You can predict how well a prompt will perform from its grammar, before spending a single inference call

arXiv 2608.27621·low signalXBlueskyLinkedInCopy link

A study of 9,000 linguistically controlled prompt variants generated from 100 seeds across 30 linguistic metrics (arXiv 2608.27621, Aug 27), evaluated with five open-source LLMs on 625 annotated requirements, found that measurable linguistic properties significantly predict prompt performance ahead of inference, with R2 between 0.38 and 0.42 at q<0.05. Syntactic and morphosyntactic features carry most of the signal; the cross-model predictors are compound dependency distribution, conjunction density, and word and sentence length. Two useful negatives: the linguistic patterns that hurt LLM performance overlap with those that make text harder for humans to read, and lexical variety turned out to be irrelevant as a quality dimension. Useful as a cheap filter in front of an expensive prompt-optimization pipeline.