Learning Using Privileged Information

The field studies how extra information available during training—such as future observations, detailed labels, other sensor modalities, or expert feedback—can help a model that will not receive it at deployment. Common approaches include teacher–student distillation, privileged-feature SVMs, representation learning, and reinforcement learning with guided or hindsight-informed rollouts. These methods are used in settings from robotics and autonomous driving to medical imaging, vision-language models, and reasoning agents. The central tension is that richer guidance can improve learning, but may also teach shortcuts or bias the student toward evidence it cannot access, so useful supervision must be relevant, reliable, and transferable to ordinary inputs.