Apple researchers find strongest LLMs are not the most reliable writing-style judges

Apple researchers released a paper, accepted at the Can We Trust AI Evaluation? workshop at NeurIPS 2026, examining how reliably large language models judge whether generated text matches a user's preferred writing style, such as an upbeat tone. They produced text with four LLMs on three long-form writing tasks, gathered human judgments of which of two responses was better, collected matching LLM judge rankings, and compared the results.

Judge accuracy tracked general model capability only weakly, with a Pearson correlation of 0.2 or less, and shifted with the writing task and the criteria being scored. Judges were more consistent and performed better when comparing response pairs than when rating responses one at a time. The most capable models did not best reproduce human labels, which the authors link to heightened sensitivity to small details.