A Rewrite That Keeps the Label Has Already Read It: Why an LLM Paraphrase Can Raise a Sentiment Classifier's Accuracy Without the Classifier Getting Better, Four Accounts of the Gain, and the Re-Annotation and Neutral-Rate Controls That Would Separate Them

Abstract

(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. A 2026 study reports that sentiment classifiers score higher on reviews rewritten by small language models than on the human originals, by up to 5.8 accuracy points, and reads the gain as stylistic alignment: the rewrite strips idiosyncrasy and hands the classifier text closer to what it was trained on. It concludes that benchmarks cleaned or paraphrased by language models will overstate classifier performance. Two older literatures point the other way. Paraphrase and style-transfer attacks drive the same classifiers down, and a 2025 study of fifty thousand rephrased tweets finds that rewriting pulls sentiment toward neutral, a pull its robustness checks found again when the prompt asked for the sentiment to be kept. In this analysis the author reviews and synthesises published results to show that neither direction measures the classifier alone, because every one of these designs scores the rewritten text against the label of the text it replaced. The paper separates four accounts of a rewrite-induced gain: alignment of style with the classifier's training data; a fall in the rate of neutral predictions, which the anchor's scoring counts as errors; the paraphraser acting as an annotator that decides the polarity and then writes it plainly; and the removal of hard, mixed and ironic items. It shows from the anchor's released code that its prompt demands one of two polarities and that a neutral prediction is always scored wrong, that its own count of label flips is mislabelled: the printed 291 is the net gain, not the number of improvements, and that a test-time rewriting study found rewritten inputs still sitting with the out-of-distribution data in embedding space. The paper consolidates the measured values in two tables and a figure, states a licensing procedure as an algorithm, counts from the anchor's released predictions that a fall in neutral predictions carries about half of the largest gain, and names the re-annotation, paraphraser-baseline and prompt-ablation controls that would settle which account holds for the rest.