Prompt Sensitivity in Tool-Augmented Model Comparisons.

arXiv:2610.31148v1 Announce Type: cross Abstract: Language models are compared on question-answering benchmarks that pose each question in one wording, and a reworded question can change which of two models scores higher. We ask whether giving the models tools