Why can fine-tuning on domain facts INCREASE hallucination?

The intuition feels right: show the model your company's facts so it learns them. In practice it often makes hallucination worse, not better. The reason is that **fine-tuning** is great at teaching behavior and weak at teaching facts. With a few hundred or few thousand examples, the model quickly picks up the style of your domain: the phrasing, the structure, the confident expert tone. That part transfers fast. What does not transfer is reliable factual memory. Pretraining saw each fact across billions of tokens, many times, in many phrasings. A small fine-tune cannot reproduce that, so the model learns to sound like an expert without actually knowing the facts. The dangerous combination is a confident voice plus shaky knowledge. When asked something outside the training set, the model makes up an answer in the same authoritative tone. It bluffs convincingly, which is harder to catch than an obvious error. The fix is to put facts where they belong, in a retrieval system. Let the model read the relevant documents at query time and answer grounded in them. Reserve fine-tuning for format and tone, and let retrieval carry the facts.

**The mechanism is a mismatch.** Fine-tuning teaches behavior far more efficiently than it teaches facts. A few thousand examples installs the domain style quickly. The same budget cannot install reliable factual memory. Pretraining wove facts into the weights across billions of token encounters. A small supervised fine-tune cannot reproduce that scale of memorization. So the model absorbs the confident expert voice of your data without absorbing the facts underneath. Now you have the worst pairing. The model speaks with domain authority but lacks the knowledge to back it. On any question outside the training set, it confabulates fluently in that trained voice. You raised its confidence without raising its competence, which is a net loss for factuality. Research from 2024 sharpened this. Fine-tuning on genuinely new factual knowledge tends to increase hallucination, especially on facts the model only half knew. The training pushes it to commit harder to shaky memories. The fix is to retrieve facts at query time and let the model generate grounded in that context. Reading and integrating retrieved text is a behavior models are good at. Reserve fine-tuning for format and domain style, layer retrieval for the facts, and you get both wins.