The default perspective of a LLM is a centroid of the training data + alignment, roughly. By cueing it to take different perspectives, you can get away from th

The default perspective of a LLM is a centroid of the training data + alignment, roughly. By cueing it to take different perspectives, you can get away from that centroid. By providing a variety of perspectives that sample the space evenly you can get a low bias estimator. I also use anti-refusal techniques to reduce the impact of alignment, however I'm not instructing the model to take any extreme perspectives so I don't think it's that big of a deal in the first place.