Definition
Perplexity probes use a model’s assigned perplexity on a piece of text as a signal for some property of that text — whether it’s in-distribution, machine-generated, memorized, or just unusual. They’re cheap, model-internal, and often surprisingly informative.