Why do instruction-tuned models hallucinate more than b...

The primary cause of increased hallucination in instruction-tuned models is the ‘citation trap.’ During Supervised Fine-Tuning (SFT), the model learns type signatures and stylistic patterns unconditionally. If the training data contains responses with citations, the model learns to mimic the structure of a citation as a stylistic output, even when it lacks the underlying factual knowledge to support it. To fix this, developers should employ on-policy training that specifically rewards verifiable correctness, or implement strict data hygiene to remove unverifiable factual statements before the SFT process begins.