LoRAs are better at steering models to produce correct answers from their data set than imparting new knowledge.