Steering given distributions towards ideal distributions, where fairness and accuracy are not at a tradeoff.

Perform input dependent steering of MLLMs by learning to predict steering vectors from input context