Can Prompting Reliably Close the Sentence-Length Dispersion Gap?
Vittoria Lanzo ⋅ Dico Angelo
Abstract
In machine-text detection, burstiness is the variation of perplexity across a document; in everyday use it is the variation of sentence length, whose obvious statistic, the variance of sentence length, grows with mean sentence length. We give the sentence-length reading a scale-free definition: sentence boundaries as a point process in word-time, scored with the Goh–Barabási burstiness coefficient under the Kim–Jo finite-$n$ correction, bounded and with no model in the loop. The protocol is paired evaluation on 200 era-, topic-, and register-matched Wikinews articles dated 2016–2021: five prompting modes, nine models, three samples per headline and mode, uniform normalization of generated text before scoring. Scored raw, markdown scaffolding manufactures dispersion: on the audited model it produces an equivalence verdict that normalization reverses (instruction-mode 0.006 scored raw, equivalent at the 0.05 margin; 0.035 on normalized text, significantly different, $p = 2\times10^{-5}$, and equivalent only from $m^{*} = 0.052$ upward), and its uninstructed gap grows from 0.053 to 0.102. Uninstructed, eight of the nine models write significantly flatter than the human articles, a median paired gap of 0.050 to 0.128 in $B_n$. Of the forty-five model-by-mode cells, thirty-two land significantly flatter than the human articles, nine show no significant difference, and four land significantly more dispersed: machine flatness is a strong central tendency that the prompts tested move but do not reliably close. Under the same instruction, five models remain significantly flatter than the human articles, three show no significant difference and one writes significantly more dispersed text; three of the nine are equivalent within $\pm 0.05$ and five within $\pm 0.08$, two of them while differing significantly (none within $\pm 0.03$), and for one model among the nine the instruction makes the text significantly flatter than its own uninstructed output. Few-shot exemplars of typical human news prose produce text significantly flatter than the instruction alone in every model.
Successful Page Load