When large language models are aligned via supervised fine-tuning, they may encounter new factual information that was not acquired through pre-training. It is often conjectured that this can teach the model the behavior of hallucinating factually incorrect responses, as the model is trained to generate facts that are not grounded in its pre-existing knowledge. In this work, we study the impact of such exposure to new knowledge on the capability of the fine-tuned model to utilize its pre-existing knowledge. To this end, we design a controlled setup, focused on closed-book QA, where we vary the proportion of the fine-tuning examples that introduce new knowledge. We demonstrate that large language models struggle to acquire new factual knowledge through fine-tuning, as fine-tuning examples that introduce new knowledge are learned significantly slower than those consistent with the model's knowledge. However, we also find that as the examples with new knowledge are eventually learned, they linearly increase the model's tendency to hallucinate. Taken together, our results highlight the risk in introducing new factual knowledge through fine-tuning, and support the view that large language models mostly acquire factual knowledge through pre-training, whereas fine-tuning teaches them to use it more efficiently.

On Wednesday, a polar vortex brought bitter cold to the Midwest. Overnight, Chicago reached a low of 21 degrees Fahrenheit below zero, making it slightly colder than Antarctica, Alaska, and the North Pole. Wind chills were 64 degrees below zero in Park Rapids, Minnesota, and 45 degrees below zero in Buffalo, North Dakota, according to the National Weather Service. Schools, restaurants, and businesses closed, and more than 1,000 flights have been canceled. Even the United States Postal Service stalled mail delivery, temporarily.