AI reviewers may reach different judgments on manuscripts that present the same underlying science in different wording, a behavior that could reward rhetorical optimization rather than genuine scientific improvement, according to a new paper posted on Hugging Face.
The authors formulate Rhetorical Robustness as the joint requirement of stability across content-preserving rewrites and discrimination across papers. To study it, they introduce RobustReview, described as a controlled full-manuscript benchmark containing 1,260 manuscript versions, and evaluate 30 reviewer configurations on it.
The benchmark reveals what the authors call false robustness, where low sensitivity to rewrites coincides with score collapse across papers. It also shows that human alignment and rhetorical robustness rank reviewers differently, and that the evaluated content-focused prompting protocol does not consistently improve robustness across backbones.
Motivated by these findings, the authors introduce SciCore, a dual-branch reviewer that averages a full-manuscript judgment with a judgment based on an extracted, structured science core. This design combines manuscript-level assessment with a content-normalized view intended to reduce rhetorical sensitivity.
In the paper's primary GPT-5.5 comparison, SciCore is reported to achieve a leading joint stability-discrimination profile among the benchmarked reviewers while maintaining competitive human alignment. The authors conclude that rhetorical robustness is a distinct evaluation target and that science-core review shows potential for improving it.