Definition
Plain language
A score for how much machine-generated text statistically resembles human writing overall.
As stated in the literature
A distributional metric comparing generated and human text distributions in an embedding space, used to evaluate open-ended text-generation quality.
Why it matters: It gives a way to measure whether generated text matches the broad statistical feel of human writing, rather than judging one sentence at a time.
For example, it compares a batch of machine-written paragraphs against human-written ones and reports how closely the two pools resemble each other overall.
Heard on the show
“There's a metric called MAUVE that measures whether your generated text, in aggregate, statistically resembles human text — does the distribution look human.”Episode 127 — What Diffusion Language Models Were Missing: A Map, Not an Algorithm