad skills and multiple behaviors in large language models, finding that steering broader skills is competitive with narrower skills and that simultaneous injection of individual steering vectors for different behaviors is promising, while combining them into one vector is largely unsuccessful.
Current large language models have dangerous capabilities, which are likely to become more problematic in the future. Activation steering techniques can be used to reduce risks from these capabilities. In this paper, we investigate the efficacy of activation steering for broad skills and multiple behaviours. First, by comparing the effects of reducing performance on general coding ability and Python-specific ability, we find that steering broader skills is competitive to steering narrower skills. Second, we steer models to become more or less myopic and wealth-seeking, among other behaviours. In our experiments, combining steering vectors for multiple different behaviours into one steering vector is largely unsuccessful. On the other hand, injecting individual steering vectors at different places in a model simultaneously is promising.