Why might a model trained on high-quality, expert-level...

When SFT examples contain complex knowledge that the model did not acquire during pre-training, it learns to mimic the output structure of the expert data. This includes adopting a confident, detailed, and cited tone. Because the training loss penalizes incorrect token prediction more heavily than it rewards honesty or uncertainty, the model learns to fill these structural templates with plausible-sounding fabrications rather than admitting it lacks the factual information.