Rhetorical Structure Theory (RST) parsing, which maps discourse relations between text spans, has traditionally been dominated by discriminative models built on pretrained encoder-only language models. These models classify relationships rather than generating text, and have set the standard for the task.

A new arXiv preprint, "Worth Their Weights? Evaluating Generative Models for Discourse Parsing," asks whether the recent availability of high-quality generative language models can shift this balance. The title itself poses the central question: do these larger, more computationally expensive models offer a real advantage over the established discriminative approach?

The abstract provided is truncated, so the paper's specific methodology, benchmarks, and conclusions are not yet visible. The full text would be needed to see how the authors define "worth" and whether generative models live up to that standard.