A recent study published on arXiv found that language models from different vendors tend to select the same subset of papers when generating citations. The researchers tested 11 models from three vendors on a panel of 120 real papers with fabricated authors and citation counts. They found that all models concentrated their citations on a small number of papers, with one component explaining up to 73% of variation across their preference maps. This suggests that current language models impose a common content-level filter on scientific attention, even when every reference is real and every paper is equally visible. --- Why it matters: This matters because it shows that the way language models generate citations can perpetuate a 'citation monoculture', where only a small number of papers are consistently cited. This can limit the diversity of ideas in academic research and make it harder for new contributions to be recognized. Source: https://arxiv.org/abs/2608.19230