What are the failure modes of perplexity-based filtering?

Perplexity-based filters have three main failure modes. First, because n-grams are local, text that is globally incoherent can still pass the filter. Second, the filter is language-model-centric, meaning it often penalizes stylistically unique but valid content such as poetry or code-switching. Finally, if the reference corpus used to train the KenLM does not align with the target domain, the filter will produce inaccurate quality scores.