Does fine-tuning on new facts increase hallucination risk?

It is often conjectured that supervised fine-tuning on facts the model never saw in pretraining teaches it to hallucinate — by training it to assert things ungrounded in its knowledge. This work tests that in a controlled closed-book QA setup, varying the proportion of fine-tuning examples that introduce new knowledge. Two findings: LLMs struggle to acquire new factual knowledge through fine-tuning — Unknown examples are fit significantly slower than examples consistent with existing knowledge; but as those Unknown examples are eventually learned, they linearly increase the model's tendency to hallucinate on pre-existing knowledge. So the harm shows up as a form of overfitting on the slow-to-fit Unknown examples.

The keeper is a concrete fine-tuning practice: prefer early-stopping over a fixed step count, and consider filtering out Unknown examples (or keeping a few to teach uncertainty expression) — because the act of forcing new facts in degrades grounded recall.

This completes a tight loop in the vault's knowledge-acquisition thread. It is the empirical mechanism behind Can models store unlimited facts without growing larger? (externalize facts rather than fine-tune them in), it complements Does teaching question patterns before document training improve knowledge access? (encoding order matters), and it shares the leakage/overfitting concern of Does repeated sensitive data in fine-tuning cause memorization?.